2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/log2.h>
35 #include <linux/slab.h>
36 #include <linux/netdevice.h>
38 #include <rdma/ib_cache.h>
39 #include <rdma/ib_pack.h>
40 #include <rdma/ib_addr.h>
41 #include <rdma/ib_mad.h>
43 #include <linux/mlx4/driver.h>
44 #include <linux/mlx4/qp.h>
49 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
50 struct mlx4_ib_cq *recv_cq);
51 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
52 struct mlx4_ib_cq *recv_cq);
55 MLX4_IB_ACK_REQ_FREQ = 8,
59 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83,
60 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
61 MLX4_IB_LINK_TYPE_IB = 0,
62 MLX4_IB_LINK_TYPE_ETH = 1
67 * Largest possible UD header: send with GRH and immediate
68 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
69 * tag. (LRH would only use 8 bytes, so Ethernet is the
72 MLX4_IB_UD_HEADER_SIZE = 82,
73 MLX4_IB_LSO_HEADER_SPARE = 128,
77 MLX4_IB_IBOE_ETHERTYPE = 0x8915
85 struct ib_ud_header ud_header;
86 u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
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_FAST_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),
117 [IB_WR_BIND_MW] = cpu_to_be32(MLX4_OPCODE_BIND_MW),
120 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
122 return container_of(mqp, struct mlx4_ib_sqp, qp);
125 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
127 if (!mlx4_is_master(dev->dev))
130 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
131 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
135 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
140 /* PPF or Native -- real SQP */
141 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
142 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
143 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
146 /* VF or PF -- proxy SQP */
147 if (mlx4_is_mfunc(dev->dev)) {
148 for (i = 0; i < dev->dev->caps.num_ports; i++) {
149 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
150 qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
159 /* used for INIT/CLOSE port logic */
160 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
165 /* PPF or Native -- real QP0 */
166 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
167 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
168 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
171 /* VF or PF -- proxy QP0 */
172 if (mlx4_is_mfunc(dev->dev)) {
173 for (i = 0; i < dev->dev->caps.num_ports; i++) {
174 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
183 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
185 return mlx4_buf_offset(&qp->buf, offset);
188 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
190 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
193 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
195 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
199 * Stamp a SQ WQE so that it is invalid if prefetched by marking the
200 * first four bytes of every 64 byte chunk with
201 * 0x7FFFFFF | (invalid_ownership_value << 31).
203 * When the max work request size is less than or equal to the WQE
204 * basic block size, as an optimization, we can stamp all WQEs with
205 * 0xffffffff, and skip the very first chunk of each WQE.
207 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
215 struct mlx4_wqe_ctrl_seg *ctrl;
217 if (qp->sq_max_wqes_per_wr > 1) {
218 s = roundup(size, 1U << qp->sq.wqe_shift);
219 for (i = 0; i < s; i += 64) {
220 ind = (i >> qp->sq.wqe_shift) + n;
221 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
222 cpu_to_be32(0xffffffff);
223 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
224 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
228 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
229 s = (ctrl->fence_size & 0x3f) << 4;
230 for (i = 64; i < s; i += 64) {
232 *wqe = cpu_to_be32(0xffffffff);
237 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
239 struct mlx4_wqe_ctrl_seg *ctrl;
240 struct mlx4_wqe_inline_seg *inl;
244 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
245 s = sizeof(struct mlx4_wqe_ctrl_seg);
247 if (qp->ibqp.qp_type == IB_QPT_UD) {
248 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
249 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
250 memset(dgram, 0, sizeof *dgram);
251 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
252 s += sizeof(struct mlx4_wqe_datagram_seg);
255 /* Pad the remainder of the WQE with an inline data segment. */
258 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
260 ctrl->srcrb_flags = 0;
261 ctrl->fence_size = size / 16;
263 * Make sure descriptor is fully written before setting ownership bit
264 * (because HW can start executing as soon as we do).
268 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
269 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
271 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
274 /* Post NOP WQE to prevent wrap-around in the middle of WR */
275 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
277 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
278 if (unlikely(s < qp->sq_max_wqes_per_wr)) {
279 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
285 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
287 struct ib_event event;
288 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
290 if (type == MLX4_EVENT_TYPE_PATH_MIG)
291 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
293 if (ibqp->event_handler) {
294 event.device = ibqp->device;
295 event.element.qp = ibqp;
297 case MLX4_EVENT_TYPE_PATH_MIG:
298 event.event = IB_EVENT_PATH_MIG;
300 case MLX4_EVENT_TYPE_COMM_EST:
301 event.event = IB_EVENT_COMM_EST;
303 case MLX4_EVENT_TYPE_SQ_DRAINED:
304 event.event = IB_EVENT_SQ_DRAINED;
306 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
307 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
309 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
310 event.event = IB_EVENT_QP_FATAL;
312 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
313 event.event = IB_EVENT_PATH_MIG_ERR;
315 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
316 event.event = IB_EVENT_QP_REQ_ERR;
318 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
319 event.event = IB_EVENT_QP_ACCESS_ERR;
322 pr_warn("Unexpected event type %d "
323 "on QP %06x\n", type, qp->qpn);
327 ibqp->event_handler(&event, ibqp->qp_context);
331 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
334 * UD WQEs must have a datagram segment.
335 * RC and UC WQEs might have a remote address segment.
336 * MLX WQEs need two extra inline data segments (for the UD
337 * header and space for the ICRC).
341 return sizeof (struct mlx4_wqe_ctrl_seg) +
342 sizeof (struct mlx4_wqe_datagram_seg) +
343 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
344 case MLX4_IB_QPT_PROXY_SMI_OWNER:
345 case MLX4_IB_QPT_PROXY_SMI:
346 case MLX4_IB_QPT_PROXY_GSI:
347 return sizeof (struct mlx4_wqe_ctrl_seg) +
348 sizeof (struct mlx4_wqe_datagram_seg) + 64;
349 case MLX4_IB_QPT_TUN_SMI_OWNER:
350 case MLX4_IB_QPT_TUN_GSI:
351 return sizeof (struct mlx4_wqe_ctrl_seg) +
352 sizeof (struct mlx4_wqe_datagram_seg);
355 return sizeof (struct mlx4_wqe_ctrl_seg) +
356 sizeof (struct mlx4_wqe_raddr_seg);
358 return sizeof (struct mlx4_wqe_ctrl_seg) +
359 sizeof (struct mlx4_wqe_atomic_seg) +
360 sizeof (struct mlx4_wqe_raddr_seg);
361 case MLX4_IB_QPT_SMI:
362 case MLX4_IB_QPT_GSI:
363 return sizeof (struct mlx4_wqe_ctrl_seg) +
364 ALIGN(MLX4_IB_UD_HEADER_SIZE +
365 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
367 sizeof (struct mlx4_wqe_inline_seg),
368 sizeof (struct mlx4_wqe_data_seg)) +
370 sizeof (struct mlx4_wqe_inline_seg),
371 sizeof (struct mlx4_wqe_data_seg));
373 return sizeof (struct mlx4_wqe_ctrl_seg);
377 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
378 int is_user, int has_rq, struct mlx4_ib_qp *qp)
380 /* Sanity check RQ size before proceeding */
381 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
382 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
386 if (cap->max_recv_wr)
389 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
391 /* HW requires >= 1 RQ entry with >= 1 gather entry */
392 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
395 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr));
396 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
397 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
400 /* leave userspace return values as they were, so as not to break ABI */
402 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt;
403 cap->max_recv_sge = qp->rq.max_gs;
405 cap->max_recv_wr = qp->rq.max_post =
406 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
407 cap->max_recv_sge = min(qp->rq.max_gs,
408 min(dev->dev->caps.max_sq_sg,
409 dev->dev->caps.max_rq_sg));
415 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
416 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
420 /* Sanity check SQ size before proceeding */
421 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
422 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
423 cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
424 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
428 * For MLX transport we need 2 extra S/G entries:
429 * one for the header and one for the checksum at the end
431 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
432 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
433 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
436 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
437 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
438 send_wqe_overhead(type, qp->flags);
440 if (s > dev->dev->caps.max_sq_desc_sz)
444 * Hermon supports shrinking WQEs, such that a single work
445 * request can include multiple units of 1 << wqe_shift. This
446 * way, work requests can differ in size, and do not have to
447 * be a power of 2 in size, saving memory and speeding up send
448 * WR posting. Unfortunately, if we do this then the
449 * wqe_index field in CQEs can't be used to look up the WR ID
450 * anymore, so we do this only if selective signaling is off.
452 * Further, on 32-bit platforms, we can't use vmap() to make
453 * the QP buffer virtually contiguous. Thus we have to use
454 * constant-sized WRs to make sure a WR is always fully within
455 * a single page-sized chunk.
