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/etherdevice.h>
37 #include <linux/slab.h>
38 #include <linux/netdevice.h>
39 #include <linux/vmalloc.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 <linux/mlx4/driver.h>
47 #include <linux/mlx4/qp.h>
50 #include <rdma/mlx4-abi.h>
52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
53 struct mlx4_ib_cq *recv_cq);
54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
55 struct mlx4_ib_cq *recv_cq);
58 MLX4_IB_ACK_REQ_FREQ = 8,
62 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83,
63 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
64 MLX4_IB_LINK_TYPE_IB = 0,
65 MLX4_IB_LINK_TYPE_ETH = 1
70 * Largest possible UD header: send with GRH and immediate
71 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
72 * tag. (LRH would only use 8 bytes, so Ethernet is the
75 MLX4_IB_UD_HEADER_SIZE = 82,
76 MLX4_IB_LSO_HEADER_SPARE = 128,
84 struct ib_ud_header ud_header;
85 u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
86 struct ib_qp *roce_v2_gsi;
90 MLX4_IB_MIN_SQ_STRIDE = 6,
91 MLX4_IB_CACHE_LINE_SIZE = 64,
96 MLX4_RAW_QP_MSGMAX = 31,
103 static const __be32 mlx4_ib_opcode[] = {
104 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND),
105 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO),
106 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
107 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
108 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
109 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
110 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
111 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
112 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
113 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
114 [IB_WR_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR),
115 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
116 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
119 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
121 return container_of(mqp, struct mlx4_ib_sqp, qp);
124 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
126 if (!mlx4_is_master(dev->dev))
129 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
130 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
134 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
139 /* PPF or Native -- real SQP */
140 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
141 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
142 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
145 /* VF or PF -- proxy SQP */
146 if (mlx4_is_mfunc(dev->dev)) {
147 for (i = 0; i < dev->dev->caps.num_ports; i++) {
148 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
149 qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
158 return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
161 /* used for INIT/CLOSE port logic */
162 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
167 /* PPF or Native -- real QP0 */
168 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
169 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
170 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
173 /* VF or PF -- proxy QP0 */
174 if (mlx4_is_mfunc(dev->dev)) {
175 for (i = 0; i < dev->dev->caps.num_ports; i++) {
176 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
185 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
187 return mlx4_buf_offset(&qp->buf, offset);
190 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
192 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
195 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
197 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
201 * Stamp a SQ WQE so that it is invalid if prefetched by marking the
202 * first four bytes of every 64 byte chunk with
203 * 0x7FFFFFF | (invalid_ownership_value << 31).
205 * When the max work request size is less than or equal to the WQE
206 * basic block size, as an optimization, we can stamp all WQEs with
207 * 0xffffffff, and skip the very first chunk of each WQE.
209 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
217 struct mlx4_wqe_ctrl_seg *ctrl;
219 if (qp->sq_max_wqes_per_wr > 1) {
220 s = roundup(size, 1U << qp->sq.wqe_shift);
221 for (i = 0; i < s; i += 64) {
222 ind = (i >> qp->sq.wqe_shift) + n;
223 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
224 cpu_to_be32(0xffffffff);
225 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
226 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
230 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
231 s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
232 for (i = 64; i < s; i += 64) {
234 *wqe = cpu_to_be32(0xffffffff);
239 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
241 struct mlx4_wqe_ctrl_seg *ctrl;
242 struct mlx4_wqe_inline_seg *inl;
246 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
247 s = sizeof(struct mlx4_wqe_ctrl_seg);
249 if (qp->ibqp.qp_type == IB_QPT_UD) {
250 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
251 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
252 memset(dgram, 0, sizeof *dgram);
253 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
254 s += sizeof(struct mlx4_wqe_datagram_seg);
257 /* Pad the remainder of the WQE with an inline data segment. */
260 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
262 ctrl->srcrb_flags = 0;
263 ctrl->qpn_vlan.fence_size = size / 16;
265 * Make sure descriptor is fully written before setting ownership bit
266 * (because HW can start executing as soon as we do).
270 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
271 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
273 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
276 /* Post NOP WQE to prevent wrap-around in the middle of WR */
277 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
279 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
280 if (unlikely(s < qp->sq_max_wqes_per_wr)) {
281 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
287 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
289 struct ib_event event;
290 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
292 if (type == MLX4_EVENT_TYPE_PATH_MIG)
293 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
295 if (ibqp->event_handler) {
296 event.device = ibqp->device;
297 event.element.qp = ibqp;
299 case MLX4_EVENT_TYPE_PATH_MIG:
300 event.event = IB_EVENT_PATH_MIG;
302 case MLX4_EVENT_TYPE_COMM_EST:
303 event.event = IB_EVENT_COMM_EST;
305 case MLX4_EVENT_TYPE_SQ_DRAINED:
306 event.event = IB_EVENT_SQ_DRAINED;
308 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
309 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
311 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
312 event.event = IB_EVENT_QP_FATAL;
314 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
315 event.event = IB_EVENT_PATH_MIG_ERR;
317 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
318 event.event = IB_EVENT_QP_REQ_ERR;
320 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
321 event.event = IB_EVENT_QP_ACCESS_ERR;
324 pr_warn("Unexpected event type %d "
325 "on QP %06x\n", type, qp->qpn);
329 ibqp->event_handler(&event, ibqp->qp_context);
333 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
336 * UD WQEs must have a datagram segment.
337 * RC and UC WQEs might have a remote address segment.
338 * MLX WQEs need two extra inline data segments (for the UD
339 * header and space for the ICRC).
343 return sizeof (struct mlx4_wqe_ctrl_seg) +
344 sizeof (struct mlx4_wqe_datagram_seg) +
345 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
346 case MLX4_IB_QPT_PROXY_SMI_OWNER:
347 case MLX4_IB_QPT_PROXY_SMI:
348 case MLX4_IB_QPT_PROXY_GSI:
349 return sizeof (struct mlx4_wqe_ctrl_seg) +
350 sizeof (struct mlx4_wqe_datagram_seg) + 64;
351 case MLX4_IB_QPT_TUN_SMI_OWNER:
352 case MLX4_IB_QPT_TUN_GSI:
353 return sizeof (struct mlx4_wqe_ctrl_seg) +
354 sizeof (struct mlx4_wqe_datagram_seg);
357 return sizeof (struct mlx4_wqe_ctrl_seg) +
358 sizeof (struct mlx4_wqe_raddr_seg);
360 return sizeof (struct mlx4_wqe_ctrl_seg) +
361 sizeof (struct mlx4_wqe_masked_atomic_seg) +
362 sizeof (struct mlx4_wqe_raddr_seg);
363 case MLX4_IB_QPT_SMI:
364 case MLX4_IB_QPT_GSI:
365 return sizeof (struct mlx4_wqe_ctrl_seg) +
366 ALIGN(MLX4_IB_UD_HEADER_SIZE +
367 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
369 sizeof (struct mlx4_wqe_inline_seg),
370 sizeof (struct mlx4_wqe_data_seg)) +
372 sizeof (struct mlx4_wqe_inline_seg),
373 sizeof (struct mlx4_wqe_data_seg));
375 return sizeof (struct mlx4_wqe_ctrl_seg);
379 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
380 int is_user, int has_rq, struct mlx4_ib_qp *qp)
382 /* Sanity check RQ size before proceeding */
383 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
384 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
388 if (cap->max_recv_wr)
391 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
393 /* HW requires >= 1 RQ entry with >= 1 gather entry */
394 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
397 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr));
398 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
399 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
402 /* leave userspace return values as they were, so as not to break ABI */
404 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt;
405 cap->max_recv_sge = qp->rq.max_gs;
407 cap->max_recv_wr = qp->rq.max_post =
408 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
409 cap->max_recv_sge = min(qp->rq.max_gs,
410 min(dev->dev->caps.max_sq_sg,
411 dev->dev->caps.max_rq_sg));
417 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
418 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
423 /* Sanity check SQ size before proceeding */
424 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
425 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
426 cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
427 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
431 * For MLX transport we need 2 extra S/G entries:
432 * one for the header and one for the checksum at the end
434 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
435 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
436 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
439 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
440 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
441 send_wqe_overhead(type, qp->flags);
443 if (s > dev->dev->caps.max_sq_desc_sz)
447 * Hermon supports shrinking WQEs, such that a single work
448 * request can include multiple units of 1 << wqe_shift. This
449 * way, work requests can differ in size, and do not have to
450 * be a power of 2 in size, saving memory and speeding up send
451 * WR posting. Unfortunately, if we do this then the
452 * wqe_index field in CQEs can't be used to look up the WR ID
453 * anymore, so we do this only if selective signaling is off.
