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>
40 #include <rdma/ib_cache.h>
41 #include <rdma/ib_pack.h>
42 #include <rdma/ib_addr.h>
43 #include <rdma/ib_mad.h>
45 #include <linux/mlx4/driver.h>
46 #include <linux/mlx4/qp.h>
49 #include <rdma/mlx4-abi.h>
51 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
52 struct mlx4_ib_cq *recv_cq);
53 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
54 struct mlx4_ib_cq *recv_cq);
55 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state);
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 enum mlx4_ib_source_type {
124 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
126 return container_of(mqp, struct mlx4_ib_sqp, qp);
129 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
131 if (!mlx4_is_master(dev->dev))
134 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
135 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
139 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
144 /* PPF or Native -- real SQP */
145 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
146 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
147 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
150 /* VF or PF -- proxy SQP */
151 if (mlx4_is_mfunc(dev->dev)) {
152 for (i = 0; i < dev->dev->caps.num_ports; i++) {
153 if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy ||
154 qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) {
163 return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
166 /* used for INIT/CLOSE port logic */
167 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
172 /* PPF or Native -- real QP0 */
173 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
174 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
175 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
178 /* VF or PF -- proxy QP0 */
179 if (mlx4_is_mfunc(dev->dev)) {
180 for (i = 0; i < dev->dev->caps.num_ports; i++) {
181 if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) {
190 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
192 return mlx4_buf_offset(&qp->buf, offset);
195 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
197 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
200 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
202 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
206 * Stamp a SQ WQE so that it is invalid if prefetched by marking the
207 * first four bytes of every 64 byte chunk with
208 * 0x7FFFFFF | (invalid_ownership_value << 31).
210 * When the max work request size is less than or equal to the WQE
211 * basic block size, as an optimization, we can stamp all WQEs with
212 * 0xffffffff, and skip the very first chunk of each WQE.
214 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
222 struct mlx4_wqe_ctrl_seg *ctrl;
224 if (qp->sq_max_wqes_per_wr > 1) {
225 s = roundup(size, 1U << qp->sq.wqe_shift);
226 for (i = 0; i < s; i += 64) {
227 ind = (i >> qp->sq.wqe_shift) + n;
228 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
229 cpu_to_be32(0xffffffff);
230 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
231 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
235 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
236 s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4;
237 for (i = 64; i < s; i += 64) {
239 *wqe = cpu_to_be32(0xffffffff);
244 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
246 struct mlx4_wqe_ctrl_seg *ctrl;
247 struct mlx4_wqe_inline_seg *inl;
251 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
252 s = sizeof(struct mlx4_wqe_ctrl_seg);
254 if (qp->ibqp.qp_type == IB_QPT_UD) {
255 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
256 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
257 memset(dgram, 0, sizeof *dgram);
258 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
259 s += sizeof(struct mlx4_wqe_datagram_seg);
262 /* Pad the remainder of the WQE with an inline data segment. */
265 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
267 ctrl->srcrb_flags = 0;
268 ctrl->qpn_vlan.fence_size = size / 16;
270 * Make sure descriptor is fully written before setting ownership bit
271 * (because HW can start executing as soon as we do).
275 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
276 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
278 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
281 /* Post NOP WQE to prevent wrap-around in the middle of WR */
282 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
284 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
285 if (unlikely(s < qp->sq_max_wqes_per_wr)) {
286 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
292 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
294 struct ib_event event;
295 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
297 if (type == MLX4_EVENT_TYPE_PATH_MIG)
298 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
300 if (ibqp->event_handler) {
301 event.device = ibqp->device;
302 event.element.qp = ibqp;
304 case MLX4_EVENT_TYPE_PATH_MIG:
305 event.event = IB_EVENT_PATH_MIG;
307 case MLX4_EVENT_TYPE_COMM_EST:
308 event.event = IB_EVENT_COMM_EST;
310 case MLX4_EVENT_TYPE_SQ_DRAINED:
311 event.event = IB_EVENT_SQ_DRAINED;
313 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
314 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
316 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
317 event.event = IB_EVENT_QP_FATAL;
319 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
320 event.event = IB_EVENT_PATH_MIG_ERR;
322 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
323 event.event = IB_EVENT_QP_REQ_ERR;
325 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
326 event.event = IB_EVENT_QP_ACCESS_ERR;
329 pr_warn("Unexpected event type %d "
330 "on QP %06x\n", type, qp->qpn);
334 ibqp->event_handler(&event, ibqp->qp_context);
338 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type)
340 pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n",
344 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
347 * UD WQEs must have a datagram segment.
348 * RC and UC WQEs might have a remote address segment.
349 * MLX WQEs need two extra inline data segments (for the UD
350 * header and space for the ICRC).
354 return sizeof (struct mlx4_wqe_ctrl_seg) +
355 sizeof (struct mlx4_wqe_datagram_seg) +
356 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
357 case MLX4_IB_QPT_PROXY_SMI_OWNER:
358 case MLX4_IB_QPT_PROXY_SMI:
359 case MLX4_IB_QPT_PROXY_GSI:
360 return sizeof (struct mlx4_wqe_ctrl_seg) +
361 sizeof (struct mlx4_wqe_datagram_seg) + 64;
362 case MLX4_IB_QPT_TUN_SMI_OWNER:
363 case MLX4_IB_QPT_TUN_GSI:
364 return sizeof (struct mlx4_wqe_ctrl_seg) +
365 sizeof (struct mlx4_wqe_datagram_seg);
368 return sizeof (struct mlx4_wqe_ctrl_seg) +
369 sizeof (struct mlx4_wqe_raddr_seg);
371 return sizeof (struct mlx4_wqe_ctrl_seg) +
372 sizeof (struct mlx4_wqe_masked_atomic_seg) +
373 sizeof (struct mlx4_wqe_raddr_seg);
374 case MLX4_IB_QPT_SMI:
375 case MLX4_IB_QPT_GSI:
376 return sizeof (struct mlx4_wqe_ctrl_seg) +
377 ALIGN(MLX4_IB_UD_HEADER_SIZE +
378 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
380 sizeof (struct mlx4_wqe_inline_seg),
381 sizeof (struct mlx4_wqe_data_seg)) +
383 sizeof (struct mlx4_wqe_inline_seg),
384 sizeof (struct mlx4_wqe_data_seg));
386 return sizeof (struct mlx4_wqe_ctrl_seg);
390 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
391 int is_user, int has_rq, struct mlx4_ib_qp *qp,
394 /* Sanity check RQ size before proceeding */
395 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
396 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
400 if (cap->max_recv_wr || inl_recv_sz)
403 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
405 u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg *
406 sizeof(struct mlx4_wqe_data_seg);
409 /* HW requires >= 1 RQ entry with >= 1 gather entry */
410 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge ||
411 inl_recv_sz > max_inl_recv_sz))
414 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr));
415 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
416 wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg);
417 qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz));
420 /* leave userspace return values as they were, so as not to break ABI */
422 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt;
423 cap->max_recv_sge = qp->rq.max_gs;
425 cap->max_recv_wr = qp->rq.max_post =
426 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
427 cap->max_recv_sge = min(qp->rq.max_gs,
428 min(dev->dev->caps.max_sq_sg,
429 dev->dev->caps.max_rq_sg));
435 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
436 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
441 /* Sanity check SQ size before proceeding */
442 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
443 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
444 cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
445 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
449 * For MLX transport we need 2 extra S/G entries:
450 * one for the header and one for the checksum at the end
452 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
453 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
454 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
457 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
458 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
459 send_wqe_overhead(type, qp->flags);
461 if (s > dev->dev->caps.max_sq_desc_sz)
465 * Hermon supports shrinking WQEs, such that a single work
466 * request can include multiple units of 1 << wqe_shift. This
467 * way, work requests can differ in size, and do not have to
468 * be a power of 2 in size, saving memory and speeding up send
469 * WR posting. Unfortunately, if we do this then the
470 * wqe_index field in CQEs can't be used to look up the WR ID
471 * anymore, so we do this only if selective signaling is off.
473 * Further, on 32-bit platforms, we can't use vmap() to make
474 * the QP buffer virtually contiguous. Thus we have to use
475 * constant-sized WRs to make sure a WR is always fully within
476 * a single page-sized chunk.
478 * Finally, we use NOP work requests to pad the end of the
479 * work queue, to avoid wrap-around in the middle of WR. We
480 * set NEC bit to avoid getting completions with error for
481 * these NOP WRs, but since NEC is only supported starting
482 * with firmware 2.2.232, we use constant-sized WRs for older
485 * And, since MLX QPs only support SEND, we use constant-sized
488 * We look for the smallest value of wqe_shift such that the
489 * resulting number of wqes does not exceed device
492 * We set WQE size to at least 64 bytes, this way stamping
493 * invalidates each WQE.
495 if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
496 qp->sq_signal_bits && BITS_PER_LONG == 64 &&
497 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
498 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
499 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
500 qp->sq.wqe_shift = ilog2(64);
502 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
505 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
508 * We need to leave 2 KB + 1 WR of headroom in the SQ to
509 * allow HW to prefetch.
511 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
512 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
513 qp->sq_max_wqes_per_wr +
516 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
519 if (qp->sq_max_wqes_per_wr <= 1)
525 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
526 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
527 send_wqe_overhead(type, qp->flags)) /
528 sizeof (struct mlx4_wqe_data_seg);
530 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
531 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
532 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
534 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
536 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
540 cap->max_send_wr = qp->sq.max_post =
541 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
542 cap->max_send_sge = min(qp->sq.max_gs,
543 min(dev->dev->caps.max_sq_sg,
544 dev->dev->caps.max_rq_sg));
545 /* We don't support inline sends for kernel QPs (yet) */
546 cap->max_inline_data = 0;
551 static int set_user_sq_size(struct mlx4_ib_dev *dev,
552 struct mlx4_ib_qp *qp,
553 struct mlx4_ib_create_qp *ucmd)
555 /* Sanity check SQ size before proceeding */
556 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes ||
557 ucmd->log_sq_stride >
558 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
559 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
562 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
563 qp->sq.wqe_shift = ucmd->log_sq_stride;
565 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
566 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
571 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
576 kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf),
578 if (!qp->sqp_proxy_rcv)
580 for (i = 0; i < qp->rq.wqe_cnt; i++) {
581 qp->sqp_proxy_rcv[i].addr =
582 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
584 if (!qp->sqp_proxy_rcv[i].addr)
586 qp->sqp_proxy_rcv[i].map =
587 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
588 sizeof (struct mlx4_ib_proxy_sqp_hdr),
590 if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
591 kfree(qp->sqp_proxy_rcv[i].addr);
600 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
601 sizeof (struct mlx4_ib_proxy_sqp_hdr),
603 kfree(qp->sqp_proxy_rcv[i].addr);
605 kfree(qp->sqp_proxy_rcv);
606 qp->sqp_proxy_rcv = NULL;
610 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
614 for (i = 0; i < qp->rq.wqe_cnt; i++) {
615 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
616 sizeof (struct mlx4_ib_proxy_sqp_hdr),
618 kfree(qp->sqp_proxy_rcv[i].addr);
620 kfree(qp->sqp_proxy_rcv);
623 static int qp_has_rq(struct ib_qp_init_attr *attr)
625 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
631 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
634 for (i = 0; i < dev->caps.num_ports; i++) {
635 if (qpn == dev->caps.spec_qps[i].qp0_proxy)
636 return !!dev->caps.spec_qps[i].qp0_qkey;
641 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
642 struct mlx4_ib_qp *qp)
644 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
645 mlx4_counter_free(dev->dev, qp->counter_index->index);
646 list_del(&qp->counter_index->list);
647 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
649 kfree(qp->counter_index);
650 qp->counter_index = NULL;
653 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx,
654 struct ib_qp_init_attr *init_attr,
655 struct mlx4_ib_create_qp_rss *ucmd)
657 rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num |
658 (init_attr->rwq_ind_tbl->log_ind_tbl_size << 24);
660 if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) &&
661 (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) {
662 memcpy(rss_ctx->rss_key, ucmd->rx_hash_key,
663 MLX4_EN_RSS_KEY_SIZE);
665 pr_debug("RX Hash function is not supported\n");
666 return (-EOPNOTSUPP);
669 if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4 |
670 MLX4_IB_RX_HASH_DST_IPV4 |
671 MLX4_IB_RX_HASH_SRC_IPV6 |
672 MLX4_IB_RX_HASH_DST_IPV6 |
673 MLX4_IB_RX_HASH_SRC_PORT_TCP |
674 MLX4_IB_RX_HASH_DST_PORT_TCP |
675 MLX4_IB_RX_HASH_SRC_PORT_UDP |
676 MLX4_IB_RX_HASH_DST_PORT_UDP |
677 MLX4_IB_RX_HASH_INNER)) {
678 pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n",
679 ucmd->rx_hash_fields_mask);
680 return (-EOPNOTSUPP);
683 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) &&
684 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
685 rss_ctx->flags = MLX4_RSS_IPV4;
686 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) ||
687 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) {
688 pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n");
689 return (-EOPNOTSUPP);
692 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) &&
693 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
694 rss_ctx->flags |= MLX4_RSS_IPV6;
695 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) ||
696 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) {
697 pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n");
698 return (-EOPNOTSUPP);
701 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) &&
702 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
703 if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) {
704 pr_debug("RX Hash fields_mask for UDP is not supported\n");
705 return (-EOPNOTSUPP);
708 if (rss_ctx->flags & MLX4_RSS_IPV4)
709 rss_ctx->flags |= MLX4_RSS_UDP_IPV4;
710 if (rss_ctx->flags & MLX4_RSS_IPV6)
711 rss_ctx->flags |= MLX4_RSS_UDP_IPV6;
712 if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
713 pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n");
714 return (-EOPNOTSUPP);
716 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) ||
717 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) {
718 pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n");
719 return (-EOPNOTSUPP);
722 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) &&
723 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
724 if (rss_ctx->flags & MLX4_RSS_IPV4)
725 rss_ctx->flags |= MLX4_RSS_TCP_IPV4;
726 if (rss_ctx->flags & MLX4_RSS_IPV6)
727 rss_ctx->flags |= MLX4_RSS_TCP_IPV6;
728 if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) {
729 pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n");
730 return (-EOPNOTSUPP);
732 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) ||
733 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) {
734 pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n");
735 return (-EOPNOTSUPP);
738 if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) {
739 if (dev->dev->caps.tunnel_offload_mode ==
740 MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
742 * Hash according to inner headers if exist, otherwise
743 * according to outer headers.