457 * Finally, we use NOP work requests to pad the end of the
458 * work queue, to avoid wrap-around in the middle of WR. We
459 * set NEC bit to avoid getting completions with error for
460 * these NOP WRs, but since NEC is only supported starting
461 * with firmware 2.2.232, we use constant-sized WRs for older
464 * And, since MLX QPs only support SEND, we use constant-sized
467 * We look for the smallest value of wqe_shift such that the
468 * resulting number of wqes does not exceed device
471 * We set WQE size to at least 64 bytes, this way stamping
472 * invalidates each WQE.
474 if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
475 qp->sq_signal_bits && BITS_PER_LONG == 64 &&
476 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
477 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
478 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
479 qp->sq.wqe_shift = ilog2(64);
481 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
484 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
487 * We need to leave 2 KB + 1 WR of headroom in the SQ to
488 * allow HW to prefetch.
490 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
491 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
492 qp->sq_max_wqes_per_wr +
495 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
498 if (qp->sq_max_wqes_per_wr <= 1)
504 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
505 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
506 send_wqe_overhead(type, qp->flags)) /
507 sizeof (struct mlx4_wqe_data_seg);
509 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
510 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
511 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
513 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
515 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
519 cap->max_send_wr = qp->sq.max_post =
520 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
521 cap->max_send_sge = min(qp->sq.max_gs,
522 min(dev->dev->caps.max_sq_sg,
523 dev->dev->caps.max_rq_sg));
524 /* We don't support inline sends for kernel QPs (yet) */
525 cap->max_inline_data = 0;
530 static int set_user_sq_size(struct mlx4_ib_dev *dev,
531 struct mlx4_ib_qp *qp,
532 struct mlx4_ib_create_qp *ucmd)
534 /* Sanity check SQ size before proceeding */
535 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes ||
536 ucmd->log_sq_stride >
537 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
538 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
541 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
542 qp->sq.wqe_shift = ucmd->log_sq_stride;
544 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
545 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
550 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
555 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
557 if (!qp->sqp_proxy_rcv)
559 for (i = 0; i < qp->rq.wqe_cnt; i++) {
560 qp->sqp_proxy_rcv[i].addr =
561 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
563 if (!qp->sqp_proxy_rcv[i].addr)
565 qp->sqp_proxy_rcv[i].map =
566 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
567 sizeof (struct mlx4_ib_proxy_sqp_hdr),
569 if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
570 kfree(qp->sqp_proxy_rcv[i].addr);
579 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
580 sizeof (struct mlx4_ib_proxy_sqp_hdr),
582 kfree(qp->sqp_proxy_rcv[i].addr);
584 kfree(qp->sqp_proxy_rcv);
585 qp->sqp_proxy_rcv = NULL;
589 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
593 for (i = 0; i < qp->rq.wqe_cnt; i++) {
594 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
595 sizeof (struct mlx4_ib_proxy_sqp_hdr),
597 kfree(qp->sqp_proxy_rcv[i].addr);
599 kfree(qp->sqp_proxy_rcv);
602 static int qp_has_rq(struct ib_qp_init_attr *attr)
604 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
610 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
613 for (i = 0; i < dev->caps.num_ports; i++) {
614 if (qpn == dev->caps.qp0_proxy[i])
615 return !!dev->caps.qp0_qkey[i];
620 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
621 struct mlx4_ib_qp *qp)
623 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
624 mlx4_counter_free(dev->dev, qp->counter_index->index);
625 list_del(&qp->counter_index->list);
626 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
628 kfree(qp->counter_index);
629 qp->counter_index = NULL;
632 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
633 struct ib_qp_init_attr *init_attr,
634 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
639 struct mlx4_ib_sqp *sqp;
640 struct mlx4_ib_qp *qp;
641 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
642 struct mlx4_ib_cq *mcq;
645 /* When tunneling special qps, we use a plain UD qp */
647 if (mlx4_is_mfunc(dev->dev) &&
648 (!mlx4_is_master(dev->dev) ||
649 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
650 if (init_attr->qp_type == IB_QPT_GSI)
651 qp_type = MLX4_IB_QPT_PROXY_GSI;
653 if (mlx4_is_master(dev->dev) ||
654 qp0_enabled_vf(dev->dev, sqpn))
655 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
657 qp_type = MLX4_IB_QPT_PROXY_SMI;
661 /* add extra sg entry for tunneling */
662 init_attr->cap.max_recv_sge++;
663 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
664 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
665 container_of(init_attr,
666 struct mlx4_ib_qp_tunnel_init_attr, init_attr);
667 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
668 tnl_init->proxy_qp_type != IB_QPT_GSI) ||
669 !mlx4_is_master(dev->dev))
671 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
672 qp_type = MLX4_IB_QPT_TUN_GSI;
673 else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
674 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
676 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
678 qp_type = MLX4_IB_QPT_TUN_SMI;
679 /* we are definitely in the PPF here, since we are creating
680 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
681 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
682 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
687 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
688 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
689 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
690 sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
694 qp->pri.vid = 0xFFFF;
695 qp->alt.vid = 0xFFFF;
697 qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
700 qp->pri.vid = 0xFFFF;
701 qp->alt.vid = 0xFFFF;
706 qp->mlx4_ib_qp_type = qp_type;
708 mutex_init(&qp->mutex);
709 spin_lock_init(&qp->sq.lock);
710 spin_lock_init(&qp->rq.lock);
711 INIT_LIST_HEAD(&qp->gid_list);
712 INIT_LIST_HEAD(&qp->steering_rules);
714 qp->state = IB_QPS_RESET;
715 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
716 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
718 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
723 struct mlx4_ib_create_qp ucmd;
725 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
730 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
732 err = set_user_sq_size(dev, qp, &ucmd);
736 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
738 if (IS_ERR(qp->umem)) {
739 err = PTR_ERR(qp->umem);
743 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
744 ilog2(qp->umem->page_size), &qp->mtt);
748 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
752 if (qp_has_rq(init_attr)) {
753 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
754 ucmd.db_addr, &qp->db);
759 qp->sq_no_prefetch = 0;
761 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
762 qp->flags |= MLX4_IB_QP_LSO;
764 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
765 if (dev->steering_support ==
766 MLX4_STEERING_MODE_DEVICE_MANAGED)
767 qp->flags |= MLX4_IB_QP_NETIF;
772 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
776 if (qp_has_rq(init_attr)) {
777 err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
784 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
789 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
794 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
798 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
799 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
800 if (!qp->sq.wrid || !qp->rq.wrid) {
807 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
808 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
809 if (alloc_proxy_bufs(pd->device, qp)) {
815 /* Raw packet QPNs may not have bits 6,7 set in their qp_num;
816 * otherwise, the WQE BlueFlame setup flow wrongly causes
818 if (init_attr->qp_type == IB_QPT_RAW_PACKET)
819 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
820 (init_attr->cap.max_send_wr ?
821 MLX4_RESERVE_ETH_BF_QP : 0) |
822 (init_attr->cap.max_recv_wr ?
823 MLX4_RESERVE_A0_QP : 0));
825 if (qp->flags & MLX4_IB_QP_NETIF)
826 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
828 err = mlx4_qp_reserve_range(dev->dev, 1, 1,
834 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
835 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
837 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
841 if (init_attr->qp_type == IB_QPT_XRC_TGT)
842 qp->mqp.qpn |= (1 << 23);
845 * Hardware wants QPN written in big-endian order (after
846 * shifting) for send doorbell. Precompute this value to save
847 * a little bit when posting sends.
