2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/log2.h>
35 #include <linux/slab.h>
36 #include <linux/netdevice.h>
37 #include <linux/bitops.h>
39 #include <rdma/ib_cache.h>
40 #include <rdma/ib_pack.h>
41 #include <rdma/ib_addr.h>
42 #include <rdma/ib_mad.h>
44 #include <linux/mlx4/qp.h>
45 #include <linux/mlx4/driver.h>
56 MLX4_IB_ACK_REQ_FREQ = 8,
60 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83,
61 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
62 MLX4_IB_LINK_TYPE_IB = 0,
63 MLX4_IB_LINK_TYPE_ETH = 1
68 * Largest possible UD header: send with GRH and immediate
69 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
70 * tag. (LRH would only use 8 bytes, so Ethernet is the
73 MLX4_IB_UD_HEADER_SIZE = 82,
74 MLX4_IB_LSO_HEADER_SPARE = 128,
78 MLX4_IB_IBOE_ETHERTYPE = 0x8915
86 struct ib_ud_header ud_header;
87 u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
91 MLX4_IB_MIN_SQ_STRIDE = 6,
92 MLX4_IB_CACHE_LINE_SIZE = 64,
97 MLX4_RAW_QP_MSGMAX = 31,
100 static const __be32 mlx4_ib_opcode[] = {
101 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND),
102 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO),
103 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
104 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
105 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
106 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
107 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
108 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
109 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
110 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
111 [IB_WR_FAST_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR),
112 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
113 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
117 #if defined(__i386__)
118 #define wc_wmb() asm volatile("lock; addl $0,0(%%esp) " ::: "memory")
119 #elif defined(__x86_64__)
120 #define wc_wmb() asm volatile("sfence" ::: "memory")
121 #elif defined(__ia64__)
122 #define wc_wmb() asm volatile("fwb" ::: "memory")
124 #define wc_wmb() wmb()
128 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
130 return container_of(mqp, struct mlx4_ib_sqp, qp);
133 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
135 if (!mlx4_is_master(dev->dev))
138 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
139 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
143 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
148 /* PPF or Native -- real SQP */
149 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
150 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
151 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
154 /* VF or PF -- proxy SQP */
155 if (mlx4_is_mfunc(dev->dev)) {
156 for (i = 0; i < dev->dev->caps.num_ports; i++) {
157 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
158 qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
167 /* used for INIT/CLOSE port logic */
168 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
173 /* PPF or Native -- real QP0 */
174 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
175 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
176 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
179 /* VF or PF -- proxy QP0 */
180 if (mlx4_is_mfunc(dev->dev)) {
181 for (i = 0; i < dev->dev->caps.num_ports; i++) {
182 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
191 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
193 return mlx4_buf_offset(&qp->buf, offset);
196 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
198 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
201 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
203 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
207 * Stamp a SQ WQE so that it is invalid if prefetched by marking the
208 * first four bytes of every 64 byte chunk with
209 * 0x7FFFFFF | (invalid_ownership_value << 31).
211 * When the max work request size is less than or equal to the WQE
212 * basic block size, as an optimization, we can stamp all WQEs with
213 * 0xffffffff, and skip the very first chunk of each WQE.
215 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
223 struct mlx4_wqe_ctrl_seg *ctrl;
225 if (qp->sq_max_wqes_per_wr > 1) {
226 s = roundup(size, 1U << qp->sq.wqe_shift);
227 for (i = 0; i < s; i += 64) {
228 ind = (i >> qp->sq.wqe_shift) + n;
229 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
230 cpu_to_be32(0xffffffff);
231 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
232 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
236 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
237 s = (ctrl->fence_size & 0x3f) << 4;
238 for (i = 64; i < s; i += 64) {
240 *wqe = cpu_to_be32(0xffffffff);
245 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
247 struct mlx4_wqe_ctrl_seg *ctrl;
248 struct mlx4_wqe_inline_seg *inl;
252 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
253 s = sizeof(struct mlx4_wqe_ctrl_seg);
255 if (qp->ibqp.qp_type == IB_QPT_UD) {
256 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
257 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
258 memset(dgram, 0, sizeof *dgram);
259 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
260 s += sizeof(struct mlx4_wqe_datagram_seg);
263 /* Pad the remainder of the WQE with an inline data segment. */
266 inl->byte_count = cpu_to_be32(1U << 31 | (size - s - sizeof *inl));
268 ctrl->srcrb_flags = 0;
269 ctrl->fence_size = size / 16;
271 * Make sure descriptor is fully written before setting ownership bit
272 * (because HW can start executing as soon as we do).
276 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
277 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
279 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
282 /* Post NOP WQE to prevent wrap-around in the middle of WR */
283 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
285 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
286 if (unlikely(s < qp->sq_max_wqes_per_wr)) {
287 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
293 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
295 struct ib_event event;
296 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
298 if (type == MLX4_EVENT_TYPE_PATH_MIG)
299 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
301 if (ibqp->event_handler) {
302 event.device = ibqp->device;
303 event.element.qp = ibqp;
305 case MLX4_EVENT_TYPE_PATH_MIG:
306 event.event = IB_EVENT_PATH_MIG;
308 case MLX4_EVENT_TYPE_COMM_EST:
309 event.event = IB_EVENT_COMM_EST;
311 case MLX4_EVENT_TYPE_SQ_DRAINED:
312 event.event = IB_EVENT_SQ_DRAINED;
314 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
315 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
317 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
318 event.event = IB_EVENT_QP_FATAL;
320 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
321 event.event = IB_EVENT_PATH_MIG_ERR;
323 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
324 event.event = IB_EVENT_QP_REQ_ERR;
326 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
327 event.event = IB_EVENT_QP_ACCESS_ERR;
330 pr_warn("Unexpected event type %d "
331 "on QP %06x\n", type, qp->qpn);
335 ibqp->event_handler(&event, ibqp->qp_context);
339 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
342 * UD WQEs must have a datagram segment.
343 * RC and UC WQEs might have a remote address segment.
344 * MLX WQEs need two extra inline data segments (for the UD
345 * header and space for the ICRC).
349 return sizeof (struct mlx4_wqe_ctrl_seg) +
350 sizeof (struct mlx4_wqe_datagram_seg) +
351 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
352 case MLX4_IB_QPT_PROXY_SMI_OWNER:
353 case MLX4_IB_QPT_PROXY_SMI:
354 case MLX4_IB_QPT_PROXY_GSI:
355 return sizeof (struct mlx4_wqe_ctrl_seg) +
356 sizeof (struct mlx4_wqe_datagram_seg) + 64;
357 case MLX4_IB_QPT_TUN_SMI_OWNER:
358 case MLX4_IB_QPT_TUN_GSI:
359 return sizeof (struct mlx4_wqe_ctrl_seg) +
360 sizeof (struct mlx4_wqe_datagram_seg);
363 return sizeof (struct mlx4_wqe_ctrl_seg) +
364 sizeof (struct mlx4_wqe_raddr_seg);
366 return sizeof (struct mlx4_wqe_ctrl_seg) +
367 sizeof (struct mlx4_wqe_masked_atomic_seg) +
368 sizeof (struct mlx4_wqe_raddr_seg);
369 case MLX4_IB_QPT_SMI:
370 case MLX4_IB_QPT_GSI:
371 return sizeof (struct mlx4_wqe_ctrl_seg) +
372 ALIGN(MLX4_IB_UD_HEADER_SIZE +
373 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
375 sizeof (struct mlx4_wqe_inline_seg),
376 sizeof (struct mlx4_wqe_data_seg)) +
378 sizeof (struct mlx4_wqe_inline_seg),
379 sizeof (struct mlx4_wqe_data_seg));
381 return sizeof (struct mlx4_wqe_ctrl_seg);
385 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
386 int is_user, int has_rq, struct mlx4_ib_qp *qp)
388 /* Sanity check RQ size before proceeding */
389 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
390 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
394 if (cap->max_recv_wr)
397 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
399 /* HW requires >= 1 RQ entry with >= 1 gather entry */
400 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
403 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr));
404 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
405 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
408 /* leave userspace return values as they were, so as not to break ABI */
410 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt;
411 cap->max_recv_sge = qp->rq.max_gs;
413 cap->max_recv_wr = qp->rq.max_post =
414 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
415 cap->max_recv_sge = min(qp->rq.max_gs,
416 min(dev->dev->caps.max_sq_sg,
417 dev->dev->caps.max_rq_sg));
423 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
424 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
428 /* Sanity check SQ size before proceeding */
429 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
430 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
431 cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
432 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
436 * For MLX transport we need 2 extra S/G entries:
437 * one for the header and one for the checksum at the end
439 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
440 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
441 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
444 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
445 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
446 send_wqe_overhead(type, qp->flags);
448 if (s > dev->dev->caps.max_sq_desc_sz)
452 * Hermon supports shrinking WQEs, such that a single work
453 * request can include multiple units of 1 << wqe_shift. This
454 * way, work requests can differ in size, and do not have to
455 * be a power of 2 in size, saving memory and speeding up send
456 * WR posting. Unfortunately, if we do this then the
457 * wqe_index field in CQEs can't be used to look up the WR ID
458 * anymore, so we do this only if selective signaling is off.