745 rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY;
747 pr_debug("RSS Hash for inner headers isn't supported\n");
748 return (-EOPNOTSUPP);
755 static int create_qp_rss(struct mlx4_ib_dev *dev,
756 struct ib_qp_init_attr *init_attr,
757 struct mlx4_ib_create_qp_rss *ucmd,
758 struct mlx4_ib_qp *qp)
763 qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
765 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage);
769 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
773 mutex_init(&qp->mutex);
775 INIT_LIST_HEAD(&qp->gid_list);
776 INIT_LIST_HEAD(&qp->steering_rules);
778 qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET;
779 qp->state = IB_QPS_RESET;
781 /* Set dummy send resources to be compatible with HV and PRM */
782 qp->sq_no_prefetch = 1;
784 qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
785 qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE;
787 (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt;
789 qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL);
795 err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd);
805 mlx4_qp_remove(dev->dev, &qp->mqp);
806 mlx4_qp_free(dev->dev, &qp->mqp);
809 mlx4_qp_release_range(dev->dev, qpn, 1);
813 static struct ib_qp *_mlx4_ib_create_qp_rss(struct ib_pd *pd,
814 struct ib_qp_init_attr *init_attr,
815 struct ib_udata *udata)
817 struct mlx4_ib_qp *qp;
818 struct mlx4_ib_create_qp_rss ucmd = {};
819 size_t required_cmd_sz;
823 pr_debug("RSS QP with NULL udata\n");
824 return ERR_PTR(-EINVAL);
828 return ERR_PTR(-EOPNOTSUPP);
830 required_cmd_sz = offsetof(typeof(ucmd), reserved1) +
831 sizeof(ucmd.reserved1);
832 if (udata->inlen < required_cmd_sz) {
833 pr_debug("invalid inlen\n");
834 return ERR_PTR(-EINVAL);
837 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
838 pr_debug("copy failed\n");
839 return ERR_PTR(-EFAULT);
842 if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)))
843 return ERR_PTR(-EOPNOTSUPP);
845 if (ucmd.comp_mask || ucmd.reserved1)
846 return ERR_PTR(-EOPNOTSUPP);
848 if (udata->inlen > sizeof(ucmd) &&
849 !ib_is_udata_cleared(udata, sizeof(ucmd),
850 udata->inlen - sizeof(ucmd))) {
851 pr_debug("inlen is not supported\n");
852 return ERR_PTR(-EOPNOTSUPP);
855 if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
856 pr_debug("RSS QP with unsupported QP type %d\n",
858 return ERR_PTR(-EOPNOTSUPP);
861 if (init_attr->create_flags) {
862 pr_debug("RSS QP doesn't support create flags\n");
863 return ERR_PTR(-EOPNOTSUPP);
866 if (init_attr->send_cq || init_attr->cap.max_send_wr) {
867 pr_debug("RSS QP with unsupported send attributes\n");
868 return ERR_PTR(-EOPNOTSUPP);
871 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
873 return ERR_PTR(-ENOMEM);
875 qp->pri.vid = 0xFFFF;
876 qp->alt.vid = 0xFFFF;
878 err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp);
884 qp->ibqp.qp_num = qp->mqp.qpn;
890 * This function allocates a WQN from a range which is consecutive and aligned
891 * to its size. In case the range is full, then it creates a new range and
892 * allocates WQN from it. The new range will be used for following allocations.
894 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context,
895 struct mlx4_ib_qp *qp, int range_size, int *wqn)
897 struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
898 struct mlx4_wqn_range *range;
901 mutex_lock(&context->wqn_ranges_mutex);
903 range = list_first_entry_or_null(&context->wqn_ranges_list,
904 struct mlx4_wqn_range, list);
906 if (!range || (range->refcount == range->size) || range->dirty) {
907 range = kzalloc(sizeof(*range), GFP_KERNEL);
913 err = mlx4_qp_reserve_range(dev->dev, range_size,
914 range_size, &range->base_wqn, 0,
921 range->size = range_size;
922 list_add(&range->list, &context->wqn_ranges_list);
923 } else if (range_size != 1) {
925 * Requesting a new range (>1) when last range is still open, is
932 qp->wqn_range = range;
934 *wqn = range->base_wqn + range->refcount;
939 mutex_unlock(&context->wqn_ranges_mutex);
944 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context,
945 struct mlx4_ib_qp *qp, bool dirty_release)
947 struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device);
948 struct mlx4_wqn_range *range;
950 mutex_lock(&context->wqn_ranges_mutex);
952 range = qp->wqn_range;
955 if (!range->refcount) {
956 mlx4_qp_release_range(dev->dev, range->base_wqn,
958 list_del(&range->list);
960 } else if (dirty_release) {
962 * A range which one of its WQNs is destroyed, won't be able to be
963 * reused for further WQN allocations.
964 * The next created WQ will allocate a new range.
969 mutex_unlock(&context->wqn_ranges_mutex);
972 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
973 enum mlx4_ib_source_type src,
974 struct ib_qp_init_attr *init_attr,
975 struct ib_udata *udata, int sqpn,
976 struct mlx4_ib_qp **caller_qp)
980 struct ib_qp_cap backup_cap;
981 struct mlx4_ib_sqp *sqp = NULL;
982 struct mlx4_ib_qp *qp;
983 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
984 struct mlx4_ib_cq *mcq;
988 /* When tunneling special qps, we use a plain UD qp */
990 if (mlx4_is_mfunc(dev->dev) &&
991 (!mlx4_is_master(dev->dev) ||
992 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
993 if (init_attr->qp_type == IB_QPT_GSI)
994 qp_type = MLX4_IB_QPT_PROXY_GSI;
996 if (mlx4_is_master(dev->dev) ||
997 qp0_enabled_vf(dev->dev, sqpn))
998 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
1000 qp_type = MLX4_IB_QPT_PROXY_SMI;
1004 /* add extra sg entry for tunneling */
1005 init_attr->cap.max_recv_sge++;
1006 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
1007 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
1008 container_of(init_attr,
1009 struct mlx4_ib_qp_tunnel_init_attr, init_attr);
1010 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
1011 tnl_init->proxy_qp_type != IB_QPT_GSI) ||
1012 !mlx4_is_master(dev->dev))
1014 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
1015 qp_type = MLX4_IB_QPT_TUN_GSI;
1016 else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
1017 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
1019 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
1021 qp_type = MLX4_IB_QPT_TUN_SMI;
1022 /* we are definitely in the PPF here, since we are creating
1023 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
1024 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
1025 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
1030 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
1031 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
1032 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
1033 sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL);
1037 qp->pri.vid = 0xFFFF;
1038 qp->alt.vid = 0xFFFF;
1040 qp = kzalloc(sizeof(struct mlx4_ib_qp), GFP_KERNEL);
1043 qp->pri.vid = 0xFFFF;
1044 qp->alt.vid = 0xFFFF;
1049 qp->mlx4_ib_qp_type = qp_type;
1051 mutex_init(&qp->mutex);
1052 spin_lock_init(&qp->sq.lock);
1053 spin_lock_init(&qp->rq.lock);
1054 INIT_LIST_HEAD(&qp->gid_list);
1055 INIT_LIST_HEAD(&qp->steering_rules);
1057 qp->state = IB_QPS_RESET;
1058 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1059 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1064 struct mlx4_ib_create_qp qp;
1065 struct mlx4_ib_create_wq wq;
1071 copy_len = (src == MLX4_IB_QP_SRC) ?
1072 sizeof(struct mlx4_ib_create_qp) :
1073 min(sizeof(struct mlx4_ib_create_wq), udata->inlen);
1075 if (ib_copy_from_udata(&ucmd, udata, copy_len)) {
1080 if (src == MLX4_IB_RWQ_SRC) {
1081 if (ucmd.wq.comp_mask || ucmd.wq.reserved[0] ||
1082 ucmd.wq.reserved[1] || ucmd.wq.reserved[2]) {
1083 pr_debug("user command isn't supported\n");
1088 if (ucmd.wq.log_range_size >
1089 ilog2(dev->dev->caps.max_rss_tbl_sz)) {
1090 pr_debug("WQN range size must be equal or smaller than %d\n",
1091 dev->dev->caps.max_rss_tbl_sz);
1095 range_size = 1 << ucmd.wq.log_range_size;
1097 qp->inl_recv_sz = ucmd.qp.inl_recv_sz;
1100 if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1101 if (!(dev->dev->caps.flags &
1102 MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
1103 pr_debug("scatter FCS is unsupported\n");
1108 qp->flags |= MLX4_IB_QP_SCATTER_FCS;
1111 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1112 qp_has_rq(init_attr), qp, qp->inl_recv_sz);
1116 if (src == MLX4_IB_QP_SRC) {
1117 qp->sq_no_prefetch = ucmd.qp.sq_no_prefetch;
1119 err = set_user_sq_size(dev, qp,
1120 (struct mlx4_ib_create_qp *)
1125 qp->sq_no_prefetch = 1;
1127 qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE;
1128 /* Allocated buffer expects to have at least that SQ
1131 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
1132 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
1135 qp->umem = ib_umem_get(pd->uobject->context,
1136 (src == MLX4_IB_QP_SRC) ? ucmd.qp.buf_addr :
1137 ucmd.wq.buf_addr, qp->buf_size, 0, 0);
1138 if (IS_ERR(qp->umem)) {
1139 err = PTR_ERR(qp->umem);
1143 n = ib_umem_page_count(qp->umem);
1144 shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
1145 err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
1150 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
1154 if (qp_has_rq(init_attr)) {
1155 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
1156 (src == MLX4_IB_QP_SRC) ? ucmd.qp.db_addr :
1157 ucmd.wq.db_addr, &qp->db);
1161 qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS;
1163 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject,
1164 qp_has_rq(init_attr), qp, 0);
1168 qp->sq_no_prefetch = 0;
1170 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1171 qp->flags |= MLX4_IB_QP_LSO;
1173 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1174 if (dev->steering_support ==
1175 MLX4_STEERING_MODE_DEVICE_MANAGED)
1176 qp->flags |= MLX4_IB_QP_NETIF;
1181 memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
1182 err = set_kernel_sq_size(dev, &init_attr->cap,
1187 if (qp_has_rq(init_attr)) {
1188 err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1195 if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
1197 memcpy(&init_attr->cap, &backup_cap,
1198 sizeof(backup_cap));
1199 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
1204 if (mlx4_buf_alloc(dev->dev, qp->buf_size,
1205 PAGE_SIZE * 2, &qp->buf)) {
1211 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1216 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1220 qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt,
1221 sizeof(u64), GFP_KERNEL);
1222 qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt,
1223 sizeof(u64), GFP_KERNEL);
1224 if (!qp->sq.wrid || !qp->rq.wrid) {
1228 qp->mqp.usage = MLX4_RES_USAGE_DRIVER;
1232 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1233 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1234 if (alloc_proxy_bufs(pd->device, qp)) {
1239 } else if (src == MLX4_IB_RWQ_SRC) {
1240 err = mlx4_ib_alloc_wqn(to_mucontext(pd->uobject->context), qp,
1245 /* Raw packet QPNs may not have bits 6,7 set in their qp_num;
1246 * otherwise, the WQE BlueFlame setup flow wrongly causes
1247 * VLAN insertion. */
1248 if (init_attr->qp_type == IB_QPT_RAW_PACKET)
1249 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
1250 (init_attr->cap.max_send_wr ?
1251 MLX4_RESERVE_ETH_BF_QP : 0) |
1252 (init_attr->cap.max_recv_wr ?