849 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
851 qp->mqp.event = mlx4_ib_qp_event;
855 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
856 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
857 to_mcq(init_attr->recv_cq));
858 /* Maintain device to QPs access, needed for further handling
861 list_add_tail(&qp->qps_list, &dev->qp_list);
862 /* Maintain CQ to QPs access, needed for further handling
865 mcq = to_mcq(init_attr->send_cq);
866 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
867 mcq = to_mcq(init_attr->recv_cq);
868 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
869 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
870 to_mcq(init_attr->recv_cq));
871 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
876 if (qp->flags & MLX4_IB_QP_NETIF)
877 mlx4_ib_steer_qp_free(dev, qpn, 1);
879 mlx4_qp_release_range(dev->dev, qpn, 1);
882 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
883 free_proxy_bufs(pd->device, qp);
886 if (qp_has_rq(init_attr))
887 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
894 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
898 ib_umem_release(qp->umem);
900 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
903 if (!pd->uobject && qp_has_rq(init_attr))
904 mlx4_db_free(dev->dev, &qp->db);
912 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
915 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
916 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
917 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
918 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
919 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
920 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
921 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
926 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
927 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
929 if (send_cq == recv_cq) {
930 spin_lock(&send_cq->lock);
931 __acquire(&recv_cq->lock);
932 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
933 spin_lock(&send_cq->lock);
934 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
936 spin_lock(&recv_cq->lock);
937 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
941 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
942 __releases(&send_cq->lock) __releases(&recv_cq->lock)
944 if (send_cq == recv_cq) {
945 __release(&recv_cq->lock);
946 spin_unlock(&send_cq->lock);
947 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
948 spin_unlock(&recv_cq->lock);
949 spin_unlock(&send_cq->lock);
951 spin_unlock(&send_cq->lock);
952 spin_unlock(&recv_cq->lock);
956 static void del_gid_entries(struct mlx4_ib_qp *qp)
958 struct mlx4_ib_gid_entry *ge, *tmp;
960 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
966 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
968 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
969 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
971 return to_mpd(qp->ibqp.pd);
974 static void get_cqs(struct mlx4_ib_qp *qp,
975 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
977 switch (qp->ibqp.qp_type) {
979 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
983 *send_cq = to_mcq(qp->ibqp.send_cq);
987 *send_cq = to_mcq(qp->ibqp.send_cq);
988 *recv_cq = to_mcq(qp->ibqp.recv_cq);
993 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
996 struct mlx4_ib_cq *send_cq, *recv_cq;
999 if (qp->state != IB_QPS_RESET) {
1000 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1001 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1002 pr_warn("modify QP %06x to RESET failed.\n",
1004 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1005 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1007 qp->pri.smac_port = 0;
1010 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1013 if (qp->pri.vid < 0x1000) {
1014 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1015 qp->pri.vid = 0xFFFF;
1016 qp->pri.candidate_vid = 0xFFFF;
1017 qp->pri.update_vid = 0;
1019 if (qp->alt.vid < 0x1000) {
1020 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1021 qp->alt.vid = 0xFFFF;
1022 qp->alt.candidate_vid = 0xFFFF;
1023 qp->alt.update_vid = 0;
1027 get_cqs(qp, &send_cq, &recv_cq);
1029 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1030 mlx4_ib_lock_cqs(send_cq, recv_cq);
1032 /* del from lists under both locks above to protect reset flow paths */
1033 list_del(&qp->qps_list);
1034 list_del(&qp->cq_send_list);
1035 list_del(&qp->cq_recv_list);
1037 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1038 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1039 if (send_cq != recv_cq)
1040 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1043 mlx4_qp_remove(dev->dev, &qp->mqp);
1045 mlx4_ib_unlock_cqs(send_cq, recv_cq);
1046 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1048 mlx4_qp_free(dev->dev, &qp->mqp);
1050 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1051 if (qp->flags & MLX4_IB_QP_NETIF)
1052 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1054 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1057 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1061 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1063 ib_umem_release(qp->umem);
1067 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1068 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1069 free_proxy_bufs(&dev->ib_dev, qp);
1070 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1072 mlx4_db_free(dev->dev, &qp->db);
1075 del_gid_entries(qp);
1078 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1081 if (!mlx4_is_mfunc(dev->dev) ||
1082 (mlx4_is_master(dev->dev) &&
1083 attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1084 return dev->dev->phys_caps.base_sqpn +
1085 (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1088 /* PF or VF -- creating proxies */
1089 if (attr->qp_type == IB_QPT_SMI)
1090 return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1092 return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1095 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1096 struct ib_qp_init_attr *init_attr,
1097 struct ib_udata *udata)
1099 struct mlx4_ib_qp *qp = NULL;
1101 int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1105 gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1106 GFP_NOIO : GFP_KERNEL;
1108 * We only support LSO, vendor flag1, and multicast loopback blocking,
1109 * and only for kernel UD QPs.
1111 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1112 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1113 MLX4_IB_SRIOV_TUNNEL_QP |
1116 MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1117 return ERR_PTR(-EINVAL);
1119 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1120 if (init_attr->qp_type != IB_QPT_UD)
1121 return ERR_PTR(-EINVAL);
1124 if (init_attr->create_flags &&
1125 ((udata && init_attr->create_flags & ~(sup_u_create_flags)) ||
1126 ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1127 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1128 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)) &&
1129 init_attr->qp_type != IB_QPT_UD) ||
1130 ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1131 init_attr->qp_type > IB_QPT_GSI)))
1132 return ERR_PTR(-EINVAL);
1134 switch (init_attr->qp_type) {
1135 case IB_QPT_XRC_TGT:
1136 pd = to_mxrcd(init_attr->xrcd)->pd;
1137 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1138 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1140 case IB_QPT_XRC_INI:
1141 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1142 return ERR_PTR(-ENOSYS);
1143 init_attr->recv_cq = init_attr->send_cq;
1147 case IB_QPT_RAW_PACKET:
1148 qp = kzalloc(sizeof *qp, gfp);
1150 return ERR_PTR(-ENOMEM);
1151 qp->pri.vid = 0xFFFF;
1152 qp->alt.vid = 0xFFFF;
1156 err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1157 udata, 0, &qp, gfp);
1159 return ERR_PTR(err);
1161 qp->ibqp.qp_num = qp->mqp.qpn;
1169 /* Userspace is not allowed to create special QPs: */
1171 return ERR_PTR(-EINVAL);
1173 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1174 get_sqp_num(to_mdev(pd->device), init_attr),
1177 return ERR_PTR(err);
1179 qp->port = init_attr->port_num;
1180 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1185 /* Don't support raw QPs */
1186 return ERR_PTR(-EINVAL);
1192 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1194 struct mlx4_ib_dev *dev = to_mdev(qp->device);
1195 struct mlx4_ib_qp *mqp = to_mqp(qp);
1196 struct mlx4_ib_pd *pd;
1198 if (is_qp0(dev, mqp))
1199 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1201 if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1202 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1203 dev->qp1_proxy[mqp->port - 1] = NULL;
1204 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1207 if (mqp->counter_index)
1208 mlx4_ib_free_qp_counter(dev, mqp);
1211 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1213 if (is_sqp(dev, mqp))
1214 kfree(to_msqp(mqp));
1221 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1224 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC;
1225 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC;
1226 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD;
1227 case MLX4_IB_QPT_XRC_INI:
1228 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC;
1229 case MLX4_IB_QPT_SMI:
1230 case MLX4_IB_QPT_GSI:
1231 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX;
1233 case MLX4_IB_QPT_PROXY_SMI_OWNER:
1234 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1235 MLX4_QP_ST_MLX : -1);
1236 case MLX4_IB_QPT_PROXY_SMI:
1237 case MLX4_IB_QPT_TUN_SMI:
1238 case MLX4_IB_QPT_PROXY_GSI:
1239 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ?
1240 MLX4_QP_ST_UD : -1);
1245 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1250 u32 hw_access_flags = 0;
1252 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1253 dest_rd_atomic = attr->max_dest_rd_atomic;
1255 dest_rd_atomic = qp->resp_depth;
1257 if (attr_mask & IB_QP_ACCESS_FLAGS)
1258 access_flags = attr->qp_access_flags;
1260 access_flags = qp->atomic_rd_en;
1262 if (!dest_rd_atomic)
1263 access_flags &= IB_ACCESS_REMOTE_WRITE;
1265 if (access_flags & IB_ACCESS_REMOTE_READ)
1266 hw_access_flags |= MLX4_QP_BIT_RRE;
1267 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1268 hw_access_flags |= MLX4_QP_BIT_RAE;
1269 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1270 hw_access_flags |= MLX4_QP_BIT_RWE;
1272 return cpu_to_be32(hw_access_flags);
1275 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1278 if (attr_mask & IB_QP_PKEY_INDEX)
1279 sqp->pkey_index = attr->pkey_index;
1280 if (attr_mask & IB_QP_QKEY)
1281 sqp->qkey = attr->qkey;
1282 if (attr_mask & IB_QP_SQ_PSN)
1283 sqp->send_psn = attr->sq_psn;
1286 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1288 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1291 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1292 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1293 struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1295 int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1296 IB_LINK_LAYER_ETHERNET;
1302 path->grh_mylmc = ah->src_path_bits & 0x7f;
1303 path->rlid = cpu_to_be16(ah->dlid);
1304 if (ah->static_rate) {
1305 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1306 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1307 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1308 --path->static_rate;
1310 path->static_rate = 0;
1312 if (ah->ah_flags & IB_AH_GRH) {
1313 int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1315 ah->grh.sgid_index);
1317 if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1318 pr_err("sgid_index (%u) too large. max is %d\n",
1319 real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1323 path->grh_mylmc |= 1 << 7;
1324 path->mgid_index = real_sgid_index;
1325 path->hop_limit = ah->grh.hop_limit;
1326 path->tclass_flowlabel =
1327 cpu_to_be32((ah->grh.traffic_class << 20) |
1328 (ah->grh.flow_label));
1329 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1333 if (!(ah->ah_flags & IB_AH_GRH))
1336 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1337 ((port - 1) << 6) | ((ah->sl & 7) << 3);
1339 path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1340 if (vlan_tag < 0x1000) {
1341 if (smac_info->vid < 0x1000) {
1342 /* both valid vlan ids */
1343 if (smac_info->vid != vlan_tag) {
1344 /* different VIDs. unreg old and reg new */
1345 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1348 smac_info->candidate_vid = vlan_tag;
1349 smac_info->candidate_vlan_index = vidx;
1350 smac_info->candidate_vlan_port = port;
1351 smac_info->update_vid = 1;
1352 path->vlan_index = vidx;
1354 path->vlan_index = smac_info->vlan_index;
1357 /* no current vlan tag in qp */
1358 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1361 smac_info->candidate_vid = vlan_tag;
1362 smac_info->candidate_vlan_index = vidx;
1363 smac_info->candidate_vlan_port = port;
1364 smac_info->update_vid = 1;
1365 path->vlan_index = vidx;
1367 path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1370 /* have current vlan tag. unregister it at modify-qp success */
1371 if (smac_info->vid < 0x1000) {
1372 smac_info->candidate_vid = 0xFFFF;
1373 smac_info->update_vid = 1;
1377 /* get smac_index for RoCE use.