460 * Further, on 32-bit platforms, we can't use vmap() to make
461 * the QP buffer virtually contiguous. Thus we have to use
462 * constant-sized WRs to make sure a WR is always fully within
463 * a single page-sized chunk.
465 * Finally, we use NOP work requests to pad the end of the
466 * work queue, to avoid wrap-around in the middle of WR. We
467 * set NEC bit to avoid getting completions with error for
468 * these NOP WRs, but since NEC is only supported starting
469 * with firmware 2.2.232, we use constant-sized WRs for older
472 * And, since MLX QPs only support SEND, we use constant-sized
475 * We look for the smallest value of wqe_shift such that the
476 * resulting number of wqes does not exceed device
479 * We set WQE size to at least 64 bytes, this way stamping
480 * invalidates each WQE.
482 if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
483 qp->sq_signal_bits && BITS_PER_LONG == 64 &&
484 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
485 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
486 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
487 qp->sq.wqe_shift = ilog2(64);
489 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
492 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
495 * We need to leave 2 KB + 1 WR of headroom in the SQ to
496 * allow HW to prefetch.
498 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
499 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
500 qp->sq_max_wqes_per_wr +
503 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
506 if (qp->sq_max_wqes_per_wr <= 1)
512 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
513 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
514 send_wqe_overhead(type, qp->flags)) /
515 sizeof (struct mlx4_wqe_data_seg);
517 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
518 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
519 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
521 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
523 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
527 cap->max_send_wr = qp->sq.max_post =
528 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
529 cap->max_send_sge = min(qp->sq.max_gs,
530 min(dev->dev->caps.max_sq_sg,
531 dev->dev->caps.max_rq_sg));
532 qp->max_inline_data = cap->max_inline_data;
537 static int set_user_sq_size(struct mlx4_ib_dev *dev,
538 struct mlx4_ib_qp *qp,
539 struct mlx4_ib_create_qp *ucmd)
541 /* Sanity check SQ size before proceeding */
542 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes ||
543 ucmd->log_sq_stride >
544 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
545 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
548 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
549 qp->sq.wqe_shift = ucmd->log_sq_stride;
551 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
552 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
557 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
562 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
564 if (!qp->sqp_proxy_rcv)
566 for (i = 0; i < qp->rq.wqe_cnt; i++) {
567 qp->sqp_proxy_rcv[i].addr =
568 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
570 if (!qp->sqp_proxy_rcv[i].addr)
572 qp->sqp_proxy_rcv[i].map =
573 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
574 sizeof (struct mlx4_ib_proxy_sqp_hdr),
582 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
583 sizeof (struct mlx4_ib_proxy_sqp_hdr),
585 kfree(qp->sqp_proxy_rcv[i].addr);
587 kfree(qp->sqp_proxy_rcv);
588 qp->sqp_proxy_rcv = NULL;
592 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
596 for (i = 0; i < qp->rq.wqe_cnt; i++) {
597 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
598 sizeof (struct mlx4_ib_proxy_sqp_hdr),
600 kfree(qp->sqp_proxy_rcv[i].addr);
602 kfree(qp->sqp_proxy_rcv);
605 static int qp_has_rq(struct ib_qp_init_attr *attr)
607 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
614 static int init_qpg_parent(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *pqp,
615 struct ib_qp_init_attr *attr, int *qpn)
617 struct mlx4_ib_qpg_data *qpg_data;
618 int tss_num, rss_num;
619 int tss_align_num, rss_align_num;
620 int tss_base, rss_base = 0;
623 /* Parent is part of the TSS range (in SW TSS ARP is sent via parent) */
624 tss_num = 1 + attr->parent_attrib.tss_child_count;
625 tss_align_num = roundup_pow_of_two(tss_num);
626 rss_num = attr->parent_attrib.rss_child_count;
627 rss_align_num = roundup_pow_of_two(rss_num);
630 /* RSS is requested */
631 if (!(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS))
633 if (rss_align_num > dev->dev->caps.max_rss_tbl_sz)
635 /* We must work with power of two */
636 attr->parent_attrib.rss_child_count = rss_align_num;
639 qpg_data = kzalloc(sizeof *qpg_data, GFP_KERNEL);
643 if(pqp->flags & MLX4_IB_QP_NETIF)
644 err = mlx4_ib_steer_qp_alloc(dev, tss_align_num, &tss_base);
646 err = mlx4_qp_reserve_range(dev->dev, tss_align_num,
647 tss_align_num, &tss_base, 1);
652 u32 alloc = BITS_TO_LONGS(tss_align_num) * sizeof(long);
653 qpg_data->tss_bitmap = kzalloc(alloc, GFP_KERNEL);
654 if (qpg_data->tss_bitmap == NULL) {
658 bitmap_fill(qpg_data->tss_bitmap, tss_num);
659 /* Note parent takes first index */
660 clear_bit(0, qpg_data->tss_bitmap);
664 u32 alloc = BITS_TO_LONGS(rss_align_num) * sizeof(long);
665 err = mlx4_qp_reserve_range(dev->dev, rss_align_num,
669 qpg_data->rss_bitmap = kzalloc(alloc, GFP_KERNEL);
670 if (qpg_data->rss_bitmap == NULL) {
674 bitmap_fill(qpg_data->rss_bitmap, rss_align_num);
677 qpg_data->tss_child_count = attr->parent_attrib.tss_child_count;
678 qpg_data->rss_child_count = attr->parent_attrib.rss_child_count;
679 qpg_data->qpg_parent = pqp;
680 qpg_data->qpg_tss_mask_sz = ilog2(tss_align_num);
681 qpg_data->tss_qpn_base = tss_base;
682 qpg_data->rss_qpn_base = rss_base;
684 pqp->qpg_data = qpg_data;
690 mlx4_qp_release_range(dev->dev, rss_base, rss_align_num);
694 kfree(qpg_data->tss_bitmap);
697 if(pqp->flags & MLX4_IB_QP_NETIF)
698 mlx4_ib_steer_qp_free(dev, tss_base, tss_align_num);
700 mlx4_qp_release_range(dev->dev, tss_base, tss_align_num);
707 static void free_qpg_parent(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *pqp)
709 struct mlx4_ib_qpg_data *qpg_data = pqp->qpg_data;
712 if (qpg_data->tss_child_count > 1)
713 kfree(qpg_data->tss_bitmap);
715 align_num = roundup_pow_of_two(1 + qpg_data->tss_child_count);
716 if(pqp->flags & MLX4_IB_QP_NETIF)
717 mlx4_ib_steer_qp_free(dev, qpg_data->tss_qpn_base, align_num);
719 mlx4_qp_release_range(dev->dev, qpg_data->tss_qpn_base, align_num);
721 if (qpg_data->rss_child_count > 1) {
722 kfree(qpg_data->rss_bitmap);
723 align_num = roundup_pow_of_two(qpg_data->rss_child_count);
724 mlx4_qp_release_range(dev->dev, qpg_data->rss_qpn_base,
731 static int alloc_qpg_qpn(struct ib_qp_init_attr *init_attr,
732 struct mlx4_ib_qp *pqp, int *qpn)
734 struct mlx4_ib_qp *mqp = to_mqp(init_attr->qpg_parent);
735 struct mlx4_ib_qpg_data *qpg_data = mqp->qpg_data;
738 switch (init_attr->qpg_type) {
739 case IB_QPG_CHILD_TX:
740 if (qpg_data->tss_child_count == 0)
743 /* Parent took index 0 */