1253 MLX4_RESERVE_A0_QP : 0),
1256 if (qp->flags & MLX4_IB_QP_NETIF)
1257 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
1259 err = mlx4_qp_reserve_range(dev->dev, 1, 1,
1260 &qpn, 0, qp->mqp.usage);
1265 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1266 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1268 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1272 if (init_attr->qp_type == IB_QPT_XRC_TGT)
1273 qp->mqp.qpn |= (1 << 23);
1276 * Hardware wants QPN written in big-endian order (after
1277 * shifting) for send doorbell. Precompute this value to save
1278 * a little bit when posting sends.
1280 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1282 qp->mqp.event = (src == MLX4_IB_QP_SRC) ? mlx4_ib_qp_event :
1288 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1289 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
1290 to_mcq(init_attr->recv_cq));
1291 /* Maintain device to QPs access, needed for further handling
1294 list_add_tail(&qp->qps_list, &dev->qp_list);
1295 /* Maintain CQ to QPs access, needed for further handling
1298 mcq = to_mcq(init_attr->send_cq);
1299 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
1300 mcq = to_mcq(init_attr->recv_cq);
1301 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
1302 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
1303 to_mcq(init_attr->recv_cq));
1304 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1309 if (qp->flags & MLX4_IB_QP_NETIF)
1310 mlx4_ib_steer_qp_free(dev, qpn, 1);
1311 else if (src == MLX4_IB_RWQ_SRC)
1312 mlx4_ib_release_wqn(to_mucontext(pd->uobject->context),
1315 mlx4_qp_release_range(dev->dev, qpn, 1);
1318 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1319 free_proxy_bufs(pd->device, qp);
1322 if (qp_has_rq(init_attr))
1323 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1325 kvfree(qp->sq.wrid);
1326 kvfree(qp->rq.wrid);
1330 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1334 ib_umem_release(qp->umem);
1336 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1339 if (!pd->uobject && qp_has_rq(init_attr))
1340 mlx4_db_free(dev->dev, &qp->db);
1345 else if (!*caller_qp)
1350 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1353 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
1354 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
1355 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
1356 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
1357 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
1358 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
1359 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
1364 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1365 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1367 if (send_cq == recv_cq) {
1368 spin_lock(&send_cq->lock);
1369 __acquire(&recv_cq->lock);
1370 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1371 spin_lock(&send_cq->lock);
1372 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1374 spin_lock(&recv_cq->lock);
1375 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1379 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1380 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1382 if (send_cq == recv_cq) {
1383 __release(&recv_cq->lock);
1384 spin_unlock(&send_cq->lock);
1385 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1386 spin_unlock(&recv_cq->lock);
1387 spin_unlock(&send_cq->lock);
1389 spin_unlock(&send_cq->lock);
1390 spin_unlock(&recv_cq->lock);
1394 static void del_gid_entries(struct mlx4_ib_qp *qp)
1396 struct mlx4_ib_gid_entry *ge, *tmp;
1398 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1399 list_del(&ge->list);
1404 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1406 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1407 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1409 return to_mpd(qp->ibqp.pd);
1412 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src,
1413 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1415 switch (qp->ibqp.qp_type) {
1416 case IB_QPT_XRC_TGT:
1417 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1418 *recv_cq = *send_cq;
1420 case IB_QPT_XRC_INI:
1421 *send_cq = to_mcq(qp->ibqp.send_cq);
1422 *recv_cq = *send_cq;
1425 *recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) :
1426 to_mcq(qp->ibwq.cq);
1427 *send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) :
1433 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1435 if (qp->state != IB_QPS_RESET) {
1438 for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size);
1440 struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i];
1441 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
1443 mutex_lock(&wq->mutex);
1447 mutex_unlock(&wq->mutex);
1450 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1451 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1452 pr_warn("modify QP %06x to RESET failed.\n",
1456 mlx4_qp_remove(dev->dev, &qp->mqp);
1457 mlx4_qp_free(dev->dev, &qp->mqp);
1458 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1459 del_gid_entries(qp);
1463 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1464 enum mlx4_ib_source_type src, int is_user)
1466 struct mlx4_ib_cq *send_cq, *recv_cq;
1467 unsigned long flags;
1469 if (qp->state != IB_QPS_RESET) {
1470 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1471 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1472 pr_warn("modify QP %06x to RESET failed.\n",
1474 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1475 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1477 qp->pri.smac_port = 0;
1480 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1483 if (qp->pri.vid < 0x1000) {
1484 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1485 qp->pri.vid = 0xFFFF;
1486 qp->pri.candidate_vid = 0xFFFF;
1487 qp->pri.update_vid = 0;
1489 if (qp->alt.vid < 0x1000) {
1490 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1491 qp->alt.vid = 0xFFFF;
1492 qp->alt.candidate_vid = 0xFFFF;
1493 qp->alt.update_vid = 0;
1497 get_cqs(qp, src, &send_cq, &recv_cq);
1499 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1500 mlx4_ib_lock_cqs(send_cq, recv_cq);
1502 /* del from lists under both locks above to protect reset flow paths */
1503 list_del(&qp->qps_list);
1504 list_del(&qp->cq_send_list);
1505 list_del(&qp->cq_recv_list);
1507 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1508 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1509 if (send_cq != recv_cq)
1510 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1513 mlx4_qp_remove(dev->dev, &qp->mqp);
1515 mlx4_ib_unlock_cqs(send_cq, recv_cq);
1516 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1518 mlx4_qp_free(dev->dev, &qp->mqp);
1520 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1521 if (qp->flags & MLX4_IB_QP_NETIF)
1522 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1523 else if (src == MLX4_IB_RWQ_SRC)
1524 mlx4_ib_release_wqn(to_mucontext(
1525 qp->ibwq.uobject->context), qp, 1);
1527 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1530 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1533 if (qp->rq.wqe_cnt) {
1534 struct mlx4_ib_ucontext *mcontext = !src ?
1535 to_mucontext(qp->ibqp.uobject->context) :
1536 to_mucontext(qp->ibwq.uobject->context);
1537 mlx4_ib_db_unmap_user(mcontext, &qp->db);
1539 ib_umem_release(qp->umem);
1541 kvfree(qp->sq.wrid);
1542 kvfree(qp->rq.wrid);
1543 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1544 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1545 free_proxy_bufs(&dev->ib_dev, qp);
1546 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1548 mlx4_db_free(dev->dev, &qp->db);
1551 del_gid_entries(qp);
1554 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1557 if (!mlx4_is_mfunc(dev->dev) ||
1558 (mlx4_is_master(dev->dev) &&
1559 attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1560 return dev->dev->phys_caps.base_sqpn +
1561 (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1564 /* PF or VF -- creating proxies */
1565 if (attr->qp_type == IB_QPT_SMI)
1566 return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy;
1568 return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy;
1571 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1572 struct ib_qp_init_attr *init_attr,
1573 struct ib_udata *udata)
1575 struct mlx4_ib_qp *qp = NULL;
1577 int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1580 if (init_attr->rwq_ind_tbl)
1581 return _mlx4_ib_create_qp_rss(pd, init_attr, udata);
1584 * We only support LSO, vendor flag1, and multicast loopback blocking,
1585 * and only for kernel UD QPs.
1587 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1588 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1589 MLX4_IB_SRIOV_TUNNEL_QP |
1592 MLX4_IB_QP_CREATE_ROCE_V2_GSI))
1593 return ERR_PTR(-EINVAL);
1595 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1596 if (init_attr->qp_type != IB_QPT_UD)
1597 return ERR_PTR(-EINVAL);
1600 if (init_attr->create_flags) {
1601 if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1602 return ERR_PTR(-EINVAL);
1604 if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1605 MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1606 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1607 init_attr->qp_type != IB_QPT_UD) ||
1608 (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1609 init_attr->qp_type > IB_QPT_GSI) ||
1610 (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1611 init_attr->qp_type != IB_QPT_GSI))
1612 return ERR_PTR(-EINVAL);
1615 switch (init_attr->qp_type) {
1616 case IB_QPT_XRC_TGT:
1617 pd = to_mxrcd(init_attr->xrcd)->pd;
1618 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1619 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1621 case IB_QPT_XRC_INI:
1622 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1623 return ERR_PTR(-ENOSYS);
1624 init_attr->recv_cq = init_attr->send_cq;
1628 case IB_QPT_RAW_PACKET:
1629 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1631 return ERR_PTR(-ENOMEM);
1632 qp->pri.vid = 0xFFFF;
1633 qp->alt.vid = 0xFFFF;
1637 err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1638 init_attr, udata, 0, &qp);
1641 return ERR_PTR(err);
1644 qp->ibqp.qp_num = qp->mqp.qpn;
1654 /* Userspace is not allowed to create special QPs: */
1656 return ERR_PTR(-EINVAL);
1657 if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1658 int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev,
1660 MLX4_RES_USAGE_DRIVER);
1663 return ERR_PTR(res);
1665 sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1668 err = create_qp_common(to_mdev(pd->device), pd, MLX4_IB_QP_SRC,
1669 init_attr, udata, sqpn, &qp);
1671 return ERR_PTR(err);
1673 qp->port = init_attr->port_num;
1674 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1675 init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1679 /* Don't support raw QPs */
1680 return ERR_PTR(-EINVAL);
1686 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1687 struct ib_qp_init_attr *init_attr,
1688 struct ib_udata *udata) {
1689 struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1691 struct mlx4_ib_dev *dev = to_mdev(device);
1693 ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1695 if (!IS_ERR(ibqp) &&
1696 (init_attr->qp_type == IB_QPT_GSI) &&
1697 !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1698 struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1699 int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1702 dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1703 init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1704 sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1706 if (IS_ERR(sqp->roce_v2_gsi)) {
1707 pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1708 sqp->roce_v2_gsi = NULL;
1710 sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1711 sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1714 init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1720 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1722 struct mlx4_ib_dev *dev = to_mdev(qp->device);
1723 struct mlx4_ib_qp *mqp = to_mqp(qp);
1725 if (is_qp0(dev, mqp))
1726 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1728 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI &&
1729 dev->qp1_proxy[mqp->port - 1] == mqp) {
1730 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1731 dev->qp1_proxy[mqp->port - 1] = NULL;
1732 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1735 if (mqp->counter_index)
1736 mlx4_ib_free_qp_counter(dev, mqp);
1738 if (qp->rwq_ind_tbl) {
1739 destroy_qp_rss(dev, mqp);
1741 struct mlx4_ib_pd *pd;
1744 destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, !!pd->ibpd.uobject);
1747 if (is_sqp(dev, mqp))
1748 kfree(to_msqp(mqp));
1755 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1757 struct mlx4_ib_qp *mqp = to_mqp(qp);
1759 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1760 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1762 if (sqp->roce_v2_gsi)
1763 ib_destroy_qp(sqp->roce_v2_gsi);
1766 return _mlx4_ib_destroy_qp(qp);
1769 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1772 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC;
1773 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC;
1774 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD;
1775 case MLX4_IB_QPT_XRC_INI:
1776 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC;
1777 case MLX4_IB_QPT_SMI:
1778 case MLX4_IB_QPT_GSI:
1779 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX;
1781 case MLX4_IB_QPT_PROXY_SMI_OWNER:
1782 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1783 MLX4_QP_ST_MLX : -1);
1784 case MLX4_IB_QPT_PROXY_SMI:
1785 case MLX4_IB_QPT_TUN_SMI:
1786 case MLX4_IB_QPT_PROXY_GSI:
1787 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ?