1378 * If no smac was yet assigned, register one.
1379 * If one was already assigned, but the new mac differs,
1380 * unregister the old one and register the new one.
1382 if ((!smac_info->smac && !smac_info->smac_port) ||
1383 smac_info->smac != smac) {
1384 /* register candidate now, unreg if needed, after success */
1385 smac_index = mlx4_register_mac(dev->dev, port, smac);
1386 if (smac_index >= 0) {
1387 smac_info->candidate_smac_index = smac_index;
1388 smac_info->candidate_smac = smac;
1389 smac_info->candidate_smac_port = port;
1394 smac_index = smac_info->smac_index;
1397 memcpy(path->dmac, ah->dmac, 6);
1398 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1399 /* put MAC table smac index for IBoE */
1400 path->grh_mylmc = (u8) (smac_index) | 0x80;
1402 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1403 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1409 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1410 enum ib_qp_attr_mask qp_attr_mask,
1411 struct mlx4_ib_qp *mqp,
1412 struct mlx4_qp_path *path, u8 port)
1414 return _mlx4_set_path(dev, &qp->ah_attr,
1415 mlx4_mac_to_u64((u8 *)qp->smac),
1416 (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1417 path, &mqp->pri, port);
1420 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1421 const struct ib_qp_attr *qp,
1422 enum ib_qp_attr_mask qp_attr_mask,
1423 struct mlx4_ib_qp *mqp,
1424 struct mlx4_qp_path *path, u8 port)
1426 return _mlx4_set_path(dev, &qp->alt_ah_attr,
1427 mlx4_mac_to_u64((u8 *)qp->alt_smac),
1428 (qp_attr_mask & IB_QP_ALT_VID) ?
1429 qp->alt_vlan_id : 0xffff,
1430 path, &mqp->alt, port);
1433 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1435 struct mlx4_ib_gid_entry *ge, *tmp;
1437 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1438 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1440 ge->port = qp->port;
1445 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1446 struct mlx4_qp_context *context)
1451 u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1453 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1454 if (!qp->pri.smac && !qp->pri.smac_port) {
1455 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1456 if (smac_index >= 0) {
1457 qp->pri.candidate_smac_index = smac_index;
1458 qp->pri.candidate_smac = u64_mac;
1459 qp->pri.candidate_smac_port = qp->port;
1460 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1468 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1470 struct counter_index *new_counter_index;
1474 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1475 IB_LINK_LAYER_ETHERNET ||
1476 !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1477 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1480 err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1484 new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1485 if (!new_counter_index) {
1486 mlx4_counter_free(dev->dev, tmp_idx);
1490 new_counter_index->index = tmp_idx;
1491 new_counter_index->allocated = 1;
1492 qp->counter_index = new_counter_index;
1494 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1495 list_add_tail(&new_counter_index->list,
1496 &dev->counters_table[qp->port - 1].counters_list);
1497 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1502 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1503 const struct ib_qp_attr *attr, int attr_mask,
1504 enum ib_qp_state cur_state, enum ib_qp_state new_state)
1506 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1507 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1508 struct mlx4_ib_pd *pd;
1509 struct mlx4_ib_cq *send_cq, *recv_cq;
1510 struct mlx4_qp_context *context;
1511 enum mlx4_qp_optpar optpar = 0;
1517 /* APM is not supported under RoCE */
1518 if (attr_mask & IB_QP_ALT_PATH &&
1519 rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1520 IB_LINK_LAYER_ETHERNET)
1523 context = kzalloc(sizeof *context, GFP_KERNEL);
1527 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1528 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1530 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1531 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1533 optpar |= MLX4_QP_OPTPAR_PM_STATE;
1534 switch (attr->path_mig_state) {
1535 case IB_MIG_MIGRATED:
1536 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1539 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1542 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1547 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1548 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1549 else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1550 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1551 else if (ibqp->qp_type == IB_QPT_UD) {
1552 if (qp->flags & MLX4_IB_QP_LSO)
1553 context->mtu_msgmax = (IB_MTU_4096 << 5) |
1554 ilog2(dev->dev->caps.max_gso_sz);
1556 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1557 } else if (attr_mask & IB_QP_PATH_MTU) {
1558 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1559 pr_err("path MTU (%u) is invalid\n",
1563 context->mtu_msgmax = (attr->path_mtu << 5) |
1564 ilog2(dev->dev->caps.max_msg_sz);
1568 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1569 context->rq_size_stride |= qp->rq.wqe_shift - 4;
1572 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1573 context->sq_size_stride |= qp->sq.wqe_shift - 4;
1575 if (new_state == IB_QPS_RESET && qp->counter_index)
1576 mlx4_ib_free_qp_counter(dev, qp);
1578 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1579 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1580 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1581 if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1582 context->param3 |= cpu_to_be32(1 << 30);
1585 if (qp->ibqp.uobject)
1586 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1588 context->usr_page = cpu_to_be32(dev->priv_uar.index);
1590 if (attr_mask & IB_QP_DEST_QPN)
1591 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1593 if (attr_mask & IB_QP_PORT) {
1594 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1595 !(attr_mask & IB_QP_AV)) {
1596 mlx4_set_sched(&context->pri_path, attr->port_num);
1597 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1601 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1602 err = create_qp_lb_counter(dev, qp);
1607 dev->counters_table[qp->port - 1].default_counter;
1608 if (qp->counter_index)
1609 counter_index = qp->counter_index->index;
1611 if (counter_index != -1) {
1612 context->pri_path.counter_index = counter_index;
1613 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1614 if (qp->counter_index) {
1615 context->pri_path.fl |=
1616 MLX4_FL_ETH_SRC_CHECK_MC_LB;
1617 context->pri_path.vlan_control |=
1618 MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1621 context->pri_path.counter_index =
1622 MLX4_SINK_COUNTER_INDEX(dev->dev);
1624 if (qp->flags & MLX4_IB_QP_NETIF) {
1625 mlx4_ib_steer_qp_reg(dev, qp, 1);
1630 if (attr_mask & IB_QP_PKEY_INDEX) {
1631 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1632 context->pri_path.disable_pkey_check = 0x40;
1633 context->pri_path.pkey_index = attr->pkey_index;
1634 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1637 if (attr_mask & IB_QP_AV) {
1638 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1639 attr_mask & IB_QP_PORT ?