744 idx = find_first_bit(qpg_data->tss_bitmap,
745 qpg_data->tss_child_count + 1);
746 if (idx >= qpg_data->tss_child_count + 1)
748 old = test_and_clear_bit(idx, qpg_data->tss_bitmap);
750 idx += qpg_data->tss_qpn_base;
752 case IB_QPG_CHILD_RX:
753 if (qpg_data->rss_child_count == 0)
756 idx = find_first_bit(qpg_data->rss_bitmap,
757 qpg_data->rss_child_count);
758 if (idx >= qpg_data->rss_child_count)
760 old = test_and_clear_bit(idx, qpg_data->rss_bitmap);
762 idx += qpg_data->rss_qpn_base;
768 pqp->qpg_data = qpg_data;
774 static void free_qpg_qpn(struct mlx4_ib_qp *mqp, int qpn)
776 struct mlx4_ib_qpg_data *qpg_data = mqp->qpg_data;
778 switch (mqp->qpg_type) {
779 case IB_QPG_CHILD_TX:
781 qpn -= qpg_data->tss_qpn_base;
782 set_bit(qpn, qpg_data->tss_bitmap);
784 case IB_QPG_CHILD_RX:
785 qpn -= qpg_data->rss_qpn_base;
786 set_bit(qpn, qpg_data->rss_bitmap);
790 pr_warn("wrong qpg type (%d)\n", mqp->qpg_type);
796 static int alloc_qpn_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
797 struct ib_qp_init_attr *attr, int *qpn)
801 switch (attr->qpg_type) {
803 /* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
804 * BlueFlame setup flow wrongly causes VLAN insertion. */
805 if (attr->qp_type == IB_QPT_RAW_PACKET) {
806 err = mlx4_qp_reserve_range(dev->dev, 1, 1, qpn, 1);
808 if(qp->flags & MLX4_IB_QP_NETIF)
809 err = mlx4_ib_steer_qp_alloc(dev, 1, qpn);
811 err = mlx4_qp_reserve_range(dev->dev, 1, 1, qpn, 0);
816 err = init_qpg_parent(dev, qp, attr, qpn);
819 case IB_QPG_CHILD_TX:
820 case IB_QPG_CHILD_RX:
822 err = alloc_qpg_qpn(attr, qp, qpn);
826 qp->qpg_type = IB_QPG_NONE;
832 qp->qpg_type = attr->qpg_type;
836 static void free_qpn_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
837 enum ib_qpg_type qpg_type, int qpn)
841 if (qp->flags & MLX4_IB_QP_NETIF)
842 mlx4_ib_steer_qp_free(dev, qpn, 1);
844 mlx4_qp_release_range(dev->dev, qpn, 1);
848 free_qpg_parent(dev, qp);
851 case IB_QPG_CHILD_TX:
852 case IB_QPG_CHILD_RX:
854 free_qpg_qpn(qp, qpn);
862 /* Revert allocation on create_qp_common */
863 static void unalloc_qpn_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
864 struct ib_qp_init_attr *attr, int qpn)
866 free_qpn_common(dev, qp, attr->qpg_type, qpn);
869 static void release_qpn_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
871 free_qpn_common(dev, qp, qp->qpg_type, qp->mqp.qpn);
874 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
875 struct ib_qp_init_attr *init_attr,
876 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp)
880 struct mlx4_ib_sqp *sqp;
881 struct mlx4_ib_qp *qp;
882 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
885 init_attr->qpg_type = IB_QPG_NONE;
888 /* When tunneling special qps, we use a plain UD qp */
890 if (mlx4_is_mfunc(dev->dev) &&
891 (!mlx4_is_master(dev->dev) ||
892 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
893 if (init_attr->qp_type == IB_QPT_GSI)
894 qp_type = MLX4_IB_QPT_PROXY_GSI;
895 else if (mlx4_is_master(dev->dev))
896 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
898 qp_type = MLX4_IB_QPT_PROXY_SMI;
901 /* add extra sg entry for tunneling */
902 init_attr->cap.max_recv_sge++;
903 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
904 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
905 container_of(init_attr,
906 struct mlx4_ib_qp_tunnel_init_attr, init_attr);
907 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
908 tnl_init->proxy_qp_type != IB_QPT_GSI) ||
909 !mlx4_is_master(dev->dev))
911 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
912 qp_type = MLX4_IB_QPT_TUN_GSI;
913 else if (tnl_init->slave == mlx4_master_func_num(dev->dev))
914 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
916 qp_type = MLX4_IB_QPT_TUN_SMI;
917 /* we are definitely in the PPF here, since we are creating
918 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
919 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
920 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
925 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
926 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
927 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
928 sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL);
932 qp->pri.vid = qp->alt.vid = 0xFFFF;
934 qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL);
937 qp->pri.vid = qp->alt.vid = 0xFFFF;
942 qp->mlx4_ib_qp_type = qp_type;
944 mutex_init(&qp->mutex);
945 spin_lock_init(&qp->sq.lock);
946 spin_lock_init(&qp->rq.lock);
947 INIT_LIST_HEAD(&qp->gid_list);
948 INIT_LIST_HEAD(&qp->steering_rules);
949 INIT_LIST_HEAD(&qp->rules_list);
951 qp->state = IB_QPS_RESET;
952 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
953 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
955 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
960 struct mlx4_ib_create_qp ucmd;
964 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
969 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
971 err = set_user_sq_size(dev, qp, &ucmd);
975 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
977 if (IS_ERR(qp->umem)) {
978 err = PTR_ERR(qp->umem);
982 n = ib_umem_page_count(qp->umem);
983 shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n);
984 err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt);
989 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
993 if (qp_has_rq(init_attr)) {
994 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
995 ucmd.db_addr, &qp->db);
1000 qp->sq_no_prefetch = 0;
1002 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1003 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1005 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1006 qp->flags |= MLX4_IB_QP_LSO;
1008 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP &&
1009 dev->dev->caps.steering_mode ==
1010 MLX4_STEERING_MODE_DEVICE_MANAGED &&
1011 !mlx4_is_mfunc(dev->dev))
1012 qp->flags |= MLX4_IB_QP_NETIF;
1014 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
1018 if (qp_has_rq(init_attr)) {
1019 err = mlx4_db_alloc(dev->dev, &qp->db, 0);
1026 if (qp->max_inline_data) {
1027 err = mlx4_bf_alloc(dev->dev, &qp->bf, 0);
1029 pr_debug("failed to allocate blue flame"
1030 " register (%d)", err);
1031 qp->bf.uar = &dev->priv_uar;
1034 qp->bf.uar = &dev->priv_uar;
1036 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
1041 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
1046 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
1050 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL);
1051 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL);
1053 if (!qp->sq.wrid || !qp->rq.wrid) {
1060 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1061 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
1062 if (alloc_proxy_bufs(pd->device, qp)) {
1068 err = alloc_qpn_common(dev, qp, init_attr, &qpn);
1073 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp);
1077 if (init_attr->qp_type == IB_QPT_XRC_TGT)
1078 qp->mqp.qpn |= (1 << 23);
1081 * Hardware wants QPN written in big-endian order (after
1082 * shifting) for send doorbell. Precompute this value to save
1083 * a little bit when posting sends.