1788 MLX4_QP_ST_UD : -1);
1793 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1798 u32 hw_access_flags = 0;
1800 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1801 dest_rd_atomic = attr->max_dest_rd_atomic;
1803 dest_rd_atomic = qp->resp_depth;
1805 if (attr_mask & IB_QP_ACCESS_FLAGS)
1806 access_flags = attr->qp_access_flags;
1808 access_flags = qp->atomic_rd_en;
1810 if (!dest_rd_atomic)
1811 access_flags &= IB_ACCESS_REMOTE_WRITE;
1813 if (access_flags & IB_ACCESS_REMOTE_READ)
1814 hw_access_flags |= MLX4_QP_BIT_RRE;
1815 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1816 hw_access_flags |= MLX4_QP_BIT_RAE;
1817 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1818 hw_access_flags |= MLX4_QP_BIT_RWE;
1820 return cpu_to_be32(hw_access_flags);
1823 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1826 if (attr_mask & IB_QP_PKEY_INDEX)
1827 sqp->pkey_index = attr->pkey_index;
1828 if (attr_mask & IB_QP_QKEY)
1829 sqp->qkey = attr->qkey;
1830 if (attr_mask & IB_QP_SQ_PSN)
1831 sqp->send_psn = attr->sq_psn;
1834 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1836 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1839 static int _mlx4_set_path(struct mlx4_ib_dev *dev,
1840 const struct rdma_ah_attr *ah,
1841 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1842 struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1848 path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f;
1849 path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah));
1850 if (rdma_ah_get_static_rate(ah)) {
1851 path->static_rate = rdma_ah_get_static_rate(ah) +
1852 MLX4_STAT_RATE_OFFSET;
1853 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1854 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1855 --path->static_rate;
1857 path->static_rate = 0;
1859 if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) {
1860 const struct ib_global_route *grh = rdma_ah_read_grh(ah);
1861 int real_sgid_index =
1862 mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr);
1864 if (real_sgid_index < 0)
1865 return real_sgid_index;
1866 if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1867 pr_err("sgid_index (%u) too large. max is %d\n",
1868 real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1872 path->grh_mylmc |= 1 << 7;
1873 path->mgid_index = real_sgid_index;
1874 path->hop_limit = grh->hop_limit;
1875 path->tclass_flowlabel =
1876 cpu_to_be32((grh->traffic_class << 20) |
1878 memcpy(path->rgid, grh->dgid.raw, 16);
1881 if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) {
1882 if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH))
1885 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1886 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3);
1888 path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1889 if (vlan_tag < 0x1000) {
1890 if (smac_info->vid < 0x1000) {
1891 /* both valid vlan ids */
1892 if (smac_info->vid != vlan_tag) {
1893 /* different VIDs. unreg old and reg new */
1894 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1897 smac_info->candidate_vid = vlan_tag;
1898 smac_info->candidate_vlan_index = vidx;
1899 smac_info->candidate_vlan_port = port;
1900 smac_info->update_vid = 1;
1901 path->vlan_index = vidx;
1903 path->vlan_index = smac_info->vlan_index;
1906 /* no current vlan tag in qp */
1907 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1910 smac_info->candidate_vid = vlan_tag;
1911 smac_info->candidate_vlan_index = vidx;
1912 smac_info->candidate_vlan_port = port;
1913 smac_info->update_vid = 1;
1914 path->vlan_index = vidx;
1916 path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1919 /* have current vlan tag. unregister it at modify-qp success */
1920 if (smac_info->vid < 0x1000) {
1921 smac_info->candidate_vid = 0xFFFF;
1922 smac_info->update_vid = 1;
1926 /* get smac_index for RoCE use.
1927 * If no smac was yet assigned, register one.
1928 * If one was already assigned, but the new mac differs,
1929 * unregister the old one and register the new one.
1931 if ((!smac_info->smac && !smac_info->smac_port) ||
1932 smac_info->smac != smac) {
1933 /* register candidate now, unreg if needed, after success */
1934 smac_index = mlx4_register_mac(dev->dev, port, smac);
1935 if (smac_index >= 0) {
1936 smac_info->candidate_smac_index = smac_index;
1937 smac_info->candidate_smac = smac;
1938 smac_info->candidate_smac_port = port;
1943 smac_index = smac_info->smac_index;
1945 memcpy(path->dmac, ah->roce.dmac, 6);
1946 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1947 /* put MAC table smac index for IBoE */
1948 path->grh_mylmc = (u8) (smac_index) | 0x80;
1950 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1951 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2);
1957 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1958 enum ib_qp_attr_mask qp_attr_mask,
1959 struct mlx4_ib_qp *mqp,
1960 struct mlx4_qp_path *path, u8 port,
1961 u16 vlan_id, u8 *smac)
1963 return _mlx4_set_path(dev, &qp->ah_attr,
1964 mlx4_mac_to_u64(smac),
1966 path, &mqp->pri, port);
1969 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1970 const struct ib_qp_attr *qp,
1971 enum ib_qp_attr_mask qp_attr_mask,
1972 struct mlx4_ib_qp *mqp,
1973 struct mlx4_qp_path *path, u8 port)
1975 return _mlx4_set_path(dev, &qp->alt_ah_attr,
1978 path, &mqp->alt, port);
1981 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1983 struct mlx4_ib_gid_entry *ge, *tmp;
1985 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1986 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1988 ge->port = qp->port;
1993 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1994 struct mlx4_ib_qp *qp,
1995 struct mlx4_qp_context *context)
2000 u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
2002 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
2003 if (!qp->pri.smac && !qp->pri.smac_port) {
2004 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
2005 if (smac_index >= 0) {
2006 qp->pri.candidate_smac_index = smac_index;
2007 qp->pri.candidate_smac = u64_mac;
2008 qp->pri.candidate_smac_port = qp->port;
2009 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
2017 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
2019 struct counter_index *new_counter_index;
2023 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
2024 IB_LINK_LAYER_ETHERNET ||
2025 !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
2026 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
2029 err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER);
2033 new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
2034 if (!new_counter_index) {
2035 mlx4_counter_free(dev->dev, tmp_idx);
2039 new_counter_index->index = tmp_idx;
2040 new_counter_index->allocated = 1;
2041 qp->counter_index = new_counter_index;
2043 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
2044 list_add_tail(&new_counter_index->list,
2045 &dev->counters_table[qp->port - 1].counters_list);
2046 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
2052 MLX4_QPC_ROCE_MODE_1 = 0,
2053 MLX4_QPC_ROCE_MODE_2 = 2,
2054 MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
2057 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
2060 case IB_GID_TYPE_ROCE:
2061 return MLX4_QPC_ROCE_MODE_1;
2062 case IB_GID_TYPE_ROCE_UDP_ENCAP:
2063 return MLX4_QPC_ROCE_MODE_2;
2065 return MLX4_QPC_ROCE_MODE_UNDEFINED;
2070 * Go over all RSS QP's childes (WQs) and apply their HW state according to
2071 * their logic state if the RSS QP is the first RSS QP associated for the WQ.
2073 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num)
2078 for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2079 struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2080 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2082 mutex_lock(&wq->mutex);
2084 /* Mlx4_ib restrictions:
2085 * WQ's is associated to a port according to the RSS QP it is
2087 * In case the WQ is associated to a different port by another
2088 * RSS QP, return a failure.
2090 if ((wq->rss_usecnt > 0) && (wq->port != port_num)) {
2092 mutex_unlock(&wq->mutex);
2095 wq->port = port_num;
2096 if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) {
2097 err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY);
2099 mutex_unlock(&wq->mutex);
2105 mutex_unlock(&wq->mutex);
2111 for (j = (i - 1); j >= 0; j--) {
2112 struct ib_wq *ibwq = ind_tbl->ind_tbl[j];
2113 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2115 mutex_lock(&wq->mutex);
2117 if ((wq->rss_usecnt == 1) &&
2118 (ibwq->state == IB_WQS_RDY))
2119 if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2120 pr_warn("failed to reverse WQN=0x%06x\n",
2124 mutex_unlock(&wq->mutex);
2131 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl)
2135 for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
2136 struct ib_wq *ibwq = ind_tbl->ind_tbl[i];
2137 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq);
2139 mutex_lock(&wq->mutex);
2141 if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY))
2142 if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET))
2143 pr_warn("failed to reverse WQN=%x\n",
2147 mutex_unlock(&wq->mutex);
2151 static void fill_qp_rss_context(struct mlx4_qp_context *context,
2152 struct mlx4_ib_qp *qp)
2154 struct mlx4_rss_context *rss_context;
2156 rss_context = (void *)context + offsetof(struct mlx4_qp_context,
2157 pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
2159 rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz);
2160 rss_context->default_qpn =
2161 cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff);
2162 if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6))
2163 rss_context->base_qpn_udp = rss_context->default_qpn;
2164 rss_context->flags = qp->rss_ctx->flags;
2165 /* Currently support just toeplitz */
2166 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2168 memcpy(rss_context->rss_key, qp->rss_ctx->rss_key,
2169 MLX4_EN_RSS_KEY_SIZE);
2172 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type,
2173 const struct ib_qp_attr *attr, int attr_mask,
2174 enum ib_qp_state cur_state, enum ib_qp_state new_state)
2176 struct ib_uobject *ibuobject;
2177 struct ib_srq *ibsrq;
2178 const struct ib_gid_attr *gid_attr = NULL;
2179 struct ib_rwq_ind_table *rwq_ind_tbl;
2180 enum ib_qp_type qp_type;
2181 struct mlx4_ib_dev *dev;
2182 struct mlx4_ib_qp *qp;
2183 struct mlx4_ib_pd *pd;
2184 struct mlx4_ib_cq *send_cq, *recv_cq;
2185 struct mlx4_qp_context *context;
2186 enum mlx4_qp_optpar optpar = 0;
2192 if (src_type == MLX4_IB_RWQ_SRC) {
2195 ibwq = (struct ib_wq *)src;
2196 ibuobject = ibwq->uobject;
2199 qp_type = IB_QPT_RAW_PACKET;
2200 qp = to_mqp((struct ib_qp *)ibwq);
2201 dev = to_mdev(ibwq->device);
2202 pd = to_mpd(ibwq->pd);
2206 ibqp = (struct ib_qp *)src;
2207 ibuobject = ibqp->uobject;
2209 rwq_ind_tbl = ibqp->rwq_ind_tbl;
2210 qp_type = ibqp->qp_type;
2212 dev = to_mdev(ibqp->device);
2216 /* APM is not supported under RoCE */
2217 if (attr_mask & IB_QP_ALT_PATH &&
2218 rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2219 IB_LINK_LAYER_ETHERNET)
2222 context = kzalloc(sizeof *context, GFP_KERNEL);
2226 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
2227 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
2229 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
2230 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2232 optpar |= MLX4_QP_OPTPAR_PM_STATE;
2233 switch (attr->path_mig_state) {
2234 case IB_MIG_MIGRATED:
2235 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
2238 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
2241 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
2246 if (qp->inl_recv_sz)
2247 context->param3 |= cpu_to_be32(1 << 25);
2249 if (qp->flags & MLX4_IB_QP_SCATTER_FCS)
2250 context->param3 |= cpu_to_be32(1 << 29);
2252 if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI)
2253 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
2254 else if (qp_type == IB_QPT_RAW_PACKET)
2255 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
2256 else if (qp_type == IB_QPT_UD) {
2257 if (qp->flags & MLX4_IB_QP_LSO)
2258 context->mtu_msgmax = (IB_MTU_4096 << 5) |
2259 ilog2(dev->dev->caps.max_gso_sz);
2261 context->mtu_msgmax = (IB_MTU_4096 << 5) | 13;
2262 } else if (attr_mask & IB_QP_PATH_MTU) {
2263 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
2264 pr_err("path MTU (%u) is invalid\n",
2268 context->mtu_msgmax = (attr->path_mtu << 5) |
2269 ilog2(dev->dev->caps.max_msg_sz);
2272 if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */
2274 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
2275 context->rq_size_stride |= qp->rq.wqe_shift - 4;
2279 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
2280 context->sq_size_stride |= qp->sq.wqe_shift - 4;
2282 if (new_state == IB_QPS_RESET && qp->counter_index)
2283 mlx4_ib_free_qp_counter(dev, qp);
2285 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2286 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
2287 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
2288 if (qp_type == IB_QPT_RAW_PACKET)
2289 context->param3 |= cpu_to_be32(1 << 30);
2293 context->usr_page = cpu_to_be32(
2294 mlx4_to_hw_uar_index(dev->dev,
2295 to_mucontext(ibuobject->context)
2298 context->usr_page = cpu_to_be32(
2299 mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
2301 if (attr_mask & IB_QP_DEST_QPN)
2302 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
2304 if (attr_mask & IB_QP_PORT) {
2305 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
2306 !(attr_mask & IB_QP_AV)) {
2307 mlx4_set_sched(&context->pri_path, attr->port_num);
2308 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
2312 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
2313 err = create_qp_lb_counter(dev, qp);
2318 dev->counters_table[qp->port - 1].default_counter;
2319 if (qp->counter_index)
2320 counter_index = qp->counter_index->index;
2322 if (counter_index != -1) {
2323 context->pri_path.counter_index = counter_index;
2324 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
2325 if (qp->counter_index) {
2326 context->pri_path.fl |=
2327 MLX4_FL_ETH_SRC_CHECK_MC_LB;
2328 context->pri_path.vlan_control |=
2329 MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
2332 context->pri_path.counter_index =
2333 MLX4_SINK_COUNTER_INDEX(dev->dev);
2335 if (qp->flags & MLX4_IB_QP_NETIF) {
2336 mlx4_ib_steer_qp_reg(dev, qp, 1);
2340 if (qp_type == IB_QPT_GSI) {
2341 enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
2342 IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
2343 u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
2345 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2349 if (attr_mask & IB_QP_PKEY_INDEX) {
2350 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2351 context->pri_path.disable_pkey_check = 0x40;
2352 context->pri_path.pkey_index = attr->pkey_index;
2353 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
2356 if (attr_mask & IB_QP_AV) {
2357 u8 port_num = mlx4_is_bonded(dev->dev) ? 1 :
2358 attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2362 rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
2363 rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
2366 gid_attr = attr->ah_attr.grh.sgid_attr;
2367 vlan = rdma_vlan_dev_vlan_id(gid_attr->ndev);
2368 memcpy(smac, gid_attr->ndev->dev_addr, ETH_ALEN);
2371 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
2372 port_num, vlan, smac))
2375 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
2376 MLX4_QP_OPTPAR_SCHED_QUEUE);
2379 (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
2380 u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type);
2382 if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
2386 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
2391 if (attr_mask & IB_QP_TIMEOUT) {
2392 context->pri_path.ackto |= attr->timeout << 3;
2393 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
2396 if (attr_mask & IB_QP_ALT_PATH) {
2397 if (attr->alt_port_num == 0 ||
2398 attr->alt_port_num > dev->dev->caps.num_ports)
2401 if (attr->alt_pkey_index >=
2402 dev->dev->caps.pkey_table_len[attr->alt_port_num])
2405 if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
2407 attr->alt_port_num))
2410 context->alt_path.pkey_index = attr->alt_pkey_index;
2411 context->alt_path.ackto = attr->alt_timeout << 3;
2412 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
2415 context->pd = cpu_to_be32(pd->pdn);
2418 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
2419 get_cqs(qp, src_type, &send_cq, &recv_cq);