455 * Further, on 32-bit platforms, we can't use vmap() to make
456 * the QP buffer virtually contiguous. Thus we have to use
457 * constant-sized WRs to make sure a WR is always fully within
458 * a single page-sized chunk.
460 * Finally, we use NOP work requests to pad the end of the
461 * work queue, to avoid wrap-around in the middle of WR. We
462 * set NEC bit to avoid getting completions with error for
463 * these NOP WRs, but since NEC is only supported starting
464 * with firmware 2.2.232, we use constant-sized WRs for older
467 * And, since MLX QPs only support SEND, we use constant-sized
470 * We look for the smallest value of wqe_shift such that the
471 * resulting number of wqes does not exceed device
474 * We set WQE size to at least 64 bytes, this way stamping
475 * invalidates each WQE.
477 if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
478 qp->sq_signal_bits && BITS_PER_LONG == 64 &&
479 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
480 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
481 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
482 qp->sq.wqe_shift = ilog2(64);
484 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
487 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
490 * We need to leave 2 KB + 1 WR of headroom in the SQ to
491 * allow HW to prefetch.
493 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
494 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
495 qp->sq_max_wqes_per_wr +
498 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
501 if (qp->sq_max_wqes_per_wr <= 1)
507 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
508 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
509 send_wqe_overhead(type, qp->flags)) /
510 sizeof (struct mlx4_wqe_data_seg);
512 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
513 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
514 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
516 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
518 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
522 cap->max_send_wr = qp->sq.max_post =
523 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
524 cap->max_send_sge = min(qp->sq.max_gs,
525 min(dev->dev->caps.max_sq_sg,
526 dev->dev->caps.max_rq_sg));
527 /* We don't support inline sends for kernel QPs (yet) */
528 cap->max_inline_data = 0;
533 static int set_user_sq_size(struct mlx4_ib_dev *dev,
534 struct mlx4_ib_qp *qp,
535 struct mlx4_ib_create_qp *ucmd)
537 /* Sanity check SQ size before proceeding */
538 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes ||
539 ucmd->log_sq_stride >
540 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
541 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
544 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
545 qp->sq.wqe_shift = ucmd->log_sq_stride;
547 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
548 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
553 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
558 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
560 if (!qp->sqp_proxy_rcv)
562 for (i = 0; i < qp->rq.wqe_cnt; i++) {
563 qp->sqp_proxy_rcv[i].addr =
564 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
566 if (!qp->sqp_proxy_rcv[i].addr)
568 qp->sqp_proxy_rcv[i].map =
569 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
570 sizeof (struct mlx4_ib_proxy_sqp_hdr),
572 if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
573 kfree(qp->sqp_proxy_rcv[i].addr);
582 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
583 sizeof (struct mlx4_ib_proxy_sqp_hdr),
585 kfree(qp->sqp_proxy_rcv[i].addr);
587 kfree(qp->sqp_proxy_rcv);
588 qp->sqp_proxy_rcv = NULL;
592 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
596 for (i = 0; i < qp->rq.wqe_cnt; i++) {
597 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
598 sizeof (struct mlx4_ib_proxy_sqp_hdr),
600 kfree(qp->sqp_proxy_rcv[i].addr);
602 kfree(qp->sqp_proxy_rcv);
605 static int qp_has_rq(struct ib_qp_init_attr *attr)
607 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
613 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
616 for (i = 0; i < dev->caps.num_ports; i++) {
617 if (qpn == dev->caps.qp0_proxy[i])
618 return !!dev->caps.qp0_qkey[i];
623 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
624 struct mlx4_ib_qp *qp)
626 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
627 mlx4_counter_free(dev->dev, qp->counter_index->index);
628 list_del(&qp->counter_index->list);
629 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
631 kfree(qp->counter_index);
632 qp->counter_index = NULL;
635 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
636 struct ib_qp_init_attr *init_attr,
637 struct ib_udata *udata, int sqpn,
638 struct mlx4_ib_qp **caller_qp)
642 struct ib_qp_cap backup_cap;
643 struct mlx4_ib_sqp *sqp = NULL;
644 struct mlx4_ib_qp *qp;
645 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
646 struct mlx4_ib_cq *mcq;
649 /* When tunneling special qps, we use a plain UD qp */
651 if (mlx4_is_mfunc(dev->dev) &&
652 (!mlx4_is_master(dev->dev) ||
653 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
654 if (init_attr->qp_type == IB_QPT_GSI)
655 qp_type = MLX4_IB_QPT_PROXY_GSI;
657 if (mlx4_is_master(dev->dev) ||
658 qp0_enabled_vf(dev->dev, sqpn))
659 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
661 qp_type = MLX4_IB_QPT_PROXY_SMI;
665 /* add extra sg entry for tunneling */
666 init_attr->cap.max_recv_sge++;
667 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
668 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
669 container_of(init_attr,
670 struct mlx4_ib_qp_tunnel_init_attr, init_attr);
671 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
672 tnl_init->proxy_qp_type != IB_QPT_GSI) ||
673 !mlx4_is_master(dev->dev))
675 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
676 qp_type = MLX4_IB_QPT_TUN_GSI;
677 else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
678 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
680 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
682 qp_type = MLX4_IB_QPT_TUN_SMI;
683 /* we are definitely in the PPF here, since we are creating
684 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
685 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
686 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
691 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
692 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
693 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
694 sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
698 qp->pri.vid = 0xFFFF;
699 qp->alt.vid = 0xFFFF;
701 qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
704 qp->pri.vid = 0xFFFF;
705 qp->alt.vid = 0xFFFF;
710 qp->mlx4_ib_qp_type = qp_type;
712 mutex_init(&qp->mutex);
713 spin_lock_init(&qp->sq.lock);
714 spin_lock_init(&qp->rq.lock);
715 INIT_LIST_HEAD(&qp->gid_list);
716 INIT_LIST_HEAD(&qp->steering_rules);
718 qp->state = IB_QPS_RESET;
719 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
720 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
722 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
727 struct mlx4_ib_create_qp ucmd;
729 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
734 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
736 err = set_user_sq_size(dev, qp, &ucmd);
740 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
742 if (IS_ERR(qp->umem)) {
743 err = PTR_ERR(qp->umem);
747 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
748 qp->umem->page_shift, &qp->mtt);
752 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
756 if (qp_has_rq(init_attr)) {
757 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
758 ucmd.db_addr, &qp->db);
763 qp->sq_no_prefetch = 0;
765 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
766 qp->flags |= MLX4_IB_QP_LSO;
768 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
769 if (dev->steering_support ==
770 MLX4_STEERING_MODE_DEVICE_MANAGED)
771 qp->flags |= MLX4_IB_QP_NETIF;
776 memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
777 err = set_kernel_sq_size(dev, &init_attr->cap,
782 if (qp_has_rq(init_attr)) {
783 err = mlx4_db_alloc(dev->dev, &qp->db, 0);
790 if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
792 memcpy(&init_attr->cap, &backup_cap,
794 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
799 if (mlx4_buf_alloc(dev->dev, qp->buf_size,
800 PAGE_SIZE * 2, &qp->buf)) {
806 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
811 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
815 qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
816 GFP_KERNEL | __GFP_NOWARN);
818 qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
819 GFP_KERNEL, PAGE_KERNEL);
820 qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
821 GFP_KERNEL | __GFP_NOWARN);
823 qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
824 GFP_KERNEL, PAGE_KERNEL);
825 if (!qp->sq.wrid || !qp->rq.wrid) {
832 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
833 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
834 if (alloc_proxy_bufs(pd->device, qp)) {
840 /* Raw packet QPNs may not have bits 6,7 set in their qp_num;
841 * otherwise, the WQE BlueFlame setup flow wrongly causes
843 if (init_attr->qp_type == IB_QPT_RAW_PACKET)
844 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
845 (init_attr->cap.max_send_wr ?
846 MLX4_RESERVE_ETH_BF_QP : 0) |
847 (init_attr->cap.max_recv_wr ?