1085 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1087 qp->mqp.event = mlx4_ib_qp_event;
1093 unalloc_qpn_common(dev, qp, init_attr, qpn);
1096 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1097 free_proxy_bufs(pd->device, qp);
1100 if (qp_has_rq(init_attr))
1101 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
1108 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1112 ib_umem_release(qp->umem);
1114 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1117 if (!pd->uobject && qp_has_rq(init_attr))
1118 mlx4_db_free(dev->dev, &qp->db);
1120 if (qp->max_inline_data)
1121 mlx4_bf_free(dev->dev, &qp->bf);
1129 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
1132 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
1133 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
1134 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
1135 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
1136 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
1137 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
1138 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
1143 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1144 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1146 if (send_cq == recv_cq) {
1147 spin_lock_irq(&send_cq->lock);
1148 (void) __acquire(&recv_cq->lock);
1149 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1150 spin_lock_irq(&send_cq->lock);
1151 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1153 spin_lock_irq(&recv_cq->lock);
1154 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1158 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
1159 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1161 if (send_cq == recv_cq) {
1162 (void) __release(&recv_cq->lock);
1163 spin_unlock_irq(&send_cq->lock);
1164 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1165 spin_unlock(&recv_cq->lock);
1166 spin_unlock_irq(&send_cq->lock);
1168 spin_unlock(&send_cq->lock);
1169 spin_unlock_irq(&recv_cq->lock);
1173 static void del_gid_entries(struct mlx4_ib_qp *qp)
1175 struct mlx4_ib_gid_entry *ge, *tmp;
1177 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1178 list_del(&ge->list);
1183 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
1185 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1186 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1188 return to_mpd(qp->ibqp.pd);
1191 static void get_cqs(struct mlx4_ib_qp *qp,
1192 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1194 switch (qp->ibqp.qp_type) {
1195 case IB_QPT_XRC_TGT:
1196 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1197 *recv_cq = *send_cq;
1199 case IB_QPT_XRC_INI:
1200 *send_cq = to_mcq(qp->ibqp.send_cq);
1201 *recv_cq = *send_cq;
1204 *send_cq = to_mcq(qp->ibqp.send_cq);
1205 *recv_cq = to_mcq(qp->ibqp.recv_cq);
1210 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1213 struct mlx4_ib_cq *send_cq, *recv_cq;
1215 if (qp->state != IB_QPS_RESET) {
1216 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1217 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1218 pr_warn("modify QP %06x to RESET failed.\n",
1221 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1225 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1228 if (qp->pri.vid < 0x1000) {
1229 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1230 qp->pri.vid = 0xFFFF;
1231 qp->pri.candidate_vid = 0xFFFF;
1232 qp->pri.update_vid = 0;
1234 if (qp->alt.vid < 0x1000) {
1235 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1236 qp->alt.vid = 0xFFFF;
1237 qp->alt.candidate_vid = 0xFFFF;
1238 qp->alt.update_vid = 0;
1242 get_cqs(qp, &send_cq, &recv_cq);
1244 mlx4_ib_lock_cqs(send_cq, recv_cq);
1247 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1248 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1249 if (send_cq != recv_cq)
1250 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1253 mlx4_qp_remove(dev->dev, &qp->mqp);
1255 mlx4_ib_unlock_cqs(send_cq, recv_cq);
1257 mlx4_qp_free(dev->dev, &qp->mqp);
1259 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp))
1260 release_qpn_common(dev, qp);
1262 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1266 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1268 ib_umem_release(qp->umem);
1272 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1273 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1274 free_proxy_bufs(&dev->ib_dev, qp);
1275 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1276 if (qp->max_inline_data)
1277 mlx4_bf_free(dev->dev, &qp->bf);
1280 mlx4_db_free(dev->dev, &qp->db);
1283 del_gid_entries(qp);
1286 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1289 if (!mlx4_is_mfunc(dev->dev) ||
1290 (mlx4_is_master(dev->dev) &&
1291 attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1292 return dev->dev->phys_caps.base_sqpn +
1293 (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1296 /* PF or VF -- creating proxies */
1297 if (attr->qp_type == IB_QPT_SMI)
1298 return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1300 return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1304 static int check_qpg_attr(struct mlx4_ib_dev *dev,
1305 struct ib_qp_init_attr *attr)
1307 if (attr->qpg_type == IB_QPG_NONE)
1310 if (attr->qp_type != IB_QPT_UD)
1313 if (attr->qpg_type == IB_QPG_PARENT) {
1314 if (attr->parent_attrib.tss_child_count == 1)
1315 return -EINVAL; /* Doesn't make sense */
1316 if (attr->parent_attrib.rss_child_count == 1)
1317 return -EINVAL; /* Doesn't make sense */
1318 if ((attr->parent_attrib.tss_child_count == 0) &&
1319 (attr->parent_attrib.rss_child_count == 0))
1320 /* Should be called with IP_QPG_NONE */
1322 if (attr->parent_attrib.rss_child_count > 1) {
1324 if (!(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS))
1326 rss_align_num = roundup_pow_of_two(
1327 attr->parent_attrib.rss_child_count);
1328 if (rss_align_num > dev->dev->caps.max_rss_tbl_sz)
1332 struct mlx4_ib_qpg_data *qpg_data;
1333 if (attr->qpg_parent == NULL)
1335 if (IS_ERR(attr->qpg_parent))
1337 qpg_data = to_mqp(attr->qpg_parent)->qpg_data;
1338 if (qpg_data == NULL)
1340 if (attr->qpg_type == IB_QPG_CHILD_TX &&
1341 !qpg_data->tss_child_count)
1343 if (attr->qpg_type == IB_QPG_CHILD_RX &&
1344 !qpg_data->rss_child_count)
1351 #define RESERVED_FLAGS_MASK ((((unsigned int)IB_QP_CREATE_RESERVED_END - 1) | IB_QP_CREATE_RESERVED_END) \
1352 & ~(IB_QP_CREATE_RESERVED_START - 1))
1354 static enum mlx4_ib_qp_flags to_mlx4_ib_qp_flags(enum ib_qp_create_flags ib_qp_flags)
1356 enum mlx4_ib_qp_flags mlx4_ib_qp_flags = 0;
1358 if (ib_qp_flags & IB_QP_CREATE_IPOIB_UD_LSO)
1359 mlx4_ib_qp_flags |= MLX4_IB_QP_LSO;
1361 if (ib_qp_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
1362 mlx4_ib_qp_flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1364 if (ib_qp_flags & IB_QP_CREATE_NETIF_QP)
1365 mlx4_ib_qp_flags |= MLX4_IB_QP_NETIF;
1367 /* reserved flags */
1368 mlx4_ib_qp_flags |= (ib_qp_flags & RESERVED_FLAGS_MASK);
1370 return mlx4_ib_qp_flags;
1373 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1374 struct ib_qp_init_attr *init_attr,
1375 struct ib_udata *udata)
1377 struct mlx4_ib_qp *qp = NULL;
1380 enum mlx4_ib_qp_flags mlx4_qp_flags = to_mlx4_ib_qp_flags(init_attr->create_flags);
1381 struct ib_device *device;
1383 /* see ib_core::ib_create_qp same handling */
1384 device = pd ? pd->device : init_attr->xrcd->device;
1386 * We only support LSO, vendor flag1, and multicast loopback blocking,
1387 * and only for kernel UD QPs.