1640 attr->port_num : qp->port))
1643 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1644 MLX4_QP_OPTPAR_SCHED_QUEUE);
1647 if (attr_mask & IB_QP_TIMEOUT) {
1648 context->pri_path.ackto |= attr->timeout << 3;
1649 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1652 if (attr_mask & IB_QP_ALT_PATH) {
1653 if (attr->alt_port_num == 0 ||
1654 attr->alt_port_num > dev->dev->caps.num_ports)
1657 if (attr->alt_pkey_index >=
1658 dev->dev->caps.pkey_table_len[attr->alt_port_num])
1661 if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1663 attr->alt_port_num))
1666 context->alt_path.pkey_index = attr->alt_pkey_index;
1667 context->alt_path.ackto = attr->alt_timeout << 3;
1668 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1672 get_cqs(qp, &send_cq, &recv_cq);
1673 context->pd = cpu_to_be32(pd->pdn);
1674 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1675 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1676 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1678 /* Set "fast registration enabled" for all kernel QPs */
1679 if (!qp->ibqp.uobject)
1680 context->params1 |= cpu_to_be32(1 << 11);
1682 if (attr_mask & IB_QP_RNR_RETRY) {
1683 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1684 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1687 if (attr_mask & IB_QP_RETRY_CNT) {
1688 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1689 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1692 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1693 if (attr->max_rd_atomic)
1695 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1696 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1699 if (attr_mask & IB_QP_SQ_PSN)
1700 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1702 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1703 if (attr->max_dest_rd_atomic)
1705 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1706 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1709 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1710 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1711 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1715 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1717 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1718 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1719 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1721 if (attr_mask & IB_QP_RQ_PSN)
1722 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1724 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1725 if (attr_mask & IB_QP_QKEY) {
1726 if (qp->mlx4_ib_qp_type &
1727 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1728 context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1730 if (mlx4_is_mfunc(dev->dev) &&
1731 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1732 (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1733 MLX4_RESERVED_QKEY_BASE) {
1734 pr_err("Cannot use reserved QKEY"
1735 " 0x%x (range 0xffff0000..0xffffffff"
1736 " is reserved)\n", attr->qkey);
1740 context->qkey = cpu_to_be32(attr->qkey);
1742 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1746 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1748 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1749 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1751 if (cur_state == IB_QPS_INIT &&
1752 new_state == IB_QPS_RTR &&
1753 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1754 ibqp->qp_type == IB_QPT_UD ||
1755 ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1756 context->pri_path.sched_queue = (qp->port - 1) << 6;
1757 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1758 qp->mlx4_ib_qp_type &
1759 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1760 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1761 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1762 context->pri_path.fl = 0x80;
1764 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1765 context->pri_path.fl = 0x80;
1766 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1768 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1769 IB_LINK_LAYER_ETHERNET) {
1770 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1771 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1772 context->pri_path.feup = 1 << 7; /* don't fsm */
1773 /* handle smac_index */
1774 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1775 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1776 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1777 err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1782 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1783 dev->qp1_proxy[qp->port - 1] = qp;
1788 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1789 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1790 MLX4_IB_LINK_TYPE_ETH;
1791 if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1792 /* set QP to receive both tunneled & non-tunneled packets */
1793 if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1794 context->srqn = cpu_to_be32(7 << 28);
1798 if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1799 int is_eth = rdma_port_get_link_layer(
1800 &dev->ib_dev, qp->port) ==
1801 IB_LINK_LAYER_ETHERNET;
1803 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1804 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1809 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
1810 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1815 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1816 context->rlkey |= (1 << 4);
1819 * Before passing a kernel QP to the HW, make sure that the
1820 * ownership bits of the send queue are set and the SQ
1821 * headroom is stamped so that the hardware doesn't start
1822 * processing stale work requests.
1824 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1825 struct mlx4_wqe_ctrl_seg *ctrl;
1828 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1829 ctrl = get_send_wqe(qp, i);
1830 ctrl->owner_opcode = cpu_to_be32(1 << 31);
1831 if (qp->sq_max_wqes_per_wr == 1)
1832 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1834 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1838 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1839 to_mlx4_state(new_state), context, optpar,
1840 sqd_event, &qp->mqp);
1844 qp->state = new_state;
1846 if (attr_mask & IB_QP_ACCESS_FLAGS)
1847 qp->atomic_rd_en = attr->qp_access_flags;
1848 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1849 qp->resp_depth = attr->max_dest_rd_atomic;
1850 if (attr_mask & IB_QP_PORT) {
1851 qp->port = attr->port_num;
1852 update_mcg_macs(dev, qp);
1854 if (attr_mask & IB_QP_ALT_PATH)
1855 qp->alt_port = attr->alt_port_num;
1857 if (is_sqp(dev, qp))
1858 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1861 * If we moved QP0 to RTR, bring the IB link up; if we moved
1862 * QP0 to RESET or ERROR, bring the link back down.
1864 if (is_qp0(dev, qp)) {
1865 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1866 if (mlx4_INIT_PORT(dev->dev, qp->port))
1867 pr_warn("INIT_PORT failed for port %d\n",
1870 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1871 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1872 mlx4_CLOSE_PORT(dev->dev, qp->port);
1876 * If we moved a kernel QP to RESET, clean up all old CQ
1877 * entries and reinitialize the QP.
1879 if (new_state == IB_QPS_RESET) {
1880 if (!ibqp->uobject) {
1881 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1882 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1883 if (send_cq != recv_cq)
1884 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1890 qp->sq_next_wqe = 0;
1894 if (qp->flags & MLX4_IB_QP_NETIF)
1895 mlx4_ib_steer_qp_reg(dev, qp, 0);
1897 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1898 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1900 qp->pri.smac_port = 0;
1903 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1906 if (qp->pri.vid < 0x1000) {
1907 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1908 qp->pri.vid = 0xFFFF;
1909 qp->pri.candidate_vid = 0xFFFF;
1910 qp->pri.update_vid = 0;
1913 if (qp->alt.vid < 0x1000) {
1914 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1915 qp->alt.vid = 0xFFFF;
1916 qp->alt.candidate_vid = 0xFFFF;
1917 qp->alt.update_vid = 0;
1921 if (err && qp->counter_index)
1922 mlx4_ib_free_qp_counter(dev, qp);
1923 if (err && steer_qp)
1924 mlx4_ib_steer_qp_reg(dev, qp, 0);
1926 if (qp->pri.candidate_smac ||
1927 (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
1929 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1931 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
1932 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1933 qp->pri.smac = qp->pri.candidate_smac;
1934 qp->pri.smac_index = qp->pri.candidate_smac_index;
1935 qp->pri.smac_port = qp->pri.candidate_smac_port;
1937 qp->pri.candidate_smac = 0;
1938 qp->pri.candidate_smac_index = 0;
1939 qp->pri.candidate_smac_port = 0;
1941 if (qp->alt.candidate_smac) {
1943 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1946 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1947 qp->alt.smac = qp->alt.candidate_smac;
1948 qp->alt.smac_index = qp->alt.candidate_smac_index;
1949 qp->alt.smac_port = qp->alt.candidate_smac_port;
1951 qp->alt.candidate_smac = 0;
1952 qp->alt.candidate_smac_index = 0;
1953 qp->alt.candidate_smac_port = 0;
1956 if (qp->pri.update_vid) {
1958 if (qp->pri.candidate_vid < 0x1000)
1959 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1960 qp->pri.candidate_vid);
1962 if (qp->pri.vid < 0x1000)
1963 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1965 qp->pri.vid = qp->pri.candidate_vid;
1966 qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1967 qp->pri.vlan_index = qp->pri.candidate_vlan_index;
1969 qp->pri.candidate_vid = 0xFFFF;
1970 qp->pri.update_vid = 0;
1973 if (qp->alt.update_vid) {
1975 if (qp->alt.candidate_vid < 0x1000)
1976 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1977 qp->alt.candidate_vid);
1979 if (qp->alt.vid < 0x1000)
1980 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1982 qp->alt.vid = qp->alt.candidate_vid;
1983 qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1984 qp->alt.vlan_index = qp->alt.candidate_vlan_index;
1986 qp->alt.candidate_vid = 0xFFFF;
1987 qp->alt.update_vid = 0;
1993 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1994 int attr_mask, struct ib_udata *udata)
1996 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1997 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1998 enum ib_qp_state cur_state, new_state;
2001 mutex_lock(&qp->mutex);
2003 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2004 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2006 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2007 ll = IB_LINK_LAYER_UNSPECIFIED;
2009 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2010 ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2013 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2015 pr_debug("qpn 0x%x: invalid attribute mask specified "
2016 "for transition %d to %d. qp_type %d,"
2017 " attr_mask 0x%x\n",
2018 ibqp->qp_num, cur_state, new_state,
2019 ibqp->qp_type, attr_mask);
2023 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2024 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2025 if ((ibqp->qp_type == IB_QPT_RC) ||
2026 (ibqp->qp_type == IB_QPT_UD) ||
2027 (ibqp->qp_type == IB_QPT_UC) ||
2028 (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2029 (ibqp->qp_type == IB_QPT_XRC_INI)) {
2030 attr->port_num = mlx4_ib_bond_next_port(dev);
2033 /* no sense in changing port_num
2034 * when ports are bonded */
2035 attr_mask &= ~IB_QP_PORT;
2039 if ((attr_mask & IB_QP_PORT) &&
2040 (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2041 pr_debug("qpn 0x%x: invalid port number (%d) specified "
2042 "for transition %d to %d. qp_type %d\n",
2043 ibqp->qp_num, attr->port_num, cur_state,
2044 new_state, ibqp->qp_type);
2048 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2049 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2050 IB_LINK_LAYER_ETHERNET))
2053 if (attr_mask & IB_QP_PKEY_INDEX) {
2054 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2055 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2056 pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2057 "for transition %d to %d. qp_type %d\n",
2058 ibqp->qp_num, attr->pkey_index, cur_state,
2059 new_state, ibqp->qp_type);
2064 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2065 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2066 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2067 "Transition %d to %d. qp_type %d\n",
2068 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2069 new_state, ibqp->qp_type);
2073 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2074 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2075 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2076 "Transition %d to %d. qp_type %d\n",
2077 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2078 new_state, ibqp->qp_type);
2082 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2087 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2089 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2093 mutex_unlock(&qp->mutex);
2097 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2100 for (i = 0; i < dev->caps.num_ports; i++) {
2101 if (qpn == dev->caps.qp0_proxy[i] ||
2102 qpn == dev->caps.qp0_tunnel[i]) {
2103 *qkey = dev->caps.qp0_qkey[i];
2110 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2111 struct ib_send_wr *wr,
2112 void *wqe, unsigned *mlx_seg_len)
2114 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2115 struct ib_device *ib_dev = &mdev->ib_dev;
2116 struct mlx4_wqe_mlx_seg *mlx = wqe;
2117 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2118 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2126 if (wr->opcode != IB_WR_SEND)
2131 for (i = 0; i < wr->num_sge; ++i)
2132 send_size += wr->sg_list[i].length;
2134 /* for proxy-qp0 sends, need to add in size of tunnel header */
2135 /* for tunnel-qp0 sends, tunnel header is already in s/g list */
2136 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2137 send_size += sizeof (struct mlx4_ib_tunnel_header);
2139 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2141 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2142 sqp->ud_header.lrh.service_level =
2143 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2144 sqp->ud_header.lrh.destination_lid =
2145 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2146 sqp->ud_header.lrh.source_lid =
2147 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2150 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2152 /* force loopback */
2153 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2154 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2156 sqp->ud_header.lrh.virtual_lane = 0;
2157 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2158 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2159 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2160 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2161 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2163 sqp->ud_header.bth.destination_qpn =
2164 cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2166 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2167 if (mlx4_is_master(mdev->dev)) {
2168 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2171 if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2174 sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2175 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2177 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2178 sqp->ud_header.immediate_present = 0;
2180 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2183 * Inline data segments may not cross a 64 byte boundary. If
2184 * our UD header is bigger than the space available up to the
2185 * next 64 byte boundary in the WQE, use two inline data
2186 * segments to hold the UD header.