848 MLX4_RESERVE_A0_QP : 0));
850 if (qp->flags & MLX4_IB_QP_NETIF)
851 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
853 err = mlx4_qp_reserve_range(dev->dev, 1, 1,
859 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
860 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
862 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
866 if (init_attr->qp_type == IB_QPT_XRC_TGT)
867 qp->mqp.qpn |= (1 << 23);
870 * Hardware wants QPN written in big-endian order (after
871 * shifting) for send doorbell. Precompute this value to save
872 * a little bit when posting sends.
874 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
876 qp->mqp.event = mlx4_ib_qp_event;
880 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
881 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
882 to_mcq(init_attr->recv_cq));
883 /* Maintain device to QPs access, needed for further handling
886 list_add_tail(&qp->qps_list, &dev->qp_list);
887 /* Maintain CQ to QPs access, needed for further handling
890 mcq = to_mcq(init_attr->send_cq);
891 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
892 mcq = to_mcq(init_attr->recv_cq);
893 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
894 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
895 to_mcq(init_attr->recv_cq));
896 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
901 if (qp->flags & MLX4_IB_QP_NETIF)
902 mlx4_ib_steer_qp_free(dev, qpn, 1);
904 mlx4_qp_release_range(dev->dev, qpn, 1);
907 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
908 free_proxy_bufs(pd->device, qp);
911 if (qp_has_rq(init_attr))
912 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
919 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
923 ib_umem_release(qp->umem);
925 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
928 if (!pd->uobject && qp_has_rq(init_attr))
929 mlx4_db_free(dev->dev, &qp->db);
934 else if (!*caller_qp)
939 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
942 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
943 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
944 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
945 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
946 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
947 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
948 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
953 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
954 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
956 if (send_cq == recv_cq) {
957 spin_lock(&send_cq->lock);
958 __acquire(&recv_cq->lock);
959 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
960 spin_lock(&send_cq->lock);
961 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
963 spin_lock(&recv_cq->lock);
964 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
968 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
969 __releases(&send_cq->lock) __releases(&recv_cq->lock)
971 if (send_cq == recv_cq) {
972 __release(&recv_cq->lock);
973 spin_unlock(&send_cq->lock);
974 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
975 spin_unlock(&recv_cq->lock);
976 spin_unlock(&send_cq->lock);
978 spin_unlock(&send_cq->lock);
979 spin_unlock(&recv_cq->lock);
983 static void del_gid_entries(struct mlx4_ib_qp *qp)
985 struct mlx4_ib_gid_entry *ge, *tmp;
987 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
993 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
995 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
996 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
998 return to_mpd(qp->ibqp.pd);
1001 static void get_cqs(struct mlx4_ib_qp *qp,
1002 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1004 switch (qp->ibqp.qp_type) {
1005 case IB_QPT_XRC_TGT:
1006 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1007 *recv_cq = *send_cq;
1009 case IB_QPT_XRC_INI:
1010 *send_cq = to_mcq(qp->ibqp.send_cq);
1011 *recv_cq = *send_cq;
1014 *send_cq = to_mcq(qp->ibqp.send_cq);
1015 *recv_cq = to_mcq(qp->ibqp.recv_cq);
1020 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1023 struct mlx4_ib_cq *send_cq, *recv_cq;
1024 unsigned long flags;
1026 if (qp->state != IB_QPS_RESET) {
1027 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1028 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1029 pr_warn("modify QP %06x to RESET failed.\n",
1031 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1032 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1034 qp->pri.smac_port = 0;
1037 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1040 if (qp->pri.vid < 0x1000) {
1041 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1042 qp->pri.vid = 0xFFFF;
1043 qp->pri.candidate_vid = 0xFFFF;
1044 qp->pri.update_vid = 0;
1046 if (qp->alt.vid < 0x1000) {
1047 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1048 qp->alt.vid = 0xFFFF;
1049 qp->alt.candidate_vid = 0xFFFF;
1050 qp->alt.update_vid = 0;
1054 get_cqs(qp, &send_cq, &recv_cq);
1056 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1057 mlx4_ib_lock_cqs(send_cq, recv_cq);
1059 /* del from lists under both locks above to protect reset flow paths */
1060 list_del(&qp->qps_list);
1061 list_del(&qp->cq_send_list);
1062 list_del(&qp->cq_recv_list);
1064 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1065 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1066 if (send_cq != recv_cq)
1067 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1070 mlx4_qp_remove(dev->dev, &qp->mqp);
1072 mlx4_ib_unlock_cqs(send_cq, recv_cq);
1073 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1075 mlx4_qp_free(dev->dev, &qp->mqp);
1077 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1078 if (qp->flags & MLX4_IB_QP_NETIF)
1079 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1081 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1084 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1088 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1090 ib_umem_release(qp->umem);
1092 kvfree(qp->sq.wrid);
1093 kvfree(qp->rq.wrid);
1094 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1095 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1096 free_proxy_bufs(&dev->ib_dev, qp);
1097 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1099 mlx4_db_free(dev->dev, &qp->db);
1102 del_gid_entries(qp);
1105 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1108 if (!mlx4_is_mfunc(dev->dev) ||
1109 (mlx4_is_master(dev->dev) &&
1110 attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1111 return dev->dev->phys_caps.base_sqpn +
1112 (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1115 /* PF or VF -- creating proxies */
1116 if (attr->qp_type == IB_QPT_SMI)
1117 return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1119 return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1122 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1123 struct ib_qp_init_attr *init_attr,
1124 struct ib_udata *udata)
1126 struct mlx4_ib_qp *qp = NULL;
1128 int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1132 * We only support LSO, vendor flag1, and multicast loopback blocking,
1133 * and only for kernel UD QPs.
1135 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1136 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1137 MLX4_IB_SRIOV_TUNNEL_QP |
1140 MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1141 return ERR_PTR(-EINVAL);
1143 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1144 if (init_attr->qp_type != IB_QPT_UD)
1145 return ERR_PTR(-EINVAL);
1148 if (init_attr->create_flags) {
1149 if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1150 return ERR_PTR(-EINVAL);
1152 if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1153 MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1154 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1155 init_attr->qp_type != IB_QPT_UD) ||
1156 (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1157 init_attr->qp_type > IB_QPT_GSI) ||
1158 (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1159 init_attr->qp_type != IB_QPT_GSI))
1160 return ERR_PTR(-EINVAL);
1163 switch (init_attr->qp_type) {
1164 case IB_QPT_XRC_TGT:
1165 pd = to_mxrcd(init_attr->xrcd)->pd;
1166 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1167 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1169 case IB_QPT_XRC_INI:
1170 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1171 return ERR_PTR(-ENOSYS);
1172 init_attr->recv_cq = init_attr->send_cq;
1176 case IB_QPT_RAW_PACKET:
1177 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1179 return ERR_PTR(-ENOMEM);
1180 qp->pri.vid = 0xFFFF;
1181 qp->alt.vid = 0xFFFF;
1185 err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1189 return ERR_PTR(err);
1192 qp->ibqp.qp_num = qp->mqp.qpn;
1202 /* Userspace is not allowed to create special QPs: */
1204 return ERR_PTR(-EINVAL);
1205 if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1206 int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1209 return ERR_PTR(res);
1211 sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1214 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1217 return ERR_PTR(err);
1219 qp->port = init_attr->port_num;
1220 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1221 init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1225 /* Don't support raw QPs */
1226 return ERR_PTR(-EINVAL);
1232 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1233 struct ib_qp_init_attr *init_attr,
1234 struct ib_udata *udata) {
1235 struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1237 struct mlx4_ib_dev *dev = to_mdev(device);
1239 ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1241 if (!IS_ERR(ibqp) &&
1242 (init_attr->qp_type == IB_QPT_GSI) &&
1243 !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1244 struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1245 int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1248 dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1249 init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1250 sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1252 if (IS_ERR(sqp->roce_v2_gsi)) {
1253 pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1254 sqp->roce_v2_gsi = NULL;
1256 sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1257 sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1260 init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1266 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1268 struct mlx4_ib_dev *dev = to_mdev(qp->device);
1269 struct mlx4_ib_qp *mqp = to_mqp(qp);
1270 struct mlx4_ib_pd *pd;
1272 if (is_qp0(dev, mqp))
1273 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1275 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1276 dev->qp1_proxy[mqp->port - 1] == mqp) {
1277 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1278 dev->qp1_proxy[mqp->port - 1] = NULL;
1279 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1282 if (mqp->counter_index)
1283 mlx4_ib_free_qp_counter(dev, mqp);
1286 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1288 if (is_sqp(dev, mqp))
1289 kfree(to_msqp(mqp));
1296 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1298 struct mlx4_ib_qp *mqp = to_mqp(qp);
1300 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1301 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1303 if (sqp->roce_v2_gsi)
1304 ib_destroy_qp(sqp->roce_v2_gsi);
1307 return _mlx4_ib_destroy_qp(qp);
1310 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1313 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC;
1314 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC;
1315 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD;
1316 case MLX4_IB_QPT_XRC_INI:
1317 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC;
1318 case MLX4_IB_QPT_SMI:
1319 case MLX4_IB_QPT_GSI:
1320 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX;
1322 case MLX4_IB_QPT_PROXY_SMI_OWNER:
1323 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1324 MLX4_QP_ST_MLX : -1);
1325 case MLX4_IB_QPT_PROXY_SMI:
1326 case MLX4_IB_QPT_TUN_SMI:
1327 case MLX4_IB_QPT_PROXY_GSI:
1328 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ?