1389 if (mlx4_qp_flags & ~(MLX4_IB_QP_LSO |
1390 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1391 MLX4_IB_SRIOV_TUNNEL_QP | MLX4_IB_SRIOV_SQP |
1393 return ERR_PTR(-EINVAL);
1395 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1396 if (init_attr->qp_type != IB_QPT_UD)
1397 return ERR_PTR(-EINVAL);
1400 if (init_attr->create_flags &&
1402 ((mlx4_qp_flags & ~MLX4_IB_SRIOV_SQP) &&
1403 init_attr->qp_type != IB_QPT_UD) ||
1404 ((mlx4_qp_flags & MLX4_IB_SRIOV_SQP) &&
1405 init_attr->qp_type > IB_QPT_GSI)))
1406 return ERR_PTR(-EINVAL);
1409 err = check_qpg_attr(to_mdev(device), init_attr);
1411 return ERR_PTR(err);
1414 switch (init_attr->qp_type) {
1415 case IB_QPT_XRC_TGT:
1416 pd = to_mxrcd(init_attr->xrcd)->pd;
1417 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1418 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1420 case IB_QPT_XRC_INI:
1421 if (!(to_mdev(device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1422 return ERR_PTR(-ENOSYS);
1423 init_attr->recv_cq = init_attr->send_cq;
1427 case IB_QPT_RAW_PACKET:
1428 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1430 return ERR_PTR(-ENOMEM);
1431 qp->pri.vid = qp->alt.vid = 0xFFFF;
1435 err = create_qp_common(to_mdev(device), pd, init_attr, udata, 0, &qp);
1438 return ERR_PTR(err);
1441 qp->ibqp.qp_num = qp->mqp.qpn;
1449 /* Userspace is not allowed to create special QPs: */
1451 return ERR_PTR(-EINVAL);
1453 err = create_qp_common(to_mdev(device), pd, init_attr, udata,
1454 get_sqp_num(to_mdev(device), init_attr),
1457 return ERR_PTR(err);
1459 qp->port = init_attr->port_num;
1460 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1465 /* Don't support raw QPs */
1466 return ERR_PTR(-EINVAL);
1472 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1474 struct mlx4_ib_dev *dev = to_mdev(qp->device);
1475 struct mlx4_ib_qp *mqp = to_mqp(qp);
1476 struct mlx4_ib_pd *pd;
1478 if (is_qp0(dev, mqp))
1479 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1482 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1484 if (is_sqp(dev, mqp))
1485 kfree(to_msqp(mqp));
1492 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1495 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC;
1496 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC;
1497 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD;
1498 case MLX4_IB_QPT_XRC_INI:
1499 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC;
1500 case MLX4_IB_QPT_SMI:
1501 case MLX4_IB_QPT_GSI:
1502 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX;
1504 case MLX4_IB_QPT_PROXY_SMI_OWNER:
1505 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1506 MLX4_QP_ST_MLX : -1);
1507 case MLX4_IB_QPT_PROXY_SMI:
1508 case MLX4_IB_QPT_TUN_SMI:
1509 case MLX4_IB_QPT_PROXY_GSI:
1510 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ?
1511 MLX4_QP_ST_UD : -1);
1516 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1521 u32 hw_access_flags = 0;
1523 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1524 dest_rd_atomic = attr->max_dest_rd_atomic;
1526 dest_rd_atomic = qp->resp_depth;
1528 if (attr_mask & IB_QP_ACCESS_FLAGS)
1529 access_flags = attr->qp_access_flags;
1531 access_flags = qp->atomic_rd_en;
1533 if (!dest_rd_atomic)
1534 access_flags &= IB_ACCESS_REMOTE_WRITE;
1536 if (access_flags & IB_ACCESS_REMOTE_READ)
1537 hw_access_flags |= MLX4_QP_BIT_RRE;
1538 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1539 hw_access_flags |= MLX4_QP_BIT_RAE;
1540 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1541 hw_access_flags |= MLX4_QP_BIT_RWE;
1543 return cpu_to_be32(hw_access_flags);
1546 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1549 if (attr_mask & IB_QP_PKEY_INDEX)
1550 sqp->pkey_index = attr->pkey_index;
1551 if (attr_mask & IB_QP_QKEY)
1552 sqp->qkey = attr->qkey;
1553 if (attr_mask & IB_QP_SQ_PSN)
1554 sqp->send_psn = attr->sq_psn;
1557 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1559 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1562 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1563 struct mlx4_ib_qp *qp, struct mlx4_qp_path *path,
1564 u8 port, int is_primary)
1566 struct net_device *ndev;
1568 int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1569 IB_LINK_LAYER_ETHERNET;
1577 struct mlx4_roce_smac_vlan_info *smac_info;
1579 path->grh_mylmc = ah->src_path_bits & 0x7f;
1580 path->rlid = cpu_to_be16(ah->dlid);
1581 if (ah->static_rate) {
1582 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1583 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1584 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1585 --path->static_rate;
1587 path->static_rate = 0;
1589 if (ah->ah_flags & IB_AH_GRH) {
1590 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1591 pr_err("sgid_index (%u) too large. max is %d\n",
1592 ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1596 path->grh_mylmc |= 1 << 7;
1597 path->mgid_index = ah->grh.sgid_index;
1598 path->hop_limit = ah->grh.hop_limit;
1599 path->tclass_flowlabel =
1600 cpu_to_be32((ah->grh.traffic_class << 20) |
1601 (ah->grh.flow_label));
1602 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1606 if (!(ah->ah_flags & IB_AH_GRH))
1609 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1610 ((port - 1) << 6) | ((ah->sl & 7) << 3);
1613 smac_info = &qp->pri;
1615 smac_info = &qp->alt;
1617 vlan_tag = rdma_get_vlan_id(&dev->iboe.gid_table[port - 1][ah->grh.sgid_index]);
1618 if (vlan_tag < 0x1000) {
1619 if (smac_info->vid < 0x1000) {
1620 /* both valid vlan ids */
1621 if (smac_info->vid != vlan_tag) {
1622 /* different VIDs. unreg old and reg new */
1623 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1626 smac_info->candidate_vid = vlan_tag;
1627 smac_info->candidate_vlan_index = vidx;
1628 smac_info->candidate_vlan_port = port;
1629 smac_info->update_vid = 1;
1630 path->vlan_index = vidx;
1633 path->vlan_index = smac_info->vlan_index;
1637 /* no current vlan tag in qp */
1638 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1641 smac_info->candidate_vid = vlan_tag;
1642 smac_info->candidate_vlan_index = vidx;
1643 smac_info->candidate_vlan_port = port;
1644 smac_info->update_vid = 1;
1645 path->vlan_index = vidx;
1649 /* have current vlan tag. unregister it at modify-qp success */
1650 if (smac_info->vid < 0x1000) {
1651 smac_info->candidate_vid = 0xFFFF;
1652 smac_info->update_vid = 1;
1656 err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port);
1660 /* get smac_index for RoCE use.
1661 * If no smac was yet assigned, register one.
1662 * If one was already assigned, but the new mac differs,
1663 * unregister the old one and register the new one.