2420 } else { /* Set dummy CQs to be compatible with HV and PRM */
2421 send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq);
2424 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
2425 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
2427 /* Set "fast registration enabled" for all kernel QPs */
2429 context->params1 |= cpu_to_be32(1 << 11);
2431 if (attr_mask & IB_QP_RNR_RETRY) {
2432 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2433 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
2436 if (attr_mask & IB_QP_RETRY_CNT) {
2437 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2438 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
2441 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2442 if (attr->max_rd_atomic)
2444 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2445 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
2448 if (attr_mask & IB_QP_SQ_PSN)
2449 context->next_send_psn = cpu_to_be32(attr->sq_psn);
2451 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2452 if (attr->max_dest_rd_atomic)
2454 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2455 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
2458 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
2459 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
2460 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
2464 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
2466 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2467 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2468 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
2470 if (attr_mask & IB_QP_RQ_PSN)
2471 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2473 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
2474 if (attr_mask & IB_QP_QKEY) {
2475 if (qp->mlx4_ib_qp_type &
2476 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
2477 context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2479 if (mlx4_is_mfunc(dev->dev) &&
2480 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
2481 (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
2482 MLX4_RESERVED_QKEY_BASE) {
2483 pr_err("Cannot use reserved QKEY"
2484 " 0x%x (range 0xffff0000..0xffffffff"
2485 " is reserved)\n", attr->qkey);
2489 context->qkey = cpu_to_be32(attr->qkey);
2491 optpar |= MLX4_QP_OPTPAR_Q_KEY;
2495 context->srqn = cpu_to_be32(1 << 24 |
2496 to_msrq(ibsrq)->msrq.srqn);
2498 if (qp->rq.wqe_cnt &&
2499 cur_state == IB_QPS_RESET &&
2500 new_state == IB_QPS_INIT)
2501 context->db_rec_addr = cpu_to_be64(qp->db.dma);
2503 if (cur_state == IB_QPS_INIT &&
2504 new_state == IB_QPS_RTR &&
2505 (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI ||
2506 qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) {
2507 context->pri_path.sched_queue = (qp->port - 1) << 6;
2508 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2509 qp->mlx4_ib_qp_type &
2510 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
2511 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
2512 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
2513 context->pri_path.fl = 0x80;
2515 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
2516 context->pri_path.fl = 0x80;
2517 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2519 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2520 IB_LINK_LAYER_ETHERNET) {
2521 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2522 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2523 context->pri_path.feup = 1 << 7; /* don't fsm */
2524 /* handle smac_index */
2525 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2526 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2527 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2528 err = handle_eth_ud_smac_index(dev, qp, context);
2533 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
2534 dev->qp1_proxy[qp->port - 1] = qp;
2539 if (qp_type == IB_QPT_RAW_PACKET) {
2540 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
2541 MLX4_IB_LINK_TYPE_ETH;
2542 if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
2543 /* set QP to receive both tunneled & non-tunneled packets */
2545 context->srqn = cpu_to_be32(7 << 28);
2549 if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
2550 int is_eth = rdma_port_get_link_layer(
2551 &dev->ib_dev, qp->port) ==
2552 IB_LINK_LAYER_ETHERNET;
2554 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
2555 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
2559 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
2560 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2566 cur_state == IB_QPS_RESET &&
2567 new_state == IB_QPS_INIT)
2568 context->rlkey_roce_mode |= (1 << 4);
2571 * Before passing a kernel QP to the HW, make sure that the
2572 * ownership bits of the send queue are set and the SQ
2573 * headroom is stamped so that the hardware doesn't start
2574 * processing stale work requests.
2577 cur_state == IB_QPS_RESET &&
2578 new_state == IB_QPS_INIT) {
2579 struct mlx4_wqe_ctrl_seg *ctrl;
2582 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2583 ctrl = get_send_wqe(qp, i);
2584 ctrl->owner_opcode = cpu_to_be32(1 << 31);
2585 if (qp->sq_max_wqes_per_wr == 1)
2586 ctrl->qpn_vlan.fence_size =
2587 1 << (qp->sq.wqe_shift - 4);
2589 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2594 cur_state == IB_QPS_RESET &&
2595 new_state == IB_QPS_INIT) {
2596 fill_qp_rss_context(context, qp);
2597 context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2600 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2601 to_mlx4_state(new_state), context, optpar,
2602 sqd_event, &qp->mqp);
2606 qp->state = new_state;
2608 if (attr_mask & IB_QP_ACCESS_FLAGS)
2609 qp->atomic_rd_en = attr->qp_access_flags;
2610 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2611 qp->resp_depth = attr->max_dest_rd_atomic;
2612 if (attr_mask & IB_QP_PORT) {
2613 qp->port = attr->port_num;
2614 update_mcg_macs(dev, qp);
2616 if (attr_mask & IB_QP_ALT_PATH)
2617 qp->alt_port = attr->alt_port_num;
2619 if (is_sqp(dev, qp))
2620 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2623 * If we moved QP0 to RTR, bring the IB link up; if we moved
2624 * QP0 to RESET or ERROR, bring the link back down.
2626 if (is_qp0(dev, qp)) {
2627 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2628 if (mlx4_INIT_PORT(dev->dev, qp->port))
2629 pr_warn("INIT_PORT failed for port %d\n",
2632 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2633 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2634 mlx4_CLOSE_PORT(dev->dev, qp->port);
2638 * If we moved a kernel QP to RESET, clean up all old CQ
2639 * entries and reinitialize the QP.
2641 if (new_state == IB_QPS_RESET) {
2643 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2644 ibsrq ? to_msrq(ibsrq) : NULL);
2645 if (send_cq != recv_cq)
2646 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2652 qp->sq_next_wqe = 0;
2656 if (qp->flags & MLX4_IB_QP_NETIF)
2657 mlx4_ib_steer_qp_reg(dev, qp, 0);
2659 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2660 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2662 qp->pri.smac_port = 0;
2665 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2668 if (qp->pri.vid < 0x1000) {
2669 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2670 qp->pri.vid = 0xFFFF;
2671 qp->pri.candidate_vid = 0xFFFF;
2672 qp->pri.update_vid = 0;
2675 if (qp->alt.vid < 0x1000) {
2676 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2677 qp->alt.vid = 0xFFFF;
2678 qp->alt.candidate_vid = 0xFFFF;
2679 qp->alt.update_vid = 0;
2683 if (err && qp->counter_index)
2684 mlx4_ib_free_qp_counter(dev, qp);
2685 if (err && steer_qp)
2686 mlx4_ib_steer_qp_reg(dev, qp, 0);
2688 if (qp->pri.candidate_smac ||
2689 (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2691 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2693 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2694 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2695 qp->pri.smac = qp->pri.candidate_smac;
2696 qp->pri.smac_index = qp->pri.candidate_smac_index;
2697 qp->pri.smac_port = qp->pri.candidate_smac_port;
2699 qp->pri.candidate_smac = 0;
2700 qp->pri.candidate_smac_index = 0;
2701 qp->pri.candidate_smac_port = 0;
2703 if (qp->alt.candidate_smac) {
2705 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2708 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2709 qp->alt.smac = qp->alt.candidate_smac;
2710 qp->alt.smac_index = qp->alt.candidate_smac_index;
2711 qp->alt.smac_port = qp->alt.candidate_smac_port;
2713 qp->alt.candidate_smac = 0;
2714 qp->alt.candidate_smac_index = 0;
2715 qp->alt.candidate_smac_port = 0;
2718 if (qp->pri.update_vid) {
2720 if (qp->pri.candidate_vid < 0x1000)
2721 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2722 qp->pri.candidate_vid);
2724 if (qp->pri.vid < 0x1000)
2725 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2727 qp->pri.vid = qp->pri.candidate_vid;
2728 qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2729 qp->pri.vlan_index = qp->pri.candidate_vlan_index;
2731 qp->pri.candidate_vid = 0xFFFF;
2732 qp->pri.update_vid = 0;
2735 if (qp->alt.update_vid) {
2737 if (qp->alt.candidate_vid < 0x1000)
2738 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2739 qp->alt.candidate_vid);
2741 if (qp->alt.vid < 0x1000)
2742 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2744 qp->alt.vid = qp->alt.candidate_vid;
2745 qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2746 qp->alt.vlan_index = qp->alt.candidate_vlan_index;
2748 qp->alt.candidate_vid = 0xFFFF;
2749 qp->alt.update_vid = 0;
2756 MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE |
2760 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2761 int attr_mask, struct ib_udata *udata)
2763 enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2764 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2765 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2766 enum ib_qp_state cur_state, new_state;
2768 mutex_lock(&qp->mutex);
2770 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2771 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2773 if (cur_state != new_state || cur_state != IB_QPS_RESET) {
2774 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2775 ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2778 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2780 pr_debug("qpn 0x%x: invalid attribute mask specified "
2781 "for transition %d to %d. qp_type %d,"
2782 " attr_mask 0x%x\n",
2783 ibqp->qp_num, cur_state, new_state,
2784 ibqp->qp_type, attr_mask);
2788 if (ibqp->rwq_ind_tbl) {
2789 if (!(((cur_state == IB_QPS_RESET) &&
2790 (new_state == IB_QPS_INIT)) ||
2791 ((cur_state == IB_QPS_INIT) &&
2792 (new_state == IB_QPS_RTR)))) {
2793 pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n",
2794 ibqp->qp_num, cur_state, new_state);
2800 if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) {
2801 pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n",
2802 ibqp->qp_num, attr_mask, cur_state, new_state);
2809 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2810 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2811 if ((ibqp->qp_type == IB_QPT_RC) ||
2812 (ibqp->qp_type == IB_QPT_UD) ||
2813 (ibqp->qp_type == IB_QPT_UC) ||
2814 (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2815 (ibqp->qp_type == IB_QPT_XRC_INI)) {
2816 attr->port_num = mlx4_ib_bond_next_port(dev);
2819 /* no sense in changing port_num
2820 * when ports are bonded */
2821 attr_mask &= ~IB_QP_PORT;
2825 if ((attr_mask & IB_QP_PORT) &&
2826 (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2827 pr_debug("qpn 0x%x: invalid port number (%d) specified "
2828 "for transition %d to %d. qp_type %d\n",
2829 ibqp->qp_num, attr->port_num, cur_state,
2830 new_state, ibqp->qp_type);
2834 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2835 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2836 IB_LINK_LAYER_ETHERNET))
2839 if (attr_mask & IB_QP_PKEY_INDEX) {
2840 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2841 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2842 pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2843 "for transition %d to %d. qp_type %d\n",
2844 ibqp->qp_num, attr->pkey_index, cur_state,
2845 new_state, ibqp->qp_type);
2850 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2851 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2852 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2853 "Transition %d to %d. qp_type %d\n",
2854 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2855 new_state, ibqp->qp_type);
2859 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2860 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2861 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2862 "Transition %d to %d. qp_type %d\n",
2863 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2864 new_state, ibqp->qp_type);
2868 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2873 if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) {
2874 err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num);
2879 err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask,
2880 cur_state, new_state);
2882 if (ibqp->rwq_ind_tbl && err)
2883 bring_down_rss_rwqs(ibqp->rwq_ind_tbl);
2885 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2889 mutex_unlock(&qp->mutex);
2893 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2894 int attr_mask, struct ib_udata *udata)
2896 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2899 ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2901 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2902 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2905 if (sqp->roce_v2_gsi)
2906 err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2908 pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2914 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2917 for (i = 0; i < dev->caps.num_ports; i++) {
2918 if (qpn == dev->caps.spec_qps[i].qp0_proxy ||
2919 qpn == dev->caps.spec_qps[i].qp0_tunnel) {
2920 *qkey = dev->caps.spec_qps[i].qp0_qkey;
2927 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2928 const struct ib_ud_wr *wr,
2929 void *wqe, unsigned *mlx_seg_len)
2931 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2932 struct ib_device *ib_dev = &mdev->ib_dev;
2933 struct mlx4_wqe_mlx_seg *mlx = wqe;
2934 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2935 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2943 if (wr->wr.opcode != IB_WR_SEND)
2948 for (i = 0; i < wr->wr.num_sge; ++i)
2949 send_size += wr->wr.sg_list[i].length;
2951 /* for proxy-qp0 sends, need to add in size of tunnel header */
2952 /* for tunnel-qp0 sends, tunnel header is already in s/g list */
2953 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2954 send_size += sizeof (struct mlx4_ib_tunnel_header);
2956 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2958 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2959 sqp->ud_header.lrh.service_level =
2960 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2961 sqp->ud_header.lrh.destination_lid =
2962 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2963 sqp->ud_header.lrh.source_lid =
2964 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2967 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2969 /* force loopback */
2970 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2971 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2973 sqp->ud_header.lrh.virtual_lane = 0;
2974 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2975 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2976 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2977 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2978 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2980 sqp->ud_header.bth.destination_qpn =
2981 cpu_to_be32(mdev->dev->caps.spec_qps[sqp->qp.port - 1].qp0_tunnel);
2983 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2984 if (mlx4_is_master(mdev->dev)) {
2985 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2988 if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2991 sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2992 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2994 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2995 sqp->ud_header.immediate_present = 0;
2997 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3000 * Inline data segments may not cross a 64 byte boundary. If
3001 * our UD header is bigger than the space available up to the
3002 * next 64 byte boundary in the WQE, use two inline data
3003 * segments to hold the UD header.