1329 MLX4_QP_ST_UD : -1);
1334 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1339 u32 hw_access_flags = 0;
1341 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1342 dest_rd_atomic = attr->max_dest_rd_atomic;
1344 dest_rd_atomic = qp->resp_depth;
1346 if (attr_mask & IB_QP_ACCESS_FLAGS)
1347 access_flags = attr->qp_access_flags;
1349 access_flags = qp->atomic_rd_en;
1351 if (!dest_rd_atomic)
1352 access_flags &= IB_ACCESS_REMOTE_WRITE;
1354 if (access_flags & IB_ACCESS_REMOTE_READ)
1355 hw_access_flags |= MLX4_QP_BIT_RRE;
1356 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1357 hw_access_flags |= MLX4_QP_BIT_RAE;
1358 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1359 hw_access_flags |= MLX4_QP_BIT_RWE;
1361 return cpu_to_be32(hw_access_flags);
1364 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1367 if (attr_mask & IB_QP_PKEY_INDEX)
1368 sqp->pkey_index = attr->pkey_index;
1369 if (attr_mask & IB_QP_QKEY)
1370 sqp->qkey = attr->qkey;
1371 if (attr_mask & IB_QP_SQ_PSN)
1372 sqp->send_psn = attr->sq_psn;
1375 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1377 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1380 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1381 const struct rdma_ah_attr *ah,
1382 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1383 struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1389 path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1390 path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1391 if (rdma_ah_get_static_rate(ah)) {
1392 path->static_rate = rdma_ah_get_static_rate(ah) +
1393 MLX4_STAT_RATE_OFFSET;
1394 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1395 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1396 --path->static_rate;
1398 path->static_rate = 0;
1400 if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1401 const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1402 int real_sgid_index =
1403 mlx4_ib_gid_index_to_real_index(dev, port,
1406 if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1407 pr_err("sgid_index (%u) too large. max is %d\n",
1408 real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1412 path->grh_mylmc |= 1 << 7;
1413 path->mgid_index = real_sgid_index;
1414 path->hop_limit = grh->hop_limit;
1415 path->tclass_flowlabel =
1416 cpu_to_be32((grh->traffic_class << 20) |
1418 memcpy(path->rgid, grh->dgid.raw, 16);
1421 if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1422 if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1425 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1426 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1428 path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1429 if (vlan_tag < 0x1000) {
1430 if (smac_info->vid < 0x1000) {
1431 /* both valid vlan ids */
1432 if (smac_info->vid != vlan_tag) {
1433 /* different VIDs. unreg old and reg new */
1434 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1437 smac_info->candidate_vid = vlan_tag;
1438 smac_info->candidate_vlan_index = vidx;
1439 smac_info->candidate_vlan_port = port;
1440 smac_info->update_vid = 1;
1441 path->vlan_index = vidx;
1443 path->vlan_index = smac_info->vlan_index;
1446 /* no current vlan tag in qp */
1447 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1450 smac_info->candidate_vid = vlan_tag;
1451 smac_info->candidate_vlan_index = vidx;
1452 smac_info->candidate_vlan_port = port;
1453 smac_info->update_vid = 1;
1454 path->vlan_index = vidx;
1456 path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1459 /* have current vlan tag. unregister it at modify-qp success */
1460 if (smac_info->vid < 0x1000) {
1461 smac_info->candidate_vid = 0xFFFF;
1462 smac_info->update_vid = 1;
1466 /* get smac_index for RoCE use.
1467 * If no smac was yet assigned, register one.
1468 * If one was already assigned, but the new mac differs,
1469 * unregister the old one and register the new one.
1471 if ((!smac_info->smac && !smac_info->smac_port) ||
1472 smac_info->smac != smac) {
1473 /* register candidate now, unreg if needed, after success */
1474 smac_index = mlx4_register_mac(dev->dev, port, smac);
1475 if (smac_index >= 0) {
1476 smac_info->candidate_smac_index = smac_index;
1477 smac_info->candidate_smac = smac;
1478 smac_info->candidate_smac_port = port;
1483 smac_index = smac_info->smac_index;
1485 memcpy(path->dmac, ah->roce.dmac, 6);
1486 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1487 /* put MAC table smac index for IBoE */
1488 path->grh_mylmc = (u8) (smac_index) | 0x80;
1490 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1491 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1497 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1498 enum ib_qp_attr_mask qp_attr_mask,
1499 struct mlx4_ib_qp *mqp,
1500 struct mlx4_qp_path *path, u8 port,
1501 u16 vlan_id, u8 *smac)
1503 return _mlx4_set_path(dev, &qp->ah_attr,
1504 mlx4_mac_to_u64(smac),
1506 path, &mqp->pri, port);
1509 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1510 const struct ib_qp_attr *qp,
1511 enum ib_qp_attr_mask qp_attr_mask,
1512 struct mlx4_ib_qp *mqp,
1513 struct mlx4_qp_path *path, u8 port)
1515 return _mlx4_set_path(dev, &qp->alt_ah_attr,
1518 path, &mqp->alt, port);
1521 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1523 struct mlx4_ib_gid_entry *ge, *tmp;
1525 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1526 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1528 ge->port = qp->port;
1533 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1534 struct mlx4_ib_qp *qp,
1535 struct mlx4_qp_context *context)
1540 u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1542 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1543 if (!qp->pri.smac && !qp->pri.smac_port) {
1544 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1545 if (smac_index >= 0) {
1546 qp->pri.candidate_smac_index = smac_index;
1547 qp->pri.candidate_smac = u64_mac;
1548 qp->pri.candidate_smac_port = qp->port;
1549 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1557 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1559 struct counter_index *new_counter_index;
1563 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1564 IB_LINK_LAYER_ETHERNET ||
1565 !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1566 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1569 err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1573 new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1574 if (!new_counter_index) {
1575 mlx4_counter_free(dev->dev, tmp_idx);
1579 new_counter_index->index = tmp_idx;
1580 new_counter_index->allocated = 1;
1581 qp->counter_index = new_counter_index;
1583 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1584 list_add_tail(&new_counter_index->list,
1585 &dev->counters_table[qp->port - 1].counters_list);
1586 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1592 MLX4_QPC_ROCE_MODE_1 = 0,
1593 MLX4_QPC_ROCE_MODE_2 = 2,
1594 MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1597 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1600 case IB_GID_TYPE_ROCE:
1601 return MLX4_QPC_ROCE_MODE_1;
1602 case IB_GID_TYPE_ROCE_UDP_ENCAP:
1603 return MLX4_QPC_ROCE_MODE_2;
1605 return MLX4_QPC_ROCE_MODE_UNDEFINED;
1609 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1610 const struct ib_qp_attr *attr, int attr_mask,
1611 enum ib_qp_state cur_state, enum ib_qp_state new_state)
1613 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1614 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1615 struct mlx4_ib_pd *pd;
1616 struct mlx4_ib_cq *send_cq, *recv_cq;
1617 struct mlx4_qp_context *context;
1618 enum mlx4_qp_optpar optpar = 0;
1624 /* APM is not supported under RoCE */
1625 if (attr_mask & IB_QP_ALT_PATH &&
1626 rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1627 IB_LINK_LAYER_ETHERNET)
1630 context = kzalloc(sizeof *context, GFP_KERNEL);
1634 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1635 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1637 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1638 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1640 optpar |= MLX4_QP_OPTPAR_PM_STATE;
1641 switch (attr->path_mig_state) {
1642 case IB_MIG_MIGRATED:
1643 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1646 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1649 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1654 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1655 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1656 else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1657 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1658 else if (ibqp->qp_type == IB_QPT_UD) {
1659 if (qp->flags & MLX4_IB_QP_LSO)
1660 context->mtu_msgmax = (IB_MTU_4096 << 5) |
1661 ilog2(dev->dev->caps.max_gso_sz);
1663 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1664 } else if (attr_mask & IB_QP_PATH_MTU) {
1665 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1666 pr_err("path MTU (%u) is invalid\n",
1670 context->mtu_msgmax = (attr->path_mtu << 5) |
1671 ilog2(dev->dev->caps.max_msg_sz);
1675 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1676 context->rq_size_stride |= qp->rq.wqe_shift - 4;
1679 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1680 context->sq_size_stride |= qp->sq.wqe_shift - 4;
1682 if (new_state == IB_QPS_RESET && qp->counter_index)
1683 mlx4_ib_free_qp_counter(dev, qp);
1685 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1686 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1687 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1688 if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1689 context->param3 |= cpu_to_be32(1 << 30);