1665 spin_lock(&dev->iboe.lock);
1666 ndev = dev->iboe.netdevs[port - 1];
1669 smac = ndev->dev_addr; /* fixme: cache this value */
1671 smac = IF_LLADDR(ndev); /* fixme: cache this value */
1674 u64_mac = mlx4_mac_to_u64(smac);
1676 u64_mac = dev->dev->caps.def_mac[port];
1677 spin_unlock(&dev->iboe.lock);
1679 if (!smac_info->smac || smac_info->smac != u64_mac) {
1680 /* register candidate now, unreg if needed, after success */
1681 smac_index = mlx4_register_mac(dev->dev, port, u64_mac);
1682 if (smac_index >= 0) {
1683 smac_info->candidate_smac_index = smac_index;
1684 smac_info->candidate_smac = u64_mac;
1685 smac_info->candidate_smac_port = port;
1689 smac_index = smac_info->smac_index;
1691 memcpy(path->dmac, mac, 6);
1692 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1693 /* put MAC table smac index for IBoE */
1694 path->grh_mylmc = (u8) (smac_index) | 0x80 ;
1697 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1698 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1703 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1705 struct mlx4_ib_gid_entry *ge, *tmp;
1707 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1708 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1710 ge->port = qp->port;
1715 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1716 struct mlx4_qp_context *context)
1718 struct net_device *ndev;
1723 ndev = dev->iboe.netdevs[qp->port - 1];
1726 smac = ndev->dev_addr; /* fixme: cache this value */
1728 smac = IF_LLADDR(ndev); /* fixme: cache this value */
1730 u64_mac = mlx4_mac_to_u64(smac);
1732 u64_mac = dev->dev->caps.def_mac[qp->port];
1734 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1735 if (!qp->pri.smac) {
1736 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1737 if (smac_index >= 0) {
1738 qp->pri.candidate_smac_index = smac_index;
1739 qp->pri.candidate_smac = u64_mac;
1740 qp->pri.candidate_smac_port = qp->port;
1741 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1747 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1748 const struct ib_qp_attr *attr, int attr_mask,
1749 enum ib_qp_state cur_state, enum ib_qp_state new_state)
1751 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1752 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1753 struct mlx4_ib_pd *pd;
1754 struct mlx4_ib_cq *send_cq, *recv_cq;
1755 struct mlx4_qp_context *context;
1756 enum mlx4_qp_optpar optpar = 0;
1762 context = kzalloc(sizeof *context, GFP_KERNEL);
1766 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1767 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1769 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1770 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1772 optpar |= MLX4_QP_OPTPAR_PM_STATE;
1773 switch (attr->path_mig_state) {
1774 case IB_MIG_MIGRATED:
1775 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1778 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1781 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1786 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1787 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1788 else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1789 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1790 else if (ibqp->qp_type == IB_QPT_UD) {
1791 if (qp->flags & MLX4_IB_QP_LSO)
1792 context->mtu_msgmax = (IB_MTU_4096 << 5) |
1793 ilog2(dev->dev->caps.max_gso_sz);
1795 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1796 } else if (attr_mask & IB_QP_PATH_MTU) {
1797 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1798 pr_err("path MTU (%u) is invalid\n",
1802 context->mtu_msgmax = (attr->path_mtu << 5) |
1803 ilog2(dev->dev->caps.max_msg_sz);
1807 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1808 context->rq_size_stride |= qp->rq.wqe_shift - 4;
1811 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1812 context->sq_size_stride |= qp->sq.wqe_shift - 4;
1814 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1815 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1816 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1817 context->param3 |= cpu_to_be32(1 << 30);
1820 if (qp->ibqp.uobject)
1821 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1823 context->usr_page = cpu_to_be32(qp->bf.uar->index);
1825 if (attr_mask & IB_QP_DEST_QPN)
1826 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1828 if (attr_mask & IB_QP_PORT) {
1829 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1830 !(attr_mask & IB_QP_AV)) {
1831 mlx4_set_sched(&context->pri_path, attr->port_num);
1832 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1836 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1837 if (dev->counters[qp->port - 1] != -1) {
1838 context->pri_path.counter_index =
1839 dev->counters[qp->port - 1];
1840 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1842 context->pri_path.counter_index = 0xff;
1844 if (qp->flags & MLX4_IB_QP_NETIF &&
1845 (qp->qpg_type == IB_QPG_NONE || qp->qpg_type == IB_QPG_PARENT)) {
1846 mlx4_ib_steer_qp_reg(dev, qp, 1);
1851 if (attr_mask & IB_QP_PKEY_INDEX) {
1852 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1853 context->pri_path.disable_pkey_check = 0x40;
1854 context->pri_path.pkey_index = attr->pkey_index;
1855 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1858 if (attr_mask & IB_QP_AV) {
1859 if (mlx4_set_path(dev, &attr->ah_attr, qp, &context->pri_path,
1860 attr_mask & IB_QP_PORT ?
1861 attr->port_num : qp->port, 1))
1864 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1865 MLX4_QP_OPTPAR_SCHED_QUEUE);
1868 if (attr_mask & IB_QP_TIMEOUT) {
1869 context->pri_path.ackto |= attr->timeout << 3;
1870 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1873 if (attr_mask & IB_QP_ALT_PATH) {
1874 if (attr->alt_port_num == 0 ||
1875 attr->alt_port_num > dev->dev->caps.num_ports)
1878 if (attr->alt_pkey_index >=
1879 dev->dev->caps.pkey_table_len[attr->alt_port_num])
1882 if (mlx4_set_path(dev, &attr->alt_ah_attr, qp, &context->alt_path,
1883 attr->alt_port_num, 0))
1886 context->alt_path.pkey_index = attr->alt_pkey_index;
1887 context->alt_path.ackto = attr->alt_timeout << 3;
1888 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1892 get_cqs(qp, &send_cq, &recv_cq);
1893 context->pd = cpu_to_be32(pd->pdn);
1894 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1895 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1896 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1898 /* Set "fast registration enabled" for all kernel QPs */
1899 if (!qp->ibqp.uobject)
1900 context->params1 |= cpu_to_be32(1 << 11);
1902 if (attr_mask & IB_QP_RNR_RETRY) {
1903 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1904 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1907 if (attr_mask & IB_QP_RETRY_CNT) {
1908 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1909 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1912 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1913 if (attr->max_rd_atomic)
1915 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1916 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1919 if (attr_mask & IB_QP_SQ_PSN)
1920 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1922 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1923 if (attr->max_dest_rd_atomic)
1925 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1926 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1929 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1930 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1931 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1934 if (attr_mask & IB_M_EXT_CLASS_1)
1935 context->params2 |= cpu_to_be32(MLX4_QP_BIT_COLL_MASTER);
1937 /* for now we enable also sqe on send */
1938 if (attr_mask & IB_M_EXT_CLASS_2) {
1939 context->params2 |= cpu_to_be32(MLX4_QP_BIT_COLL_SYNC_SQ);
1940 context->params2 |= cpu_to_be32(MLX4_QP_BIT_COLL_MASTER);
1943 if (attr_mask & IB_M_EXT_CLASS_3)
1944 context->params2 |= cpu_to_be32(MLX4_QP_BIT_COLL_SYNC_RQ);
1947 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1949 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1950 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1951 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1953 if (attr_mask & IB_QP_RQ_PSN)
1954 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1956 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1957 if (attr_mask & IB_QP_QKEY) {
1958 if (qp->mlx4_ib_qp_type &
1959 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1960 context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1962 if (mlx4_is_mfunc(dev->dev) &&
1963 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1964 (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1965 MLX4_RESERVED_QKEY_BASE) {
1966 pr_err("Cannot use reserved QKEY"
1967 " 0x%x (range 0xffff0000..0xffffffff"
1968 " is reserved)\n", attr->qkey);
1972 context->qkey = cpu_to_be32(attr->qkey);
1974 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1978 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1980 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1981 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1983 if (cur_state == IB_QPS_INIT &&
1984 new_state == IB_QPS_RTR &&