3005 spc = MLX4_INLINE_ALIGN -
3006 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3007 if (header_size <= spc) {
3008 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3009 memcpy(inl + 1, sqp->header_buf, header_size);
3012 inl->byte_count = cpu_to_be32(1 << 31 | spc);
3013 memcpy(inl + 1, sqp->header_buf, spc);
3015 inl = (void *) (inl + 1) + spc;
3016 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3018 * Need a barrier here to make sure all the data is
3019 * visible before the byte_count field is set.
3020 * Otherwise the HCA prefetcher could grab the 64-byte
3021 * chunk with this inline segment and get a valid (!=
3022 * 0xffffffff) byte count but stale data, and end up
3023 * generating a packet with bad headers.
3025 * The first inline segment's byte_count field doesn't
3026 * need a barrier, because it comes after a
3027 * control/MLX segment and therefore is at an offset
3031 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3036 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3040 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
3042 union sl2vl_tbl_to_u64 tmp_vltab;
3047 tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
3048 vl = tmp_vltab.sl8[sl >> 1];
3056 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num,
3057 int index, union ib_gid *gid,
3058 enum ib_gid_type *gid_type)
3060 struct mlx4_ib_iboe *iboe = &ibdev->iboe;
3061 struct mlx4_port_gid_table *port_gid_table;
3062 unsigned long flags;
3064 port_gid_table = &iboe->gids[port_num - 1];
3065 spin_lock_irqsave(&iboe->lock, flags);
3066 memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid));
3067 *gid_type = port_gid_table->gids[index].gid_type;
3068 spin_unlock_irqrestore(&iboe->lock, flags);
3069 if (rdma_is_zero_gid(gid))
3075 #define MLX4_ROCEV2_QP1_SPORT 0xC000
3076 static int build_mlx_header(struct mlx4_ib_sqp *sqp, const struct ib_ud_wr *wr,
3077 void *wqe, unsigned *mlx_seg_len)
3079 struct ib_device *ib_dev = sqp->qp.ibqp.device;
3080 struct mlx4_ib_dev *ibdev = to_mdev(ib_dev);
3081 struct mlx4_wqe_mlx_seg *mlx = wqe;
3082 struct mlx4_wqe_ctrl_seg *ctrl = wqe;
3083 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
3084 struct mlx4_ib_ah *ah = to_mah(wr->ah);
3094 bool is_vlan = false;
3096 bool is_udp = false;
3100 for (i = 0; i < wr->wr.num_sge; ++i)
3101 send_size += wr->wr.sg_list[i].length;
3103 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
3104 is_grh = mlx4_ib_ah_grh_present(ah);
3106 enum ib_gid_type gid_type;
3107 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3108 /* When multi-function is enabled, the ib_core gid
3109 * indexes don't necessarily match the hw ones, so
3110 * we must use our own cache */
3111 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
3112 be32_to_cpu(ah->av.ib.port_pd) >> 24,
3113 ah->av.ib.gid_index, &sgid.raw[0]);
3117 err = fill_gid_by_hw_index(ibdev, sqp->qp.port,
3118 ah->av.ib.gid_index,
3121 is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
3123 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
3133 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
3134 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
3138 err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
3139 ip_version, is_udp, 0, &sqp->ud_header);
3144 sqp->ud_header.lrh.service_level =
3145 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
3146 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
3147 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
3150 if (is_grh || (ip_version == 6)) {
3151 sqp->ud_header.grh.traffic_class =
3152 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3153 sqp->ud_header.grh.flow_label =
3154 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
3155 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit;
3157 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
3159 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
3160 /* When multi-function is enabled, the ib_core gid
3161 * indexes don't necessarily match the hw ones, so
3162 * we must use our own cache
3164 sqp->ud_header.grh.source_gid.global.subnet_prefix =
3165 cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
3166 demux[sqp->qp.port - 1].
3168 sqp->ud_header.grh.source_gid.global.interface_id =
3169 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
3170 guid_cache[ah->av.ib.gid_index];
3172 sqp->ud_header.grh.source_gid =
3173 ah->ibah.sgid_attr->gid;
3176 memcpy(sqp->ud_header.grh.destination_gid.raw,
3177 ah->av.ib.dgid, 16);
3180 if (ip_version == 4) {
3181 sqp->ud_header.ip4.tos =
3182 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
3183 sqp->ud_header.ip4.id = 0;
3184 sqp->ud_header.ip4.frag_off = htons(IP_DF);
3185 sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
3187 memcpy(&sqp->ud_header.ip4.saddr,
3189 memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
3190 sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
3194 sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
3195 sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
3196 sqp->ud_header.udp.csum = 0;
3199 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
3202 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
3203 (sqp->ud_header.lrh.destination_lid ==
3204 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
3205 (sqp->ud_header.lrh.service_level << 8));
3206 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
3207 mlx->flags |= cpu_to_be32(0x1); /* force loopback */
3208 mlx->rlid = sqp->ud_header.lrh.destination_lid;
3211 switch (wr->wr.opcode) {
3213 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
3214 sqp->ud_header.immediate_present = 0;
3216 case IB_WR_SEND_WITH_IMM:
3217 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
3218 sqp->ud_header.immediate_present = 1;
3219 sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
3226 struct in6_addr in6;
3228 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
3230 ether_type = (!is_udp) ? ETH_P_IBOE:
3231 (ip_version == 4 ? ETH_P_IP : ETH_P_IPV6);
3233 mlx->sched_prio = cpu_to_be16(pcp);
3235 ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
3236 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
3237 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
3238 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
3239 memcpy(&in6, sgid.raw, sizeof(in6));
3242 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
3243 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
3245 sqp->ud_header.eth.type = cpu_to_be16(ether_type);
3247 sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
3248 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
3251 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 :
3252 sl_to_vl(to_mdev(ib_dev),
3253 sqp->ud_header.lrh.service_level,
3255 if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
3257 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
3258 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
3260 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
3261 if (!sqp->qp.ibqp.qp_num)
3262 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
3264 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
3265 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
3266 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
3267 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
3268 sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
3269 sqp->qkey : wr->remote_qkey);
3270 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
3272 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
3275 pr_err("built UD header of size %d:\n", header_size);
3276 for (i = 0; i < header_size / 4; ++i) {
3278 pr_err(" [%02x] ", i * 4);
3280 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
3281 if ((i + 1) % 8 == 0)
3288 * Inline data segments may not cross a 64 byte boundary. If
3289 * our UD header is bigger than the space available up to the
3290 * next 64 byte boundary in the WQE, use two inline data
3291 * segments to hold the UD header.
3293 spc = MLX4_INLINE_ALIGN -
3294 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3295 if (header_size <= spc) {
3296 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
3297 memcpy(inl + 1, sqp->header_buf, header_size);
3300 inl->byte_count = cpu_to_be32(1 << 31 | spc);
3301 memcpy(inl + 1, sqp->header_buf, spc);
3303 inl = (void *) (inl + 1) + spc;
3304 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
3306 * Need a barrier here to make sure all the data is
3307 * visible before the byte_count field is set.
3308 * Otherwise the HCA prefetcher could grab the 64-byte
3309 * chunk with this inline segment and get a valid (!=
3310 * 0xffffffff) byte count but stale data, and end up
3311 * generating a packet with bad headers.
3313 * The first inline segment's byte_count field doesn't
3314 * need a barrier, because it comes after a
3315 * control/MLX segment and therefore is at an offset
3319 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
3324 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
3328 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
3331 struct mlx4_ib_cq *cq;
3333 cur = wq->head - wq->tail;
3334 if (likely(cur + nreq < wq->max_post))
3338 spin_lock(&cq->lock);
3339 cur = wq->head - wq->tail;
3340 spin_unlock(&cq->lock);
3342 return cur + nreq >= wq->max_post;
3345 static __be32 convert_access(int acc)
3347 return (acc & IB_ACCESS_REMOTE_ATOMIC ?
3348 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) |
3349 (acc & IB_ACCESS_REMOTE_WRITE ?
3350 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
3351 (acc & IB_ACCESS_REMOTE_READ ?
3352 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) |
3353 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) |
3354 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
3357 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
3358 const struct ib_reg_wr *wr)
3360 struct mlx4_ib_mr *mr = to_mmr(wr->mr);
3362 fseg->flags = convert_access(wr->access);
3363 fseg->mem_key = cpu_to_be32(wr->key);
3364 fseg->buf_list = cpu_to_be64(mr->page_map);
3365 fseg->start_addr = cpu_to_be64(mr->ibmr.iova);
3366 fseg->reg_len = cpu_to_be64(mr->ibmr.length);
3367 fseg->offset = 0; /* XXX -- is this just for ZBVA? */
3368 fseg->page_size = cpu_to_be32(ilog2(mr->ibmr.page_size));
3369 fseg->reserved[0] = 0;
3370 fseg->reserved[1] = 0;
3373 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
3375 memset(iseg, 0, sizeof(*iseg));
3376 iseg->mem_key = cpu_to_be32(rkey);
3379 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
3380 u64 remote_addr, u32 rkey)
3382 rseg->raddr = cpu_to_be64(remote_addr);
3383 rseg->rkey = cpu_to_be32(rkey);
3387 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
3388 const struct ib_atomic_wr *wr)
3390 if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
3391 aseg->swap_add = cpu_to_be64(wr->swap);
3392 aseg->compare = cpu_to_be64(wr->compare_add);
3393 } else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
3394 aseg->swap_add = cpu_to_be64(wr->compare_add);
3395 aseg->compare = cpu_to_be64(wr->compare_add_mask);
3397 aseg->swap_add = cpu_to_be64(wr->compare_add);
3403 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
3404 const struct ib_atomic_wr *wr)
3406 aseg->swap_add = cpu_to_be64(wr->swap);
3407 aseg->swap_add_mask = cpu_to_be64(wr->swap_mask);
3408 aseg->compare = cpu_to_be64(wr->compare_add);
3409 aseg->compare_mask = cpu_to_be64(wr->compare_add_mask);
3412 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
3413 const struct ib_ud_wr *wr)
3415 memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
3416 dseg->dqpn = cpu_to_be32(wr->remote_qpn);
3417 dseg->qkey = cpu_to_be32(wr->remote_qkey);
3418 dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
3419 memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
3422 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
3423 struct mlx4_wqe_datagram_seg *dseg,
3424 const struct ib_ud_wr *wr,
3425 enum mlx4_ib_qp_type qpt)
3427 union mlx4_ext_av *av = &to_mah(wr->ah)->av;
3428 struct mlx4_av sqp_av = {0};
3429 int port = *((u8 *) &av->ib.port_pd) & 0x3;
3431 /* force loopback */
3432 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
3433 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
3434 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
3435 cpu_to_be32(0xf0000000);
3437 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
3438 if (qpt == MLX4_IB_QPT_PROXY_GSI)
3439 dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel);
3441 dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel);
3442 /* Use QKEY from the QP context, which is set by master */
3443 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
3446 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe,
3447 unsigned *mlx_seg_len)
3449 struct mlx4_wqe_inline_seg *inl = wqe;
3450 struct mlx4_ib_tunnel_header hdr;
3451 struct mlx4_ib_ah *ah = to_mah(wr->ah);
3455 memcpy(&hdr.av, &ah->av, sizeof hdr.av);
3456 hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
3457 hdr.pkey_index = cpu_to_be16(wr->pkey_index);
3458 hdr.qkey = cpu_to_be32(wr->remote_qkey);
3459 memcpy(hdr.mac, ah->av.eth.mac, 6);
3460 hdr.vlan = ah->av.eth.vlan;
3462 spc = MLX4_INLINE_ALIGN -
3463 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
3464 if (sizeof (hdr) <= spc) {
3465 memcpy(inl + 1, &hdr, sizeof (hdr));
3467 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
3470 memcpy(inl + 1, &hdr, spc);
3472 inl->byte_count = cpu_to_be32(1 << 31 | spc);
3474 inl = (void *) (inl + 1) + spc;
3475 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
3477 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
3482 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
3485 static void set_mlx_icrc_seg(void *dseg)
3488 struct mlx4_wqe_inline_seg *iseg = dseg;
3493 * Need a barrier here before writing the byte_count field to
3494 * make sure that all the data is visible before the
3495 * byte_count field is set. Otherwise, if the segment begins
3496 * a new cacheline, the HCA prefetcher could grab the 64-byte
3497 * chunk and get a valid (!= * 0xffffffff) byte count but
3498 * stale data, and end up sending the wrong data.