1692 if (qp->ibqp.uobject)
1693 context->usr_page = cpu_to_be32(
1694 mlx4_to_hw_uar_index(dev->dev,
1695 to_mucontext(ibqp->uobject->context)->uar.index));
1697 context->usr_page = cpu_to_be32(
1698 mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
1700 if (attr_mask & IB_QP_DEST_QPN)
1701 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1703 if (attr_mask & IB_QP_PORT) {
1704 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1705 !(attr_mask & IB_QP_AV)) {
1706 mlx4_set_sched(&context->pri_path, attr->port_num);
1707 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1711 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1712 err = create_qp_lb_counter(dev, qp);
1717 dev->counters_table[qp->port - 1].default_counter;
1718 if (qp->counter_index)
1719 counter_index = qp->counter_index->index;
1721 if (counter_index != -1) {
1722 context->pri_path.counter_index = counter_index;
1723 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1724 if (qp->counter_index) {
1725 context->pri_path.fl |=
1726 MLX4_FL_ETH_SRC_CHECK_MC_LB;
1727 context->pri_path.vlan_control |=
1728 MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1731 context->pri_path.counter_index =
1732 MLX4_SINK_COUNTER_INDEX(dev->dev);
1734 if (qp->flags & MLX4_IB_QP_NETIF) {
1735 mlx4_ib_steer_qp_reg(dev, qp, 1);
1739 if (ibqp->qp_type == IB_QPT_GSI) {
1740 enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1741 IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1742 u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1744 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1748 if (attr_mask & IB_QP_PKEY_INDEX) {
1749 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1750 context->pri_path.disable_pkey_check = 0x40;
1751 context->pri_path.pkey_index = attr->pkey_index;
1752 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1755 if (attr_mask & IB_QP_AV) {
1756 u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1757 attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1759 struct ib_gid_attr gid_attr = {.gid_type = IB_GID_TYPE_IB};
1764 rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1765 rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
1769 rdma_ah_read_grh(&attr->ah_attr)->sgid_index;
1771 status = ib_get_cached_gid(ibqp->device, port_num,
1772 index, &gid, &gid_attr);
1773 if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1775 if (!status && gid_attr.ndev) {
1776 vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1777 memcpy(smac, gid_attr.ndev->dev_addr, ETH_ALEN);
1778 dev_put(gid_attr.ndev);
1784 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1785 port_num, vlan, smac))
1788 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1789 MLX4_QP_OPTPAR_SCHED_QUEUE);
1792 (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
1793 u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
1795 if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
1799 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1804 if (attr_mask & IB_QP_TIMEOUT) {
1805 context->pri_path.ackto |= attr->timeout << 3;
1806 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1809 if (attr_mask & IB_QP_ALT_PATH) {
1810 if (attr->alt_port_num == 0 ||
1811 attr->alt_port_num > dev->dev->caps.num_ports)
1814 if (attr->alt_pkey_index >=
1815 dev->dev->caps.pkey_table_len[attr->alt_port_num])
1818 if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1820 attr->alt_port_num))
1823 context->alt_path.pkey_index = attr->alt_pkey_index;
1824 context->alt_path.ackto = attr->alt_timeout << 3;
1825 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1829 get_cqs(qp, &send_cq, &recv_cq);
1830 context->pd = cpu_to_be32(pd->pdn);
1831 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1832 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1833 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1835 /* Set "fast registration enabled" for all kernel QPs */
1836 if (!qp->ibqp.uobject)
1837 context->params1 |= cpu_to_be32(1 << 11);
1839 if (attr_mask & IB_QP_RNR_RETRY) {
1840 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1841 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1844 if (attr_mask & IB_QP_RETRY_CNT) {
1845 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1846 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1849 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1850 if (attr->max_rd_atomic)
1852 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1853 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1856 if (attr_mask & IB_QP_SQ_PSN)
1857 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1859 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1860 if (attr->max_dest_rd_atomic)
1862 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1863 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1866 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1867 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1868 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1872 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1874 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1875 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1876 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1878 if (attr_mask & IB_QP_RQ_PSN)
1879 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1881 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1882 if (attr_mask & IB_QP_QKEY) {
1883 if (qp->mlx4_ib_qp_type &
1884 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1885 context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1887 if (mlx4_is_mfunc(dev->dev) &&
1888 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1889 (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1890 MLX4_RESERVED_QKEY_BASE) {
1891 pr_err("Cannot use reserved QKEY"
1892 " 0x%x (range 0xffff0000..0xffffffff"
1893 " is reserved)\n", attr->qkey);
1897 context->qkey = cpu_to_be32(attr->qkey);
1899 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1903 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1905 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1906 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1908 if (cur_state == IB_QPS_INIT &&
1909 new_state == IB_QPS_RTR &&
1910 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1911 ibqp->qp_type == IB_QPT_UD ||
1912 ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1913 context->pri_path.sched_queue = (qp->port - 1) << 6;
1914 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1915 qp->mlx4_ib_qp_type &
1916 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1917 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1918 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1919 context->pri_path.fl = 0x80;
1921 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1922 context->pri_path.fl = 0x80;
1923 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1925 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1926 IB_LINK_LAYER_ETHERNET) {
1927 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1928 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1929 context->pri_path.feup = 1 << 7; /* don't fsm */
1930 /* handle smac_index */
1931 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1932 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1933 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1934 err = handle_eth_ud_smac_index(dev, qp, context);
1939 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1940 dev->qp1_proxy[qp->port - 1] = qp;
1945 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1946 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1947 MLX4_IB_LINK_TYPE_ETH;
1948 if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1949 /* set QP to receive both tunneled & non-tunneled packets */
1950 if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1951 context->srqn = cpu_to_be32(7 << 28);
1955 if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1956 int is_eth = rdma_port_get_link_layer(
1957 &dev->ib_dev, qp->port) ==
1958 IB_LINK_LAYER_ETHERNET;
1960 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1961 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1966 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
1967 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1972 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1973 context->rlkey_roce_mode |= (1 << 4);
1976 * Before passing a kernel QP to the HW, make sure that the
1977 * ownership bits of the send queue are set and the SQ
1978 * headroom is stamped so that the hardware doesn't start
1979 * processing stale work requests.
1981 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1982 struct mlx4_wqe_ctrl_seg *ctrl;
1985 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1986 ctrl = get_send_wqe(qp, i);
1987 ctrl->owner_opcode = cpu_to_be32(1 << 31);
1988 if (qp->sq_max_wqes_per_wr == 1)
1989 ctrl->qpn_vlan.fence_size =
1990 1 << (qp->sq.wqe_shift - 4);
1992 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1996 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1997 to_mlx4_state(new_state), context, optpar,
1998 sqd_event, &qp->mqp);
2002 qp->state = new_state;
2004 if (attr_mask & IB_QP_ACCESS_FLAGS)
2005 qp->atomic_rd_en = attr->qp_access_flags;
2006 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2007 qp->resp_depth = attr->max_dest_rd_atomic;
2008 if (attr_mask & IB_QP_PORT) {
2009 qp->port = attr->port_num;
2010 update_mcg_macs(dev, qp);
2012 if (attr_mask & IB_QP_ALT_PATH)
2013 qp->alt_port = attr->alt_port_num;
2015 if (is_sqp(dev, qp))
2016 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2019 * If we moved QP0 to RTR, bring the IB link up; if we moved
2020 * QP0 to RESET or ERROR, bring the link back down.