1985 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1986 ibqp->qp_type == IB_QPT_UD ||
1987 ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1988 context->pri_path.sched_queue = (qp->port - 1) << 6;
1989 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1990 qp->mlx4_ib_qp_type &
1991 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1992 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1993 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1994 context->pri_path.fl = 0x80;
1996 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1997 context->pri_path.fl = 0x80;
1998 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
2000 is_eth = rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
2001 IB_LINK_LAYER_ETHERNET;
2003 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
2004 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
2005 context->pri_path.feup = 1 << 7; /* don't fsm */
2006 /* handle smac_index */
2007 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
2008 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
2009 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
2010 err = handle_eth_ud_smac_index(dev, qp, context);
2017 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
2018 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2023 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2024 context->rlkey |= (1 << 4);
2026 if ((attr_mask & IB_QP_GROUP_RSS) &&
2027 (qp->qpg_data->rss_child_count > 1)) {
2028 struct mlx4_ib_qpg_data *qpg_data = qp->qpg_data;
2029 void *rss_context_base = &context->pri_path;
2030 struct mlx4_rss_context *rss_context =
2031 (struct mlx4_rss_context *) (rss_context_base
2032 + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH);
2034 context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET);
2036 /* This should be tbl_sz_base_qpn */
2037 rss_context->base_qpn = cpu_to_be32(qpg_data->rss_qpn_base |
2038 (ilog2(qpg_data->rss_child_count) << 24));
2039 rss_context->default_qpn = cpu_to_be32(qpg_data->rss_qpn_base);
2040 /* This should be flags_hash_fn */
2041 rss_context->flags = MLX4_RSS_TCP_IPV6 |
2043 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS) {
2044 rss_context->base_qpn_udp = rss_context->default_qpn;
2045 rss_context->flags |= MLX4_RSS_IPV6 |
2050 if (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP) {
2051 static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B,
2052 0x1983A2FC, 0x943E1ADB, 0xD9389E6B, 0xD1039C2C,
2053 0xA74499AD, 0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
2054 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
2055 memcpy(rss_context->rss_key, rsskey,
2056 sizeof(rss_context->rss_key));
2058 rss_context->hash_fn = MLX4_RSS_HASH_XOR;
2059 memset(rss_context->rss_key, 0,
2060 sizeof(rss_context->rss_key));
2064 * Before passing a kernel QP to the HW, make sure that the
2065 * ownership bits of the send queue are set and the SQ
2066 * headroom is stamped so that the hardware doesn't start
2067 * processing stale work requests.
2069 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2070 struct mlx4_wqe_ctrl_seg *ctrl;
2073 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
2074 ctrl = get_send_wqe(qp, i);
2075 ctrl->owner_opcode = cpu_to_be32(1 << 31);
2076 if (qp->sq_max_wqes_per_wr == 1)
2077 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
2079 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2083 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2084 to_mlx4_state(new_state), context, optpar,
2085 sqd_event, &qp->mqp);
2089 qp->state = new_state;
2091 if (attr_mask & IB_QP_ACCESS_FLAGS)
2092 qp->atomic_rd_en = attr->qp_access_flags;
2093 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2094 qp->resp_depth = attr->max_dest_rd_atomic;
2095 if (attr_mask & IB_QP_PORT) {
2096 qp->port = attr->port_num;
2097 update_mcg_macs(dev, qp);
2099 if (attr_mask & IB_QP_ALT_PATH)
2100 qp->alt_port = attr->alt_port_num;
2102 if (is_sqp(dev, qp))
2103 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2105 /* Set 'ignore_cq_overrun' bits for collectives offload */
2106 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2107 if (attr_mask & (IB_M_EXT_CLASS_2 | IB_M_EXT_CLASS_3)) {
2108 err = mlx4_ib_ignore_overrun_cq(ibqp->send_cq);
2110 pr_err("Failed to set ignore CQ "
2111 "overrun for QP 0x%x's send CQ\n",
2116 if (ibqp->recv_cq != ibqp->send_cq) {
2117 err = mlx4_ib_ignore_overrun_cq(ibqp->recv_cq);
2119 pr_err("Failed to set ignore "
2120 "CQ overrun for QP 0x%x's recv "
2121 "CQ\n", ibqp->qp_num);
2129 * If we moved QP0 to RTR, bring the IB link up; if we moved
2130 * QP0 to RESET or ERROR, bring the link back down.
2132 if (is_qp0(dev, qp)) {
2133 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2134 if (mlx4_INIT_PORT(dev->dev, qp->port))
2135 pr_warn("INIT_PORT failed for port %d\n",
2138 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2139 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2140 mlx4_CLOSE_PORT(dev->dev, qp->port);
2144 * If we moved a kernel QP to RESET, clean up all old CQ
2145 * entries and reinitialize the QP.
2147 if (new_state == IB_QPS_RESET) {
2148 if (!ibqp->uobject) {
2149 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2150 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2151 if (send_cq != recv_cq)
2152 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2158 qp->sq_next_wqe = 0;
2162 if (qp->flags & MLX4_IB_QP_NETIF &&
2163 (qp->qpg_type == IB_QPG_NONE ||
2164 qp->qpg_type == IB_QPG_PARENT))
2165 mlx4_ib_steer_qp_reg(dev, qp, 0);
2168 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2172 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2175 if (qp->pri.vid < 0x1000) {
2176 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2177 qp->pri.vid = 0xFFFF;
2178 qp->pri.candidate_vid = 0xFFFF;
2179 qp->pri.update_vid = 0;
2182 if (qp->alt.vid < 0x1000) {
2183 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2184 qp->alt.vid = 0xFFFF;
2185 qp->alt.candidate_vid = 0xFFFF;
2186 qp->alt.update_vid = 0;
2191 if (err && steer_qp)
2192 mlx4_ib_steer_qp_reg(dev, qp, 0);
2194 if (qp->pri.candidate_smac) {
2196 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2199 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2201 qp->pri.smac = qp->pri.candidate_smac;
2202 qp->pri.smac_index = qp->pri.candidate_smac_index;
2203 qp->pri.smac_port = qp->pri.candidate_smac_port;
2206 qp->pri.candidate_smac = 0;
2207 qp->pri.candidate_smac_index = 0;
2208 qp->pri.candidate_smac_port = 0;
2210 if (qp->alt.candidate_smac) {
2212 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->pri.candidate_smac);
2215 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2217 qp->alt.smac = qp->alt.candidate_smac;
2218 qp->alt.smac_index = qp->alt.candidate_smac_index;
2219 qp->alt.smac_port = qp->alt.candidate_smac_port;
2222 qp->pri.candidate_smac = 0;
2223 qp->pri.candidate_smac_index = 0;
2224 qp->pri.candidate_smac_port = 0;
2227 if (qp->pri.update_vid) {
2229 if (qp->pri.candidate_vid < 0x1000)
2230 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2231 qp->pri.candidate_vid);
2233 if (qp->pri.vid < 0x1000)
2234 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2236 qp->pri.vid = qp->pri.candidate_vid;
2237 qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2238 qp->pri.vlan_index = qp->pri.candidate_vlan_index;
2240 qp->pri.candidate_vid = 0xFFFF;
2241 qp->pri.update_vid = 0;
2244 if (qp->alt.update_vid) {
2246 if (qp->alt.candidate_vid < 0x1000)
2247 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2248 qp->alt.candidate_vid);
2250 if (qp->alt.vid < 0x1000)
2251 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2253 qp->alt.vid = qp->alt.candidate_vid;
2254 qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2255 qp->alt.vlan_index = qp->alt.candidate_vlan_index;
2257 qp->alt.candidate_vid = 0xFFFF;
2258 qp->alt.update_vid = 0;
2264 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2265 int attr_mask, struct ib_udata *udata)
2267 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2268 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2269 enum ib_qp_state cur_state, new_state;
2272 mutex_lock(&qp->mutex);
2274 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2275 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2277 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2278 attr_mask & ~IB_M_QP_MOD_VEND_MASK)) {
2279 pr_debug("qpn 0x%x: invalid attribute mask specified "
2280 "for transition %d to %d. qp_type %d,"
2281 " attr_mask 0x%x\n",
2282 ibqp->qp_num, cur_state, new_state,
2283 ibqp->qp_type, attr_mask);
2287 if ((attr_mask & IB_M_QP_MOD_VEND_MASK) && !dev->dev->caps.sync_qp) {
2288 pr_err("extended verbs are not supported by %s\n",
2293 if ((attr_mask & IB_QP_PORT) &&
2294 (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2295 pr_debug("qpn 0x%x: invalid port number (%d) specified "
2296 "for transition %d to %d. qp_type %d\n",
2297 ibqp->qp_num, attr->port_num, cur_state,
2298 new_state, ibqp->qp_type);
2302 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2303 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2304 IB_LINK_LAYER_ETHERNET))
2307 if (attr_mask & IB_QP_PKEY_INDEX) {