2188 spc = MLX4_INLINE_ALIGN -
2189 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2190 if (header_size <= spc) {
2191 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2192 memcpy(inl + 1, sqp->header_buf, header_size);
2195 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2196 memcpy(inl + 1, sqp->header_buf, spc);
2198 inl = (void *) (inl + 1) + spc;
2199 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2201 * Need a barrier here to make sure all the data is
2202 * visible before the byte_count field is set.
2203 * Otherwise the HCA prefetcher could grab the 64-byte
2204 * chunk with this inline segment and get a valid (!=
2205 * 0xffffffff) byte count but stale data, and end up
2206 * generating a packet with bad headers.
2208 * The first inline segment's byte_count field doesn't
2209 * need a barrier, because it comes after a
2210 * control/MLX segment and therefore is at an offset
2214 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2219 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2223 static void mlx4_u64_to_smac(u8 *dst_mac, u64 src_mac)
2227 for (i = ETH_ALEN; i; i--) {
2228 dst_mac[i - 1] = src_mac & 0xff;
2233 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2234 void *wqe, unsigned *mlx_seg_len)
2236 struct ib_device *ib_dev = sqp->qp.ibqp.device;
2237 struct mlx4_wqe_mlx_seg *mlx = wqe;
2238 struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2239 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2240 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2250 bool is_vlan = false;
2254 for (i = 0; i < wr->num_sge; ++i)
2255 send_size += wr->sg_list[i].length;
2257 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2258 is_grh = mlx4_ib_ah_grh_present(ah);
2260 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2261 /* When multi-function is enabled, the ib_core gid
2262 * indexes don't necessarily match the hw ones, so
2263 * we must use our own cache */
2264 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2265 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2266 ah->av.ib.gid_index, &sgid.raw[0]);
2270 err = ib_get_cached_gid(ib_dev,
2271 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2272 ah->av.ib.gid_index, &sgid,
2278 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2279 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2283 ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2286 sqp->ud_header.lrh.service_level =
2287 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2288 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2289 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2293 sqp->ud_header.grh.traffic_class =
2294 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2295 sqp->ud_header.grh.flow_label =
2296 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2297 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit;
2299 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2301 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2302 /* When multi-function is enabled, the ib_core gid
2303 * indexes don't necessarily match the hw ones, so
2304 * we must use our own cache */
2305 sqp->ud_header.grh.source_gid.global.subnet_prefix =
2306 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2308 sqp->ud_header.grh.source_gid.global.interface_id =
2309 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2310 guid_cache[ah->av.ib.gid_index];
2312 ib_get_cached_gid(ib_dev,
2313 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2314 ah->av.ib.gid_index,
2315 &sqp->ud_header.grh.source_gid, NULL);
2317 memcpy(sqp->ud_header.grh.destination_gid.raw,
2318 ah->av.ib.dgid, 16);
2321 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2324 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2325 (sqp->ud_header.lrh.destination_lid ==
2326 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2327 (sqp->ud_header.lrh.service_level << 8));
2328 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2329 mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2330 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2333 switch (wr->opcode) {
2335 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2336 sqp->ud_header.immediate_present = 0;
2338 case IB_WR_SEND_WITH_IMM:
2339 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2340 sqp->ud_header.immediate_present = 1;
2341 sqp->ud_header.immediate_data = wr->ex.imm_data;
2348 struct in6_addr in6;
2350 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2352 mlx->sched_prio = cpu_to_be16(pcp);
2354 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2355 /* FIXME: cache smac value? */
2356 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2357 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2358 memcpy(&in6, sgid.raw, sizeof(in6));
2360 if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2361 u64 mac = atomic64_read(&to_mdev(ib_dev)->iboe.mac[sqp->qp.port - 1]);
2364 mlx4_u64_to_smac(smac, mac);
2365 memcpy(sqp->ud_header.eth.smac_h, smac, ETH_ALEN);
2367 /* use the src mac of the tunnel */
2368 memcpy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac, ETH_ALEN);
2371 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2372 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2374 sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2376 sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2377 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2380 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
2381 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2382 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2384 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2385 if (!sqp->qp.ibqp.qp_num)
2386 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2388 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2389 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2390 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2391 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2392 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2393 sqp->qkey : wr->wr.ud.remote_qkey);
2394 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2396 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2399 pr_err("built UD header of size %d:\n", header_size);
2400 for (i = 0; i < header_size / 4; ++i) {
2402 pr_err(" [%02x] ", i * 4);
2404 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2405 if ((i + 1) % 8 == 0)
2412 * Inline data segments may not cross a 64 byte boundary. If
2413 * our UD header is bigger than the space available up to the
2414 * next 64 byte boundary in the WQE, use two inline data
2415 * segments to hold the UD header.
2417 spc = MLX4_INLINE_ALIGN -
2418 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2419 if (header_size <= spc) {
2420 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2421 memcpy(inl + 1, sqp->header_buf, header_size);
2424 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2425 memcpy(inl + 1, sqp->header_buf, spc);
2427 inl = (void *) (inl + 1) + spc;
2428 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2430 * Need a barrier here to make sure all the data is
2431 * visible before the byte_count field is set.
2432 * Otherwise the HCA prefetcher could grab the 64-byte
2433 * chunk with this inline segment and get a valid (!=
2434 * 0xffffffff) byte count but stale data, and end up
2435 * generating a packet with bad headers.
2437 * The first inline segment's byte_count field doesn't
2438 * need a barrier, because it comes after a
2439 * control/MLX segment and therefore is at an offset
2443 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2448 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2452 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2455 struct mlx4_ib_cq *cq;
2457 cur = wq->head - wq->tail;
2458 if (likely(cur + nreq < wq->max_post))
2462 spin_lock(&cq->lock);
2463 cur = wq->head - wq->tail;
2464 spin_unlock(&cq->lock);
2466 return cur + nreq >= wq->max_post;
2469 static __be32 convert_access(int acc)
2471 return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2472 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) |
2473 (acc & IB_ACCESS_REMOTE_WRITE ?
2474 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2475 (acc & IB_ACCESS_REMOTE_READ ?