2022 if (is_qp0(dev, qp)) {
2023 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2024 if (mlx4_INIT_PORT(dev->dev, qp->port))
2025 pr_warn("INIT_PORT failed for port %d\n",
2028 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2029 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2030 mlx4_CLOSE_PORT(dev->dev, qp->port);
2034 * If we moved a kernel QP to RESET, clean up all old CQ
2035 * entries and reinitialize the QP.
2037 if (new_state == IB_QPS_RESET) {
2038 if (!ibqp->uobject) {
2039 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2040 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2041 if (send_cq != recv_cq)
2042 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2048 qp->sq_next_wqe = 0;
2052 if (qp->flags & MLX4_IB_QP_NETIF)
2053 mlx4_ib_steer_qp_reg(dev, qp, 0);
2055 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2056 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2058 qp->pri.smac_port = 0;
2061 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2064 if (qp->pri.vid < 0x1000) {
2065 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2066 qp->pri.vid = 0xFFFF;
2067 qp->pri.candidate_vid = 0xFFFF;
2068 qp->pri.update_vid = 0;
2071 if (qp->alt.vid < 0x1000) {
2072 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2073 qp->alt.vid = 0xFFFF;
2074 qp->alt.candidate_vid = 0xFFFF;
2075 qp->alt.update_vid = 0;
2079 if (err && qp->counter_index)
2080 mlx4_ib_free_qp_counter(dev, qp);
2081 if (err && steer_qp)
2082 mlx4_ib_steer_qp_reg(dev, qp, 0);
2084 if (qp->pri.candidate_smac ||
2085 (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2087 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2089 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2090 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2091 qp->pri.smac = qp->pri.candidate_smac;
2092 qp->pri.smac_index = qp->pri.candidate_smac_index;
2093 qp->pri.smac_port = qp->pri.candidate_smac_port;
2095 qp->pri.candidate_smac = 0;
2096 qp->pri.candidate_smac_index = 0;
2097 qp->pri.candidate_smac_port = 0;
2099 if (qp->alt.candidate_smac) {
2101 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2104 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2105 qp->alt.smac = qp->alt.candidate_smac;
2106 qp->alt.smac_index = qp->alt.candidate_smac_index;
2107 qp->alt.smac_port = qp->alt.candidate_smac_port;
2109 qp->alt.candidate_smac = 0;
2110 qp->alt.candidate_smac_index = 0;
2111 qp->alt.candidate_smac_port = 0;
2114 if (qp->pri.update_vid) {
2116 if (qp->pri.candidate_vid < 0x1000)
2117 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2118 qp->pri.candidate_vid);
2120 if (qp->pri.vid < 0x1000)
2121 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2123 qp->pri.vid = qp->pri.candidate_vid;
2124 qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2125 qp->pri.vlan_index = qp->pri.candidate_vlan_index;
2127 qp->pri.candidate_vid = 0xFFFF;
2128 qp->pri.update_vid = 0;
2131 if (qp->alt.update_vid) {
2133 if (qp->alt.candidate_vid < 0x1000)
2134 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2135 qp->alt.candidate_vid);
2137 if (qp->alt.vid < 0x1000)
2138 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2140 qp->alt.vid = qp->alt.candidate_vid;
2141 qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2142 qp->alt.vlan_index = qp->alt.candidate_vlan_index;
2144 qp->alt.candidate_vid = 0xFFFF;
2145 qp->alt.update_vid = 0;
2151 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2152 int attr_mask, struct ib_udata *udata)
2154 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2155 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2156 enum ib_qp_state cur_state, new_state;
2159 mutex_lock(&qp->mutex);
2161 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2162 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2164 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2165 ll = IB_LINK_LAYER_UNSPECIFIED;
2167 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2168 ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2171 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2173 pr_debug("qpn 0x%x: invalid attribute mask specified "
2174 "for transition %d to %d. qp_type %d,"
2175 " attr_mask 0x%x\n",
2176 ibqp->qp_num, cur_state, new_state,
2177 ibqp->qp_type, attr_mask);
2181 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2182 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2183 if ((ibqp->qp_type == IB_QPT_RC) ||
2184 (ibqp->qp_type == IB_QPT_UD) ||
2185 (ibqp->qp_type == IB_QPT_UC) ||
2186 (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2187 (ibqp->qp_type == IB_QPT_XRC_INI)) {
2188 attr->port_num = mlx4_ib_bond_next_port(dev);
2191 /* no sense in changing port_num
2192 * when ports are bonded */
2193 attr_mask &= ~IB_QP_PORT;
2197 if ((attr_mask & IB_QP_PORT) &&
2198 (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2199 pr_debug("qpn 0x%x: invalid port number (%d) specified "
2200 "for transition %d to %d. qp_type %d\n",
2201 ibqp->qp_num, attr->port_num, cur_state,
2202 new_state, ibqp->qp_type);
2206 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2207 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2208 IB_LINK_LAYER_ETHERNET))
2211 if (attr_mask & IB_QP_PKEY_INDEX) {
2212 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2213 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2214 pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2215 "for transition %d to %d. qp_type %d\n",
2216 ibqp->qp_num, attr->pkey_index, cur_state,
2217 new_state, ibqp->qp_type);
2222 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2223 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2224 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2225 "Transition %d to %d. qp_type %d\n",
2226 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2227 new_state, ibqp->qp_type);
2231 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2232 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2233 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2234 "Transition %d to %d. qp_type %d\n",
2235 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2236 new_state, ibqp->qp_type);
2240 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2245 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2247 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2251 mutex_unlock(&qp->mutex);
2255 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2256 int attr_mask, struct ib_udata *udata)
2258 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2261 ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2263 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2264 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2267 if (sqp->roce_v2_gsi)
2268 err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2270 pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2276 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2279 for (i = 0; i < dev->caps.num_ports; i++) {
2280 if (qpn == dev->caps.qp0_proxy[i] ||
2281 qpn == dev->caps.qp0_tunnel[i]) {
2282 *qkey = dev->caps.qp0_qkey[i];
2289 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2290 struct ib_ud_wr *wr,
2291 void *wqe, unsigned *mlx_seg_len)
2293 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2294 struct ib_device *ib_dev = &mdev->ib_dev;
2295 struct mlx4_wqe_mlx_seg *mlx = wqe;
2296 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2297 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2305 if (wr->wr.opcode != IB_WR_SEND)
2310 for (i = 0; i < wr->wr.num_sge; ++i)
2311 send_size += wr->wr.sg_list[i].length;
2313 /* for proxy-qp0 sends, need to add in size of tunnel header */
2314 /* for tunnel-qp0 sends, tunnel header is already in s/g list */
2315 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2316 send_size += sizeof (struct mlx4_ib_tunnel_header);
2318 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2320 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2321 sqp->ud_header.lrh.service_level =
2322 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2323 sqp->ud_header.lrh.destination_lid =
2324 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2325 sqp->ud_header.lrh.source_lid =
2326 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2329 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2331 /* force loopback */
2332 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2333 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2335 sqp->ud_header.lrh.virtual_lane = 0;
2336 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2337 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2338 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2339 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2340 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2342 sqp->ud_header.bth.destination_qpn =
2343 cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2345 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2346 if (mlx4_is_master(mdev->dev)) {
2347 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2350 if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2353 sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2354 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2356 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2357 sqp->ud_header.immediate_present = 0;
2359 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2362 * Inline data segments may not cross a 64 byte boundary. If
2363 * our UD header is bigger than the space available up to the
2364 * next 64 byte boundary in the WQE, use two inline data
2365 * segments to hold the UD header.