3502 iseg->byte_count = cpu_to_be32((1 << 31) | 4);
3505 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3507 dseg->lkey = cpu_to_be32(sg->lkey);
3508 dseg->addr = cpu_to_be64(sg->addr);
3511 * Need a barrier here before writing the byte_count field to
3512 * make sure that all the data is visible before the
3513 * byte_count field is set. Otherwise, if the segment begins
3514 * a new cacheline, the HCA prefetcher could grab the 64-byte
3515 * chunk and get a valid (!= * 0xffffffff) byte count but
3516 * stale data, and end up sending the wrong data.
3520 dseg->byte_count = cpu_to_be32(sg->length);
3523 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
3525 dseg->byte_count = cpu_to_be32(sg->length);
3526 dseg->lkey = cpu_to_be32(sg->lkey);
3527 dseg->addr = cpu_to_be64(sg->addr);
3530 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe,
3531 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp,
3532 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh)
3534 unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
3536 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
3537 *blh = cpu_to_be32(1 << 6);
3539 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
3540 wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
3543 memcpy(wqe->header, wr->header, wr->hlen);
3545 *lso_hdr_sz = cpu_to_be32(wr->mss << 16 | wr->hlen);
3546 *lso_seg_len = halign;
3550 static __be32 send_ieth(const struct ib_send_wr *wr)
3552 switch (wr->opcode) {
3553 case IB_WR_SEND_WITH_IMM:
3554 case IB_WR_RDMA_WRITE_WITH_IMM:
3555 return wr->ex.imm_data;
3557 case IB_WR_SEND_WITH_INV:
3558 return cpu_to_be32(wr->ex.invalidate_rkey);
3565 static void add_zero_len_inline(void *wqe)
3567 struct mlx4_wqe_inline_seg *inl = wqe;
3569 inl->byte_count = cpu_to_be32(1 << 31);
3572 static int _mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3573 struct ib_send_wr **bad_wr, bool drain)
3575 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3577 struct mlx4_wqe_ctrl_seg *ctrl;
3578 struct mlx4_wqe_data_seg *dseg;
3579 unsigned long flags;
3583 int uninitialized_var(stamp);
3584 int uninitialized_var(size);
3585 unsigned uninitialized_var(seglen);
3588 __be32 uninitialized_var(lso_hdr_sz);
3591 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3593 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
3594 struct mlx4_ib_sqp *sqp = to_msqp(qp);
3596 if (sqp->roce_v2_gsi) {
3597 struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
3598 enum ib_gid_type gid_type;
3601 if (!fill_gid_by_hw_index(mdev, sqp->qp.port,
3602 ah->av.ib.gid_index,
3604 qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
3605 to_mqp(sqp->roce_v2_gsi) : qp;
3607 pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
3608 ah->av.ib.gid_index);
3612 spin_lock_irqsave(&qp->sq.lock, flags);
3613 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3621 ind = qp->sq_next_wqe;
3623 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3627 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3633 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3639 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3640 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3643 (wr->send_flags & IB_SEND_SIGNALED ?
3644 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3645 (wr->send_flags & IB_SEND_SOLICITED ?
3646 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3647 ((wr->send_flags & IB_SEND_IP_CSUM) ?
3648 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3649 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3652 ctrl->imm = send_ieth(wr);
3654 wqe += sizeof *ctrl;
3655 size = sizeof *ctrl / 16;
3657 switch (qp->mlx4_ib_qp_type) {
3658 case MLX4_IB_QPT_RC:
3659 case MLX4_IB_QPT_UC:
3660 switch (wr->opcode) {
3661 case IB_WR_ATOMIC_CMP_AND_SWP:
3662 case IB_WR_ATOMIC_FETCH_AND_ADD:
3663 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3664 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3665 atomic_wr(wr)->rkey);
3666 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3668 set_atomic_seg(wqe, atomic_wr(wr));
3669 wqe += sizeof (struct mlx4_wqe_atomic_seg);
3671 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3672 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3676 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3677 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3678 atomic_wr(wr)->rkey);
3679 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3681 set_masked_atomic_seg(wqe, atomic_wr(wr));
3682 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg);
3684 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3685 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3689 case IB_WR_RDMA_READ:
3690 case IB_WR_RDMA_WRITE:
3691 case IB_WR_RDMA_WRITE_WITH_IMM:
3692 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3694 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3695 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3698 case IB_WR_LOCAL_INV:
3699 ctrl->srcrb_flags |=
3700 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3701 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3702 wqe += sizeof (struct mlx4_wqe_local_inval_seg);
3703 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3707 ctrl->srcrb_flags |=
3708 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3709 set_reg_seg(wqe, reg_wr(wr));
3710 wqe += sizeof(struct mlx4_wqe_fmr_seg);
3711 size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3715 /* No extra segments required for sends */
3720 case MLX4_IB_QPT_TUN_SMI_OWNER:
3721 err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3723 if (unlikely(err)) {
3728 size += seglen / 16;
3730 case MLX4_IB_QPT_TUN_SMI:
3731 case MLX4_IB_QPT_TUN_GSI:
3732 /* this is a UD qp used in MAD responses to slaves. */
3733 set_datagram_seg(wqe, ud_wr(wr));
3734 /* set the forced-loopback bit in the data seg av */
3735 *(__be32 *) wqe |= cpu_to_be32(0x80000000);
3736 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3737 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3739 case MLX4_IB_QPT_UD:
3740 set_datagram_seg(wqe, ud_wr(wr));
3741 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3742 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3744 if (wr->opcode == IB_WR_LSO) {
3745 err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3747 if (unlikely(err)) {
3751 lso_wqe = (__be32 *) wqe;
3753 size += seglen / 16;
3757 case MLX4_IB_QPT_PROXY_SMI_OWNER:
3758 err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3760 if (unlikely(err)) {
3765 size += seglen / 16;
3766 /* to start tunnel header on a cache-line boundary */
3767 add_zero_len_inline(wqe);
3770 build_tunnel_header(ud_wr(wr), wqe, &seglen);
3772 size += seglen / 16;
3774 case MLX4_IB_QPT_PROXY_SMI:
3775 case MLX4_IB_QPT_PROXY_GSI:
3776 /* If we are tunneling special qps, this is a UD qp.
3777 * In this case we first add a UD segment targeting
3778 * the tunnel qp, and then add a header with address
3780 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3782 qp->mlx4_ib_qp_type);
3783 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3784 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3785 build_tunnel_header(ud_wr(wr), wqe, &seglen);
3787 size += seglen / 16;
3790 case MLX4_IB_QPT_SMI:
3791 case MLX4_IB_QPT_GSI:
3792 err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3794 if (unlikely(err)) {
3799 size += seglen / 16;
3807 * Write data segments in reverse order, so as to
3808 * overwrite cacheline stamp last within each
3809 * cacheline. This avoids issues with WQE
3814 dseg += wr->num_sge - 1;
3815 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3817 /* Add one more inline data segment for ICRC for MLX sends */
3818 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3819 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3820 qp->mlx4_ib_qp_type &
3821 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3822 set_mlx_icrc_seg(dseg + 1);
3823 size += sizeof (struct mlx4_wqe_data_seg) / 16;
3826 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3827 set_data_seg(dseg, wr->sg_list + i);
3830 * Possibly overwrite stamping in cacheline with LSO
3831 * segment only after making sure all data segments
3835 *lso_wqe = lso_hdr_sz;
3837 ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ?
3838 MLX4_WQE_CTRL_FENCE : 0) | size;
3841 * Make sure descriptor is fully written before
3842 * setting ownership bit (because HW can start
3843 * executing as soon as we do).
3847 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3853 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3854 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3856 stamp = ind + qp->sq_spare_wqes;
3857 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3860 * We can improve latency by not stamping the last
3861 * send queue WQE until after ringing the doorbell, so
3862 * only stamp here if there are still more WQEs to post.
3864 * Same optimization applies to padding with NOP wqe
3865 * in case of WQE shrinking (used to prevent wrap-around
3866 * in the middle of WR).
3869 stamp_send_wqe(qp, stamp, size * 16);
3870 ind = pad_wraparound(qp, ind);
3876 qp->sq.head += nreq;
3879 * Make sure that descriptors are written before
3884 writel_relaxed(qp->doorbell_qpn,
3885 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3888 * Make sure doorbells don't leak out of SQ spinlock
3889 * and reach the HCA out of order.
3893 stamp_send_wqe(qp, stamp, size * 16);
3895 ind = pad_wraparound(qp, ind);
3896 qp->sq_next_wqe = ind;
3899 spin_unlock_irqrestore(&qp->sq.lock, flags);
3904 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3905 struct ib_send_wr **bad_wr)
3907 return _mlx4_ib_post_send(ibqp, wr, bad_wr, false);
3910 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3911 struct ib_recv_wr **bad_wr, bool drain)
3913 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3914 struct mlx4_wqe_data_seg *scat;
3915 unsigned long flags;
3921 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3923 max_gs = qp->rq.max_gs;
3924 spin_lock_irqsave(&qp->rq.lock, flags);
3926 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR &&
3934 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3936 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3937 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3943 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3949 scat = get_recv_wqe(qp, ind);
3951 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3952 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3953 ib_dma_sync_single_for_device(ibqp->device,
3954 qp->sqp_proxy_rcv[ind].map,
3955 sizeof (struct mlx4_ib_proxy_sqp_hdr),
3958 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3959 /* use dma lkey from upper layer entry */
3960 scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3961 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3966 for (i = 0; i < wr->num_sge; ++i)
3967 __set_data_seg(scat + i, wr->sg_list + i);
3970 scat[i].byte_count = 0;
3971 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
3975 qp->rq.wrid[ind] = wr->wr_id;
3977 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3982 qp->rq.head += nreq;
3985 * Make sure that descriptors are written before
3990 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3993 spin_unlock_irqrestore(&qp->rq.lock, flags);
3998 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3999 struct ib_recv_wr **bad_wr)
4001 return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false);
4004 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
4006 switch (mlx4_state) {
4007 case MLX4_QP_STATE_RST: return IB_QPS_RESET;
4008 case MLX4_QP_STATE_INIT: return IB_QPS_INIT;
4009 case MLX4_QP_STATE_RTR: return IB_QPS_RTR;
4010 case MLX4_QP_STATE_RTS: return IB_QPS_RTS;
4011 case MLX4_QP_STATE_SQ_DRAINING:
4012 case MLX4_QP_STATE_SQD: return IB_QPS_SQD;
4013 case MLX4_QP_STATE_SQER: return IB_QPS_SQE;
4014 case MLX4_QP_STATE_ERR: return IB_QPS_ERR;
4019 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
4021 switch (mlx4_mig_state) {
4022 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED;
4023 case MLX4_QP_PM_REARM: return IB_MIG_REARM;
4024 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
4029 static int to_ib_qp_access_flags(int mlx4_flags)
4033 if (mlx4_flags & MLX4_QP_BIT_RRE)
4034 ib_flags |= IB_ACCESS_REMOTE_READ;
4035 if (mlx4_flags & MLX4_QP_BIT_RWE)
4036 ib_flags |= IB_ACCESS_REMOTE_WRITE;
4037 if (mlx4_flags & MLX4_QP_BIT_RAE)
4038 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4043 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev,
4044 struct rdma_ah_attr *ah_attr,
4045 struct mlx4_qp_path *path)
4047 struct mlx4_dev *dev = ibdev->dev;
4048 u8 port_num = path->sched_queue & 0x40 ? 2 : 1;
4050 memset(ah_attr, 0, sizeof(*ah_attr));
4051 if (port_num == 0 || port_num > dev->caps.num_ports)
4053 ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num);
4055 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE)
4056 rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) |
4057 ((path->sched_queue & 4) << 1));
4059 rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf);
4060 rdma_ah_set_port_num(ah_attr, port_num);
4062 rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid));
4063 rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f);
4064 rdma_ah_set_static_rate(ah_attr,
4065 path->static_rate ? path->static_rate - 5 : 0);
4066 if (path->grh_mylmc & (1 << 7)) {
4067 rdma_ah_set_grh(ah_attr, NULL,
4068 be32_to_cpu(path->tclass_flowlabel) & 0xfffff,
4071 (be32_to_cpu(path->tclass_flowlabel)
4073 rdma_ah_set_dgid_raw(ah_attr, path->rgid);
4077 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
4078 struct ib_qp_init_attr *qp_init_attr)
4080 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
4081 struct mlx4_ib_qp *qp = to_mqp(ibqp);
4082 struct mlx4_qp_context context;
4086 if (ibqp->rwq_ind_tbl)
4089 mutex_lock(&qp->mutex);
4091 if (qp->state == IB_QPS_RESET) {
4092 qp_attr->qp_state = IB_QPS_RESET;
4096 err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
4102 mlx4_state = be32_to_cpu(context.flags) >> 28;
4104 qp->state = to_ib_qp_state(mlx4_state);
4105 qp_attr->qp_state = qp->state;
4106 qp_attr->path_mtu = context.mtu_msgmax >> 5;
4107 qp_attr->path_mig_state =
4108 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
4109 qp_attr->qkey = be32_to_cpu(context.qkey);
4110 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
4111 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff;
4112 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff;
4113 qp_attr->qp_access_flags =
4114 to_ib_qp_access_flags(be32_to_cpu(context.params2));
4116 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4117 to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
4118 to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
4119 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
4120 qp_attr->alt_port_num =
4121 rdma_ah_get_port_num(&qp_attr->alt_ah_attr);
4124 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
4125 if (qp_attr->qp_state == IB_QPS_INIT)
4126 qp_attr->port_num = qp->port;
4128 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
4130 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4131 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
4133 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
4135 qp_attr->max_dest_rd_atomic =
4136 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
4137 qp_attr->min_rnr_timer =
4138 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
4139 qp_attr->timeout = context.pri_path.ackto >> 3;
4140 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7;
4141 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7;
4142 qp_attr->alt_timeout = context.alt_path.ackto >> 3;
4145 qp_attr->cur_qp_state = qp_attr->qp_state;
4146 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
4147 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
4149 if (!ibqp->uobject) {
4150 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
4151 qp_attr->cap.max_send_sge = qp->sq.max_gs;
4153 qp_attr->cap.max_send_wr = 0;
4154 qp_attr->cap.max_send_sge = 0;
4158 * We don't support inline sends for kernel QPs (yet), and we
4159 * don't know what userspace's value should be.