2476 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) |
2477 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) |
2478 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2481 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2483 struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2486 for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2487 mfrpl->mapped_page_list[i] =
2488 cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2489 MLX4_MTT_FLAG_PRESENT);
2491 fseg->flags = convert_access(wr->wr.fast_reg.access_flags);
2492 fseg->mem_key = cpu_to_be32(wr->wr.fast_reg.rkey);
2493 fseg->buf_list = cpu_to_be64(mfrpl->map);
2494 fseg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2495 fseg->reg_len = cpu_to_be64(wr->wr.fast_reg.length);
2496 fseg->offset = 0; /* XXX -- is this just for ZBVA? */
2497 fseg->page_size = cpu_to_be32(wr->wr.fast_reg.page_shift);
2498 fseg->reserved[0] = 0;
2499 fseg->reserved[1] = 0;
2502 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2505 convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2506 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ |
2507 MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2508 MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2510 if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2511 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2512 if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2513 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2514 bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2515 bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2516 bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2517 bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2520 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2522 memset(iseg, 0, sizeof(*iseg));
2523 iseg->mem_key = cpu_to_be32(rkey);
2526 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2527 u64 remote_addr, u32 rkey)
2529 rseg->raddr = cpu_to_be64(remote_addr);
2530 rseg->rkey = cpu_to_be32(rkey);
2534 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2536 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2537 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2538 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add);
2539 } else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2540 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2541 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2543 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2549 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2550 struct ib_send_wr *wr)
2552 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2553 aseg->swap_add_mask = cpu_to_be64(wr->wr.atomic.swap_mask);
2554 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add);
2555 aseg->compare_mask = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2558 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2559 struct ib_send_wr *wr)
2561 memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2562 dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2563 dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2564 dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2565 memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2568 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2569 struct mlx4_wqe_datagram_seg *dseg,
2570 struct ib_send_wr *wr,
2571 enum mlx4_ib_qp_type qpt)
2573 union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2574 struct mlx4_av sqp_av = {0};
2575 int port = *((u8 *) &av->ib.port_pd) & 0x3;
2577 /* force loopback */
2578 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2579 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2580 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2581 cpu_to_be32(0xf0000000);
2583 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2584 if (qpt == MLX4_IB_QPT_PROXY_GSI)
2585 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2587 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2588 /* Use QKEY from the QP context, which is set by master */
2589 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2592 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2594 struct mlx4_wqe_inline_seg *inl = wqe;
2595 struct mlx4_ib_tunnel_header hdr;
2596 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2600 memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2601 hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2602 hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2603 hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2604 memcpy(hdr.mac, ah->av.eth.mac, 6);
2605 hdr.vlan = ah->av.eth.vlan;
2607 spc = MLX4_INLINE_ALIGN -
2608 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2609 if (sizeof (hdr) <= spc) {
2610 memcpy(inl + 1, &hdr, sizeof (hdr));
2612 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2615 memcpy(inl + 1, &hdr, spc);
2617 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2619 inl = (void *) (inl + 1) + spc;
2620 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2622 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2627 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2630 static void set_mlx_icrc_seg(void *dseg)
2633 struct mlx4_wqe_inline_seg *iseg = dseg;
2638 * Need a barrier here before writing the byte_count field to
2639 * make sure that all the data is visible before the
2640 * byte_count field is set. Otherwise, if the segment begins
2641 * a new cacheline, the HCA prefetcher could grab the 64-byte
2642 * chunk and get a valid (!= * 0xffffffff) byte count but
2643 * stale data, and end up sending the wrong data.
2647 iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2650 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2652 dseg->lkey = cpu_to_be32(sg->lkey);
2653 dseg->addr = cpu_to_be64(sg->addr);
2656 * Need a barrier here before writing the byte_count field to
2657 * make sure that all the data is visible before the
2658 * byte_count field is set. Otherwise, if the segment begins
2659 * a new cacheline, the HCA prefetcher could grab the 64-byte
2660 * chunk and get a valid (!= * 0xffffffff) byte count but
2661 * stale data, and end up sending the wrong data.
2665 dseg->byte_count = cpu_to_be32(sg->length);
2668 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2670 dseg->byte_count = cpu_to_be32(sg->length);
2671 dseg->lkey = cpu_to_be32(sg->lkey);
2672 dseg->addr = cpu_to_be64(sg->addr);
2675 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2676 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2677 __be32 *lso_hdr_sz, __be32 *blh)
2679 unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2681 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2682 *blh = cpu_to_be32(1 << 6);
2684 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2685 wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2688 memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2690 *lso_hdr_sz = cpu_to_be32(wr->wr.ud.mss << 16 | wr->wr.ud.hlen);
2691 *lso_seg_len = halign;
2695 static __be32 send_ieth(struct ib_send_wr *wr)
2697 switch (wr->opcode) {
2698 case IB_WR_SEND_WITH_IMM:
2699 case IB_WR_RDMA_WRITE_WITH_IMM:
2700 return wr->ex.imm_data;
2702 case IB_WR_SEND_WITH_INV:
2703 return cpu_to_be32(wr->ex.invalidate_rkey);
2710 static void add_zero_len_inline(void *wqe)
2712 struct mlx4_wqe_inline_seg *inl = wqe;
2714 inl->byte_count = cpu_to_be32(1 << 31);
2717 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2718 struct ib_send_wr **bad_wr)
2720 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2722 struct mlx4_wqe_ctrl_seg *ctrl;
2723 struct mlx4_wqe_data_seg *dseg;
2724 unsigned long flags;
2728 int uninitialized_var(stamp);
2729 int uninitialized_var(size);
2730 unsigned uninitialized_var(seglen);
2733 __be32 uninitialized_var(lso_hdr_sz);
2736 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2738 spin_lock_irqsave(&qp->sq.lock, flags);
2739 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2746 ind = qp->sq_next_wqe;
2748 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2752 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2758 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2764 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2765 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2768 (wr->send_flags & IB_SEND_SIGNALED ?
2769 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2770 (wr->send_flags & IB_SEND_SOLICITED ?
2771 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2772 ((wr->send_flags & IB_SEND_IP_CSUM) ?
2773 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2774 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2777 ctrl->imm = send_ieth(wr);
2779 wqe += sizeof *ctrl;
2780 size = sizeof *ctrl / 16;
2782 switch (qp->mlx4_ib_qp_type) {
2783 case MLX4_IB_QPT_RC:
2784 case MLX4_IB_QPT_UC:
2785 switch (wr->opcode) {
2786 case IB_WR_ATOMIC_CMP_AND_SWP:
2787 case IB_WR_ATOMIC_FETCH_AND_ADD:
2788 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2789 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2790 wr->wr.atomic.rkey);
2791 wqe += sizeof (struct mlx4_wqe_raddr_seg);
2793 set_atomic_seg(wqe, wr);
2794 wqe += sizeof (struct mlx4_wqe_atomic_seg);
2796 size += (sizeof (struct mlx4_wqe_raddr_seg) +
2797 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2801 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2802 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2803 wr->wr.atomic.rkey);
2804 wqe += sizeof (struct mlx4_wqe_raddr_seg);
2806 set_masked_atomic_seg(wqe, wr);
2807 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg);
2809 size += (sizeof (struct mlx4_wqe_raddr_seg) +
2810 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2814 case IB_WR_RDMA_READ:
2815 case IB_WR_RDMA_WRITE:
2816 case IB_WR_RDMA_WRITE_WITH_IMM:
2817 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2819 wqe += sizeof (struct mlx4_wqe_raddr_seg);
2820 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2823 case IB_WR_LOCAL_INV:
2824 ctrl->srcrb_flags |=
2825 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2826 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2827 wqe += sizeof (struct mlx4_wqe_local_inval_seg);
2828 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2831 case IB_WR_FAST_REG_MR:
2832 ctrl->srcrb_flags |=
2833 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2834 set_fmr_seg(wqe, wr);
2835 wqe += sizeof (struct mlx4_wqe_fmr_seg);
2836 size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2840 ctrl->srcrb_flags |=
2841 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2842 set_bind_seg(wqe, wr);
2843 wqe += sizeof(struct mlx4_wqe_bind_seg);
2844 size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2847 /* No extra segments required for sends */
2852 case MLX4_IB_QPT_TUN_SMI_OWNER:
2853 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2854 if (unlikely(err)) {
2859 size += seglen / 16;
2861 case MLX4_IB_QPT_TUN_SMI:
2862 case MLX4_IB_QPT_TUN_GSI:
2863 /* this is a UD qp used in MAD responses to slaves. */
2864 set_datagram_seg(wqe, wr);
2865 /* set the forced-loopback bit in the data seg av */
2866 *(__be32 *) wqe |= cpu_to_be32(0x80000000);
2867 wqe += sizeof (struct mlx4_wqe_datagram_seg);
2868 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2870 case MLX4_IB_QPT_UD:
2871 set_datagram_seg(wqe, wr);
2872 wqe += sizeof (struct mlx4_wqe_datagram_seg);
2873 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2875 if (wr->opcode == IB_WR_LSO) {
2876 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2877 if (unlikely(err)) {
2881 lso_wqe = (__be32 *) wqe;
2883 size += seglen / 16;
2887 case MLX4_IB_QPT_PROXY_SMI_OWNER:
2888 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2889 if (unlikely(err)) {
2894 size += seglen / 16;
2895 /* to start tunnel header on a cache-line boundary */
2896 add_zero_len_inline(wqe);
2899 build_tunnel_header(wr, wqe, &seglen);
2901 size += seglen / 16;
2903 case MLX4_IB_QPT_PROXY_SMI:
2904 case MLX4_IB_QPT_PROXY_GSI:
2905 /* If we are tunneling special qps, this is a UD qp.