2367 spc = MLX4_INLINE_ALIGN -
2368 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2369 if (header_size <= spc) {
2370 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2371 memcpy(inl + 1, sqp->header_buf, header_size);
2374 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2375 memcpy(inl + 1, sqp->header_buf, spc);
2377 inl = (void *) (inl + 1) + spc;
2378 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2380 * Need a barrier here to make sure all the data is
2381 * visible before the byte_count field is set.
2382 * Otherwise the HCA prefetcher could grab the 64-byte
2383 * chunk with this inline segment and get a valid (!=
2384 * 0xffffffff) byte count but stale data, and end up
2385 * generating a packet with bad headers.
2387 * The first inline segment's byte_count field doesn't
2388 * need a barrier, because it comes after a
2389 * control/MLX segment and therefore is at an offset
2393 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2398 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2402 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2404 union sl2vl_tbl_to_u64 tmp_vltab;
2409 tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2410 vl = tmp_vltab.sl8[sl >> 1];
2418 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
2419 int index, union ib_gid *gid,
2420 enum ib_gid_type *gid_type)
2422 struct mlx4_ib_iboe *iboe = &ibdev->iboe;
2423 struct mlx4_port_gid_table *port_gid_table;
2424 unsigned long flags;
2426 port_gid_table = &iboe->gids[port_num - 1];
2427 spin_lock_irqsave(&iboe->lock, flags);
2428 memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
2429 *gid_type = port_gid_table->gids[index].gid_type;
2430 spin_unlock_irqrestore(&iboe->lock, flags);
2431 if (!memcmp(gid, &zgid, sizeof(*gid)))
2437 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2438 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2439 void *wqe, unsigned *mlx_seg_len)
2441 struct ib_device *ib_dev = sqp->qp.ibqp.device;
2442 struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
2443 struct mlx4_wqe_mlx_seg *mlx = wqe;
2444 struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2445 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2446 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2456 bool is_vlan = false;
2458 bool is_udp = false;
2462 for (i = 0; i < wr->wr.num_sge; ++i)
2463 send_size += wr->wr.sg_list[i].length;
2465 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2466 is_grh = mlx4_ib_ah_grh_present(ah);
2468 enum ib_gid_type gid_type;
2469 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2470 /* When multi-function is enabled, the ib_core gid
2471 * indexes don't necessarily match the hw ones, so
2472 * we must use our own cache */
2473 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2474 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2475 ah->av.ib.gid_index, &sgid.raw[0]);
2479 err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
2480 ah->av.ib.gid_index,
2483 is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2485 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2495 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2496 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2500 err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2501 ip_version, is_udp, 0, &sqp->ud_header);
2506 sqp->ud_header.lrh.service_level =
2507 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2508 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2509 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2512 if (is_grh || (ip_version == 6)) {
2513 sqp->ud_header.grh.traffic_class =
2514 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2515 sqp->ud_header.grh.flow_label =
2516 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2517 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit;
2519 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2521 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2522 /* When multi-function is enabled, the ib_core gid
2523 * indexes don't necessarily match the hw ones, so
2524 * we must use our own cache
2526 sqp->ud_header.grh.source_gid.global.subnet_prefix =
2527 cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
2528 demux[sqp->qp.port - 1].
2530 sqp->ud_header.grh.source_gid.global.interface_id =
2531 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2532 guid_cache[ah->av.ib.gid_index];
2534 ib_get_cached_gid(ib_dev,
2535 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2536 ah->av.ib.gid_index,
2537 &sqp->ud_header.grh.source_gid, NULL);
2540 memcpy(sqp->ud_header.grh.destination_gid.raw,
2541 ah->av.ib.dgid, 16);
2544 if (ip_version == 4) {
2545 sqp->ud_header.ip4.tos =
2546 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2547 sqp->ud_header.ip4.id = 0;
2548 sqp->ud_header.ip4.frag_off = htons(IP_DF);
2549 sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2551 memcpy(&sqp->ud_header.ip4.saddr,
2553 memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2554 sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2558 sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2559 sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2560 sqp->ud_header.udp.csum = 0;
2563 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2566 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2567 (sqp->ud_header.lrh.destination_lid ==
2568 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2569 (sqp->ud_header.lrh.service_level << 8));
2570 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2571 mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2572 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2575 switch (wr->wr.opcode) {
2577 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2578 sqp->ud_header.immediate_present = 0;
2580 case IB_WR_SEND_WITH_IMM:
2581 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2582 sqp->ud_header.immediate_present = 1;
2583 sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
2590 struct in6_addr in6;
2592 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2594 ether_type = (!is_udp) ? ETH_P_IBOE:
2595 (ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
2597 mlx->sched_prio = cpu_to_be16(pcp);
2599 ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2600 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2601 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2602 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2603 memcpy(&in6, sgid.raw, sizeof(in6));
2606 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2607 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2609 sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2611 sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2612 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2615 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 :
2616 sl_to_vl(to_mdev(ib_dev),
2617 sqp->ud_header.lrh.service_level,
2619 if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
2621 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2622 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2624 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2625 if (!sqp->qp.ibqp.qp_num)
2626 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2628 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2629 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2630 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2631 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2632 sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2633 sqp->qkey : wr->remote_qkey);
2634 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2636 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2639 pr_err("built UD header of size %d:\n", header_size);
2640 for (i = 0; i < header_size / 4; ++i) {
2642 pr_err(" [%02x] ", i * 4);
2644 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2645 if ((i + 1) % 8 == 0)
2652 * Inline data segments may not cross a 64 byte boundary. If
2653 * our UD header is bigger than the space available up to the
2654 * next 64 byte boundary in the WQE, use two inline data
2655 * segments to hold the UD header.
2657 spc = MLX4_INLINE_ALIGN -
2658 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2659 if (header_size <= spc) {
2660 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2661 memcpy(inl + 1, sqp->header_buf, header_size);
2664 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2665 memcpy(inl + 1, sqp->header_buf, spc);
2667 inl = (void *) (inl + 1) + spc;
2668 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2670 * Need a barrier here to make sure all the data is
2671 * visible before the byte_count field is set.
2672 * Otherwise the HCA prefetcher could grab the 64-byte
2673 * chunk with this inline segment and get a valid (!=
2674 * 0xffffffff) byte count but stale data, and end up
2675 * generating a packet with bad headers.
2677 * The first inline segment's byte_count field doesn't
2678 * need a barrier, because it comes after a
2679 * control/MLX segment and therefore is at an offset
2683 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2688 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2692 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2695 struct mlx4_ib_cq *cq;
2697 cur = wq->head - wq->tail;
2698 if (likely(cur + nreq < wq->max_post))
2702 spin_lock(&cq->lock);
2703 cur = wq->head - wq->tail;
2704 spin_unlock(&cq->lock);
2706 return cur + nreq >= wq->max_post;
2709 static __be32 convert_access(int acc)
2711 return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2712 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) |
2713 (acc & IB_ACCESS_REMOTE_WRITE ?
2714 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2715 (acc & IB_ACCESS_REMOTE_READ ?