4161 qp_attr->cap.max_inline_data = 0;
4163 qp_init_attr->cap = qp_attr->cap;
4165 qp_init_attr->create_flags = 0;
4166 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4167 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4169 if (qp->flags & MLX4_IB_QP_LSO)
4170 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
4172 if (qp->flags & MLX4_IB_QP_NETIF)
4173 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
4175 qp_init_attr->sq_sig_type =
4176 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
4177 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4180 mutex_unlock(&qp->mutex);
4184 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd,
4185 struct ib_wq_init_attr *init_attr,
4186 struct ib_udata *udata)
4188 struct mlx4_ib_dev *dev;
4189 struct ib_qp_init_attr ib_qp_init_attr;
4190 struct mlx4_ib_qp *qp;
4191 struct mlx4_ib_create_wq ucmd;
4192 int err, required_cmd_sz;
4194 if (!(udata && pd->uobject))
4195 return ERR_PTR(-EINVAL);
4197 required_cmd_sz = offsetof(typeof(ucmd), comp_mask) +
4198 sizeof(ucmd.comp_mask);
4199 if (udata->inlen < required_cmd_sz) {
4200 pr_debug("invalid inlen\n");
4201 return ERR_PTR(-EINVAL);
4204 if (udata->inlen > sizeof(ucmd) &&
4205 !ib_is_udata_cleared(udata, sizeof(ucmd),
4206 udata->inlen - sizeof(ucmd))) {
4207 pr_debug("inlen is not supported\n");
4208 return ERR_PTR(-EOPNOTSUPP);
4212 return ERR_PTR(-EOPNOTSUPP);
4214 dev = to_mdev(pd->device);
4216 if (init_attr->wq_type != IB_WQT_RQ) {
4217 pr_debug("unsupported wq type %d\n", init_attr->wq_type);
4218 return ERR_PTR(-EOPNOTSUPP);
4221 if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS) {
4222 pr_debug("unsupported create_flags %u\n",
4223 init_attr->create_flags);
4224 return ERR_PTR(-EOPNOTSUPP);
4227 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
4229 return ERR_PTR(-ENOMEM);
4231 qp->pri.vid = 0xFFFF;
4232 qp->alt.vid = 0xFFFF;
4234 memset(&ib_qp_init_attr, 0, sizeof(ib_qp_init_attr));
4235 ib_qp_init_attr.qp_context = init_attr->wq_context;
4236 ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET;
4237 ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr;
4238 ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge;
4239 ib_qp_init_attr.recv_cq = init_attr->cq;
4240 ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */
4242 if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS)
4243 ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS;
4245 err = create_qp_common(dev, pd, MLX4_IB_RWQ_SRC, &ib_qp_init_attr,
4249 return ERR_PTR(err);
4252 qp->ibwq.event_handler = init_attr->event_handler;
4253 qp->ibwq.wq_num = qp->mqp.qpn;
4254 qp->ibwq.state = IB_WQS_RESET;
4259 static int ib_wq2qp_state(enum ib_wq_state state)
4263 return IB_QPS_RESET;
4271 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state)
4273 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4274 enum ib_qp_state qp_cur_state;
4275 enum ib_qp_state qp_new_state;
4279 /* ib_qp.state represents the WQ HW state while ib_wq.state represents
4280 * the WQ logic state.
4282 qp_cur_state = qp->state;
4283 qp_new_state = ib_wq2qp_state(new_state);
4285 if (ib_wq2qp_state(new_state) == qp_cur_state)
4288 if (new_state == IB_WQS_RDY) {
4289 struct ib_qp_attr attr = {};
4291 attr.port_num = qp->port;
4292 attr_mask = IB_QP_PORT;
4294 err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr,
4295 attr_mask, IB_QPS_RESET, IB_QPS_INIT);
4297 pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n",
4302 qp_cur_state = IB_QPS_INIT;
4306 err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask,
4307 qp_cur_state, qp_new_state);
4309 if (err && (qp_cur_state == IB_QPS_INIT)) {
4310 qp_new_state = IB_QPS_RESET;
4311 if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL,
4312 attr_mask, IB_QPS_INIT, IB_QPS_RESET)) {
4313 pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n",
4315 qp_new_state = IB_QPS_INIT;
4319 qp->state = qp_new_state;
4324 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr,
4325 u32 wq_attr_mask, struct ib_udata *udata)
4327 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4328 struct mlx4_ib_modify_wq ucmd = {};
4329 size_t required_cmd_sz;
4330 enum ib_wq_state cur_state, new_state;
4333 required_cmd_sz = offsetof(typeof(ucmd), reserved) +
4334 sizeof(ucmd.reserved);
4335 if (udata->inlen < required_cmd_sz)
4338 if (udata->inlen > sizeof(ucmd) &&
4339 !ib_is_udata_cleared(udata, sizeof(ucmd),
4340 udata->inlen - sizeof(ucmd)))
4343 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4346 if (ucmd.comp_mask || ucmd.reserved)
4349 if (wq_attr_mask & IB_WQ_FLAGS)
4352 cur_state = wq_attr_mask & IB_WQ_CUR_STATE ? wq_attr->curr_wq_state :
4354 new_state = wq_attr_mask & IB_WQ_STATE ? wq_attr->wq_state : cur_state;
4356 if (cur_state < IB_WQS_RESET || cur_state > IB_WQS_ERR ||
4357 new_state < IB_WQS_RESET || new_state > IB_WQS_ERR)
4360 if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR))
4363 if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET))
4366 /* Need to protect against the parent RSS which also may modify WQ
4369 mutex_lock(&qp->mutex);
4371 /* Can update HW state only if a RSS QP has already associated to this
4372 * WQ, so we can apply its port on the WQ.
4375 err = _mlx4_ib_modify_wq(ibwq, new_state);
4378 ibwq->state = new_state;
4380 mutex_unlock(&qp->mutex);
4385 int mlx4_ib_destroy_wq(struct ib_wq *ibwq)
4387 struct mlx4_ib_dev *dev = to_mdev(ibwq->device);
4388 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq);
4390 if (qp->counter_index)
4391 mlx4_ib_free_qp_counter(dev, qp);
4393 destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, 1);
4400 struct ib_rwq_ind_table
4401 *mlx4_ib_create_rwq_ind_table(struct ib_device *device,
4402 struct ib_rwq_ind_table_init_attr *init_attr,
4403 struct ib_udata *udata)
4405 struct ib_rwq_ind_table *rwq_ind_table;
4406 struct mlx4_ib_create_rwq_ind_tbl_resp resp = {};
4407 unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size;
4408 unsigned int base_wqn;
4409 size_t min_resp_len;
4413 if (udata->inlen > 0 &&
4414 !ib_is_udata_cleared(udata, 0,
4416 return ERR_PTR(-EOPNOTSUPP);
4418 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4419 if (udata->outlen && udata->outlen < min_resp_len)
4420 return ERR_PTR(-EINVAL);
4423 device->attrs.rss_caps.max_rwq_indirection_table_size) {
4424 pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n",
4426 device->attrs.rss_caps.max_rwq_indirection_table_size);
4427 return ERR_PTR(-EINVAL);
4430 base_wqn = init_attr->ind_tbl[0]->wq_num;
4432 if (base_wqn % ind_tbl_size) {
4433 pr_debug("WQN=0x%x isn't aligned with indirection table size\n",
4435 return ERR_PTR(-EINVAL);
4438 for (i = 1; i < ind_tbl_size; i++) {
4439 if (++base_wqn != init_attr->ind_tbl[i]->wq_num) {
4440 pr_debug("indirection table's WQNs aren't consecutive\n");
4441 return ERR_PTR(-EINVAL);
4445 rwq_ind_table = kzalloc(sizeof(*rwq_ind_table), GFP_KERNEL);
4447 return ERR_PTR(-ENOMEM);
4449 if (udata->outlen) {
4450 resp.response_length = offsetof(typeof(resp), response_length) +
4451 sizeof(resp.response_length);
4452 err = ib_copy_to_udata(udata, &resp, resp.response_length);
4457 return rwq_ind_table;
4460 kfree(rwq_ind_table);
4461 return ERR_PTR(err);
4464 int mlx4_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4466 kfree(ib_rwq_ind_tbl);
4470 struct mlx4_ib_drain_cqe {
4472 struct completion done;
4475 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
4477 struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe,
4478 struct mlx4_ib_drain_cqe,
4481 complete(&cqe->done);
4484 /* This function returns only once the drained WR was completed */
4485 static void handle_drain_completion(struct ib_cq *cq,
4486 struct mlx4_ib_drain_cqe *sdrain,
4487 struct mlx4_ib_dev *dev)
4489 struct mlx4_dev *mdev = dev->dev;
4491 if (cq->poll_ctx == IB_POLL_DIRECT) {
4492 while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0)
4493 ib_process_cq_direct(cq, -1);
4497 if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4498 struct mlx4_ib_cq *mcq = to_mcq(cq);
4499 bool triggered = false;
4500 unsigned long flags;
4502 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
4503 /* Make sure that the CQ handler won't run if wasn't run yet */
4504 if (!mcq->mcq.reset_notify_added)
4505 mcq->mcq.reset_notify_added = 1;
4508 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
4511 /* Wait for any scheduled/running task to be ended */
4512 switch (cq->poll_ctx) {
4513 case IB_POLL_SOFTIRQ:
4514 irq_poll_disable(&cq->iop);
4515 irq_poll_enable(&cq->iop);
4517 case IB_POLL_WORKQUEUE:
4518 cancel_work_sync(&cq->work);
4525 /* Run the CQ handler - this makes sure that the drain WR will
4526 * be processed if wasn't processed yet.
4528 mcq->mcq.comp(&mcq->mcq);
4531 wait_for_completion(&sdrain->done);
4534 void mlx4_ib_drain_sq(struct ib_qp *qp)
4536 struct ib_cq *cq = qp->send_cq;
4537 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4538 struct mlx4_ib_drain_cqe sdrain;
4539 struct ib_send_wr *bad_swr;
4540 struct ib_rdma_wr swr = {
4543 { .wr_cqe = &sdrain.cqe, },
4544 .opcode = IB_WR_RDMA_WRITE,
4548 struct mlx4_ib_dev *dev = to_mdev(qp->device);
4549 struct mlx4_dev *mdev = dev->dev;
4551 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4552 if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4553 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4557 sdrain.cqe.done = mlx4_ib_drain_qp_done;
4558 init_completion(&sdrain.done);
4560 ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true);
4562 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
4566 handle_drain_completion(cq, &sdrain, dev);
4569 void mlx4_ib_drain_rq(struct ib_qp *qp)
4571 struct ib_cq *cq = qp->recv_cq;
4572 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
4573 struct mlx4_ib_drain_cqe rdrain;
4574 struct ib_recv_wr rwr = {}, *bad_rwr;
4576 struct mlx4_ib_dev *dev = to_mdev(qp->device);
4577 struct mlx4_dev *mdev = dev->dev;
4579 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
4580 if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) {
4581 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4585 rwr.wr_cqe = &rdrain.cqe;
4586 rdrain.cqe.done = mlx4_ib_drain_qp_done;
4587 init_completion(&rdrain.done);
4589 ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true);
4591 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
4595 handle_drain_completion(cq, &rdrain, dev);