2906 * In this case we first add a UD segment targeting
2907 * the tunnel qp, and then add a header with address
2909 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr,
2910 qp->mlx4_ib_qp_type);
2911 wqe += sizeof (struct mlx4_wqe_datagram_seg);
2912 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2913 build_tunnel_header(wr, wqe, &seglen);
2915 size += seglen / 16;
2918 case MLX4_IB_QPT_SMI:
2919 case MLX4_IB_QPT_GSI:
2920 err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2921 if (unlikely(err)) {
2926 size += seglen / 16;
2934 * Write data segments in reverse order, so as to
2935 * overwrite cacheline stamp last within each
2936 * cacheline. This avoids issues with WQE
2941 dseg += wr->num_sge - 1;
2942 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2944 /* Add one more inline data segment for ICRC for MLX sends */
2945 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2946 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2947 qp->mlx4_ib_qp_type &
2948 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2949 set_mlx_icrc_seg(dseg + 1);
2950 size += sizeof (struct mlx4_wqe_data_seg) / 16;
2953 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2954 set_data_seg(dseg, wr->sg_list + i);
2957 * Possibly overwrite stamping in cacheline with LSO
2958 * segment only after making sure all data segments
2962 *lso_wqe = lso_hdr_sz;
2964 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2965 MLX4_WQE_CTRL_FENCE : 0) | size;
2968 * Make sure descriptor is fully written before
2969 * setting ownership bit (because HW can start
2970 * executing as soon as we do).
2974 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2980 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2981 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2983 stamp = ind + qp->sq_spare_wqes;
2984 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2987 * We can improve latency by not stamping the last
2988 * send queue WQE until after ringing the doorbell, so
2989 * only stamp here if there are still more WQEs to post.
2991 * Same optimization applies to padding with NOP wqe
2992 * in case of WQE shrinking (used to prevent wrap-around
2993 * in the middle of WR).
2996 stamp_send_wqe(qp, stamp, size * 16);
2997 ind = pad_wraparound(qp, ind);
3003 qp->sq.head += nreq;
3006 * Make sure that descriptors are written before
3011 writel(qp->doorbell_qpn,
3012 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3015 * Make sure doorbells don't leak out of SQ spinlock
3016 * and reach the HCA out of order.
3020 stamp_send_wqe(qp, stamp, size * 16);
3022 ind = pad_wraparound(qp, ind);
3023 qp->sq_next_wqe = ind;
3026 spin_unlock_irqrestore(&qp->sq.lock, flags);
3031 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3032 struct ib_recv_wr **bad_wr)
3034 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3035 struct mlx4_wqe_data_seg *scat;
3036 unsigned long flags;
3042 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3044 max_gs = qp->rq.max_gs;
3045 spin_lock_irqsave(&qp->rq.lock, flags);
3047 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3054 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3056 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3057 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3063 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3069 scat = get_recv_wqe(qp, ind);
3071 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3072 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3073 ib_dma_sync_single_for_device(ibqp->device,
3074 qp->sqp_proxy_rcv[ind].map,
3075 sizeof (struct mlx4_ib_proxy_sqp_hdr),
3078 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3079 /* use dma lkey from upper layer entry */
3080 scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3081 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3086 for (i = 0; i < wr->num_sge; ++i)
3087 __set_data_seg(scat + i, wr->sg_list + i);
3090 scat[i].byte_count = 0;
3091 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
3095 qp->rq.wrid[ind] = wr->wr_id;
3097 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3102 qp->rq.head += nreq;
3105 * Make sure that descriptors are written before
3110 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3113 spin_unlock_irqrestore(&qp->rq.lock, flags);
3118 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3120 switch (mlx4_state) {
3121 case MLX4_QP_STATE_RST: return IB_QPS_RESET;
3122 case MLX4_QP_STATE_INIT: return IB_QPS_INIT;
3123 case MLX4_QP_STATE_RTR: return IB_QPS_RTR;
3124 case MLX4_QP_STATE_RTS: return IB_QPS_RTS;
3125 case MLX4_QP_STATE_SQ_DRAINING:
3126 case MLX4_QP_STATE_SQD: return IB_QPS_SQD;
3127 case MLX4_QP_STATE_SQER: return IB_QPS_SQE;
3128 case MLX4_QP_STATE_ERR: return IB_QPS_ERR;
3133 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3135 switch (mlx4_mig_state) {
3136 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED;
3137 case MLX4_QP_PM_REARM: return IB_MIG_REARM;
3138 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
3143 static int to_ib_qp_access_flags(int mlx4_flags)
3147 if (mlx4_flags & MLX4_QP_BIT_RRE)
3148 ib_flags |= IB_ACCESS_REMOTE_READ;
3149 if (mlx4_flags & MLX4_QP_BIT_RWE)
3150 ib_flags |= IB_ACCESS_REMOTE_WRITE;
3151 if (mlx4_flags & MLX4_QP_BIT_RAE)
3152 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3157 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3158 struct mlx4_qp_path *path)
3160 struct mlx4_dev *dev = ibdev->dev;
3163 memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3164 ib_ah_attr->port_num = path->sched_queue & 0x40 ? 2 : 1;
3166 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3169 is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3170 IB_LINK_LAYER_ETHERNET;
3172 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3173 ((path->sched_queue & 4) << 1);
3175 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3177 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
3178 ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3179 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
3180 ib_ah_attr->ah_flags = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3181 if (ib_ah_attr->ah_flags) {
3182 ib_ah_attr->grh.sgid_index = path->mgid_index;
3183 ib_ah_attr->grh.hop_limit = path->hop_limit;
3184 ib_ah_attr->grh.traffic_class =
3185 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3186 ib_ah_attr->grh.flow_label =
3187 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3188 memcpy(ib_ah_attr->grh.dgid.raw,
3189 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3193 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3194 struct ib_qp_init_attr *qp_init_attr)
3196 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3197 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3198 struct mlx4_qp_context context;
3202 mutex_lock(&qp->mutex);
3204 if (qp->state == IB_QPS_RESET) {
3205 qp_attr->qp_state = IB_QPS_RESET;
3209 err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3215 mlx4_state = be32_to_cpu(context.flags) >> 28;
3217 qp->state = to_ib_qp_state(mlx4_state);
3218 qp_attr->qp_state = qp->state;
3219 qp_attr->path_mtu = context.mtu_msgmax >> 5;
3220 qp_attr->path_mig_state =
3221 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3222 qp_attr->qkey = be32_to_cpu(context.qkey);
3223 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3224 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff;
3225 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff;
3226 qp_attr->qp_access_flags =
3227 to_ib_qp_access_flags(be32_to_cpu(context.params2));
3229 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3230 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3231 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3232 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3233 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
3236 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3237 if (qp_attr->qp_state == IB_QPS_INIT)
3238 qp_attr->port_num = qp->port;
3240 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3242 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3243 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3245 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3247 qp_attr->max_dest_rd_atomic =
3248 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3249 qp_attr->min_rnr_timer =
3250 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3251 qp_attr->timeout = context.pri_path.ackto >> 3;
3252 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7;
3253 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7;
3254 qp_attr->alt_timeout = context.alt_path.ackto >> 3;
3257 qp_attr->cur_qp_state = qp_attr->qp_state;
3258 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
3259 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
3261 if (!ibqp->uobject) {
3262 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
3263 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3265 qp_attr->cap.max_send_wr = 0;
3266 qp_attr->cap.max_send_sge = 0;
3270 * We don't support inline sends for kernel QPs (yet), and we
3271 * don't know what userspace's value should be.
3273 qp_attr->cap.max_inline_data = 0;
3275 qp_init_attr->cap = qp_attr->cap;
3277 qp_init_attr->create_flags = 0;
3278 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3279 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3281 if (qp->flags & MLX4_IB_QP_LSO)
3282 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3284 if (qp->flags & MLX4_IB_QP_NETIF)
3285 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3287 qp_init_attr->sq_sig_type =
3288 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3289 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3292 mutex_unlock(&qp->mutex);