2716 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) |
2717 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) |
2718 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2721 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2722 struct ib_reg_wr *wr)
2724 struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2726 fseg->flags = convert_access(wr->access);
2727 fseg->mem_key = cpu_to_be32(wr->key);
2728 fseg->buf_list = cpu_to_be64(mr->page_map);
2729 fseg->start_addr = cpu_to_be64(mr->ibmr.iova);
2730 fseg->reg_len = cpu_to_be64(mr->ibmr.length);
2731 fseg->offset = 0; /* XXX -- is this just for ZBVA? */
2732 fseg->page_size = cpu_to_be32(ilog2(mr->ibmr.page_size));
2733 fseg->reserved[0] = 0;
2734 fseg->reserved[1] = 0;
2737 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2739 memset(iseg, 0, sizeof(*iseg));
2740 iseg->mem_key = cpu_to_be32(rkey);
2743 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2744 u64 remote_addr, u32 rkey)
2746 rseg->raddr = cpu_to_be64(remote_addr);
2747 rseg->rkey = cpu_to_be32(rkey);
2751 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2752 struct ib_atomic_wr *wr)
2754 if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2755 aseg->swap_add = cpu_to_be64(wr->swap);
2756 aseg->compare = cpu_to_be64(wr->compare_add);
2757 } else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2758 aseg->swap_add = cpu_to_be64(wr->compare_add);
2759 aseg->compare = cpu_to_be64(wr->compare_add_mask);
2761 aseg->swap_add = cpu_to_be64(wr->compare_add);
2767 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2768 struct ib_atomic_wr *wr)
2770 aseg->swap_add = cpu_to_be64(wr->swap);
2771 aseg->swap_add_mask = cpu_to_be64(wr->swap_mask);
2772 aseg->compare = cpu_to_be64(wr->compare_add);
2773 aseg->compare_mask = cpu_to_be64(wr->compare_add_mask);
2776 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2777 struct ib_ud_wr *wr)
2779 memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2780 dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2781 dseg->qkey = cpu_to_be32(wr->remote_qkey);
2782 dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2783 memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2786 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2787 struct mlx4_wqe_datagram_seg *dseg,
2788 struct ib_ud_wr *wr,
2789 enum mlx4_ib_qp_type qpt)
2791 union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2792 struct mlx4_av sqp_av = {0};
2793 int port = *((u8 *) &av->ib.port_pd) & 0x3;
2795 /* force loopback */
2796 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2797 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2798 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2799 cpu_to_be32(0xf0000000);
2801 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2802 if (qpt == MLX4_IB_QPT_PROXY_GSI)
2803 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2805 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2806 /* Use QKEY from the QP context, which is set by master */
2807 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2810 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2812 struct mlx4_wqe_inline_seg *inl = wqe;
2813 struct mlx4_ib_tunnel_header hdr;
2814 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2818 memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2819 hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2820 hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2821 hdr.qkey = cpu_to_be32(wr->remote_qkey);
2822 memcpy(hdr.mac, ah->av.eth.mac, 6);
2823 hdr.vlan = ah->av.eth.vlan;
2825 spc = MLX4_INLINE_ALIGN -
2826 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2827 if (sizeof (hdr) <= spc) {
2828 memcpy(inl + 1, &hdr, sizeof (hdr));
2830 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2833 memcpy(inl + 1, &hdr, spc);
2835 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2837 inl = (void *) (inl + 1) + spc;
2838 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2840 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2845 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2848 static void set_mlx_icrc_seg(void *dseg)
2851 struct mlx4_wqe_inline_seg *iseg = dseg;
2856 * Need a barrier here before writing the byte_count field to
2857 * make sure that all the data is visible before the
2858 * byte_count field is set. Otherwise, if the segment begins
2859 * a new cacheline, the HCA prefetcher could grab the 64-byte
2860 * chunk and get a valid (!= * 0xffffffff) byte count but
2861 * stale data, and end up sending the wrong data.
2865 iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2868 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2870 dseg->lkey = cpu_to_be32(sg->lkey);
2871 dseg->addr = cpu_to_be64(sg->addr);
2874 * Need a barrier here before writing the byte_count field to
2875 * make sure that all the data is visible before the
2876 * byte_count field is set. Otherwise, if the segment begins
2877 * a new cacheline, the HCA prefetcher could grab the 64-byte
2878 * chunk and get a valid (!= * 0xffffffff) byte count but
2879 * stale data, and end up sending the wrong data.
2883 dseg->byte_count = cpu_to_be32(sg->length);
2886 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2888 dseg->byte_count = cpu_to_be32(sg->length);
2889 dseg->lkey = cpu_to_be32(sg->lkey);
2890 dseg->addr = cpu_to_be64(sg->addr);
2893 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2894 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2895 __be32 *lso_hdr_sz, __be32 *blh)
2897 unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2899 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2900 *blh = cpu_to_be32(1 << 6);
2902 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2903 wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2906 memcpy(wqe->header, wr->header, wr->hlen);
2908 *lso_hdr_sz = cpu_to_be32(wr->mss << 16 | wr->hlen);
2909 *lso_seg_len = halign;
2913 static __be32 send_ieth(struct ib_send_wr *wr)
2915 switch (wr->opcode) {
2916 case IB_WR_SEND_WITH_IMM:
2917 case IB_WR_RDMA_WRITE_WITH_IMM:
2918 return wr->ex.imm_data;
2920 case IB_WR_SEND_WITH_INV:
2921 return cpu_to_be32(wr->ex.invalidate_rkey);
2928 static void add_zero_len_inline(void *wqe)
2930 struct mlx4_wqe_inline_seg *inl = wqe;
2932 inl->byte_count = cpu_to_be32(1 << 31);
2935 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2936 struct ib_send_wr **bad_wr)
2938 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2940 struct mlx4_wqe_ctrl_seg *ctrl;
2941 struct mlx4_wqe_data_seg *dseg;
2942 unsigned long flags;
2946 int uninitialized_var(stamp);
2947 int uninitialized_var(size);
2948 unsigned uninitialized_var(seglen);
2951 __be32 uninitialized_var(lso_hdr_sz);
2954 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2956 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2957 struct mlx4_ib_sqp *sqp = to_msqp(qp);
2959 if (sqp->roce_v2_gsi) {
2960 struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2961 enum ib_gid_type gid_type;
2964 if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
2965 ah->av.ib.gid_index,
2967 qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2968 to_mqp(sqp->roce_v2_gsi) : qp;
2970 pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2971 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->qpn_vlan.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(1 << 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_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
3393 struct rdma_ah_attr *ah_attr,
3394 struct mlx4_qp_path *path)
3396 struct mlx4_dev *dev = ibdev->dev;
3397 u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
3399 memset(ah_attr, 0, sizeof(*ah_attr));
3400 ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
3401 if (port_num == 0 || port_num > dev->caps.num_ports)
3404 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
3405 rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
3406 ((path->sched_queue & 4) << 1));
3408 rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
3409 rdma_ah_set_port_num(ah_attr, port_num);
3411 rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
3412 rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
3413 rdma_ah_set_static_rate(ah_attr,
3414 path->static_rate ? path->static_rate - 5 : 0);
3415 if (path->grh_mylmc & (1 << 7)) {
3416 rdma_ah_set_grh(ah_attr, NULL,
3417 be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
3420 (be32_to_cpu(path->tclass_flowlabel)
3422 rdma_ah_set_dgid_raw(ah_attr, path->rgid);
3426 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3427 struct ib_qp_init_attr *qp_init_attr)
3429 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3430 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3431 struct mlx4_qp_context context;
3435 mutex_lock(&qp->mutex);
3437 if (qp->state == IB_QPS_RESET) {
3438 qp_attr->qp_state = IB_QPS_RESET;
3442 err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3448 mlx4_state = be32_to_cpu(context.flags) >> 28;
3450 qp->state = to_ib_qp_state(mlx4_state);
3451 qp_attr->qp_state = qp->state;
3452 qp_attr->path_mtu = context.mtu_msgmax >> 5;
3453 qp_attr->path_mig_state =
3454 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3455 qp_attr->qkey = be32_to_cpu(context.qkey);
3456 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3457 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff;
3458 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff;
3459 qp_attr->qp_access_flags =
3460 to_ib_qp_access_flags(be32_to_cpu(context.params2));
3462 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3463 to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3464 to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3465 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3466 qp_attr->alt_port_num =
3467 rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
3470 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3471 if (qp_attr->qp_state == IB_QPS_INIT)
3472 qp_attr->port_num = qp->port;
3474 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3476 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3477 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3479 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3481 qp_attr->max_dest_rd_atomic =
3482 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3483 qp_attr->min_rnr_timer =
3484 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3485 qp_attr->timeout = context.pri_path.ackto >> 3;
3486 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7;
3487 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7;
3488 qp_attr->alt_timeout = context.alt_path.ackto >> 3;
3491 qp_attr->cur_qp_state = qp_attr->qp_state;
3492 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
3493 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
3495 if (!ibqp->uobject) {
3496 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
3497 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3499 qp_attr->cap.max_send_wr = 0;
3500 qp_attr->cap.max_send_sge = 0;
3504 * We don't support inline sends for kernel QPs (yet), and we
3505 * don't know what userspace's value should be.
3507 qp_attr->cap.max_inline_data = 0;
3509 qp_init_attr->cap = qp_attr->cap;
3511 qp_init_attr->create_flags = 0;
3512 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3513 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3515 if (qp->flags & MLX4_IB_QP_LSO)
3516 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3518 if (qp->flags & MLX4_IB_QP_NETIF)
3519 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3521 qp_init_attr->sq_sig_type =
3522 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3523 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3526 mutex_unlock(&qp->mutex);