2308 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2309 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2310 pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2311 "for transition %d to %d. qp_type %d\n",
2312 ibqp->qp_num, attr->pkey_index, cur_state,
2313 new_state, ibqp->qp_type);
2318 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2319 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2320 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2321 "Transition %d to %d. qp_type %d\n",
2322 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2323 new_state, ibqp->qp_type);
2327 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2328 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2329 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2330 "Transition %d to %d. qp_type %d\n",
2331 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2332 new_state, ibqp->qp_type);
2336 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2341 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2344 mutex_unlock(&qp->mutex);
2348 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2349 struct ib_send_wr *wr,
2350 void *wqe, unsigned *mlx_seg_len)
2352 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2353 struct ib_device *ib_dev = &mdev->ib_dev;
2354 struct mlx4_wqe_mlx_seg *mlx = wqe;
2355 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2356 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2364 if (wr->opcode != IB_WR_SEND)
2369 for (i = 0; i < wr->num_sge; ++i)
2370 send_size += wr->sg_list[i].length;
2372 /* for proxy-qp0 sends, need to add in size of tunnel header */
2373 /* for tunnel-qp0 sends, tunnel header is already in s/g list */
2374 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2375 send_size += sizeof (struct mlx4_ib_tunnel_header);
2377 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
2379 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2380 sqp->ud_header.lrh.service_level =
2381 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2382 sqp->ud_header.lrh.destination_lid =
2383 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2384 sqp->ud_header.lrh.source_lid =
2385 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2388 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2390 /* force loopback */
2391 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2392 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2394 sqp->ud_header.lrh.virtual_lane = 0;
2395 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2396 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2397 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2398 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2399 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2401 sqp->ud_header.bth.destination_qpn =
2402 cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2404 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2405 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2407 sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2408 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2410 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2411 sqp->ud_header.immediate_present = 0;
2413 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2416 * Inline data segments may not cross a 64 byte boundary. If
2417 * our UD header is bigger than the space available up to the
2418 * next 64 byte boundary in the WQE, use two inline data
2419 * segments to hold the UD header.
2421 spc = MLX4_INLINE_ALIGN -
2422 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2423 if (header_size <= spc) {
2424 inl->byte_count = cpu_to_be32(1U << 31 | header_size);
2425 memcpy(inl + 1, sqp->header_buf, header_size);
2428 inl->byte_count = cpu_to_be32(1U << 31 | spc);
2429 memcpy(inl + 1, sqp->header_buf, spc);
2431 inl = (void *) (inl + 1) + spc;
2432 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2434 * Need a barrier here to make sure all the data is
2435 * visible before the byte_count field is set.
2436 * Otherwise the HCA prefetcher could grab the 64-byte
2437 * chunk with this inline segment and get a valid (!=
2438 * 0xffffffff) byte count but stale data, and end up
2439 * generating a packet with bad headers.
2441 * The first inline segment's byte_count field doesn't
2442 * need a barrier, because it comes after a
2443 * control/MLX segment and therefore is at an offset
2447 inl->byte_count = cpu_to_be32(1U << 31 | (header_size - spc));
2452 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2456 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2457 void *wqe, unsigned *mlx_seg_len)
2459 struct ib_device *ib_dev = sqp->qp.ibqp.device;
2460 struct mlx4_wqe_mlx_seg *mlx = wqe;
2461 struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2462 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2463 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2477 for (i = 0; i < wr->num_sge; ++i)
2478 send_size += wr->sg_list[i].length;
2480 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2481 is_grh = mlx4_ib_ah_grh_present(ah);
2483 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2484 /* When multi-function is enabled, the ib_core gid
2485 * indexes don't necessarily match the hw ones, so
2486 * we must use our own cache */
2487 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2488 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2489 ah->av.ib.gid_index, &sgid.raw[0]);
2493 err = ib_get_cached_gid(ib_dev,
2494 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2495 ah->av.ib.gid_index, &sgid);
2500 vlan = rdma_get_vlan_id(&sgid);
2501 is_vlan = vlan < 0x1000;
2503 ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2506 sqp->ud_header.lrh.service_level =
2507 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2508 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2509 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2513 sqp->ud_header.grh.traffic_class =
2514 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2515 sqp->ud_header.grh.flow_label =
2516 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2517 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit;
2519 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2521 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2522 /* When multi-function is enabled, the ib_core gid
2523 * indexes don't necessarily match the hw ones, so
2524 * we must use our own cache */
2525 sqp->ud_header.grh.source_gid.global.subnet_prefix =
2526 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2528 sqp->ud_header.grh.source_gid.global.interface_id =
2529 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2530 guid_cache[ah->av.ib.gid_index];
2532 ib_get_cached_gid(ib_dev,
2533 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2534 ah->av.ib.gid_index,
2535 &sqp->ud_header.grh.source_gid);
2537 memcpy(sqp->ud_header.grh.destination_gid.raw,
2538 ah->av.ib.dgid, 16);
2541 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2544 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2545 (sqp->ud_header.lrh.destination_lid ==
2546 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2547 (sqp->ud_header.lrh.service_level << 8));
2548 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2549 mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2550 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2553 switch (wr->opcode) {
2555 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2556 sqp->ud_header.immediate_present = 0;
2558 case IB_WR_SEND_WITH_IMM:
2559 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2560 sqp->ud_header.immediate_present = 1;
2561 sqp->ud_header.immediate_data = wr->ex.imm_data;
2569 struct in6_addr in6;
2571 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2573 mlx->sched_prio = cpu_to_be16(pcp);
2575 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2576 /* FIXME: cache smac value? */
2577 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2578 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2579 memcpy(&in6, sgid.raw, sizeof(in6));
2580 rdma_get_ll_mac(&in6, smac);
2581 memcpy(sqp->ud_header.eth.smac_h, smac, 6);
2582 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2583 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2585 sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2587 sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2588 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2591 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
2592 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2593 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2595 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2596 if (!sqp->qp.ibqp.qp_num)
2597 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2599 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2600 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2601 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2602 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2603 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2604 sqp->qkey : wr->wr.ud.remote_qkey);
2605 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2607 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2610 pr_err("built UD header of size %d:\n", header_size);
2611 for (i = 0; i < header_size / 4; ++i) {
2613 pr_err(" [%02x] ", i * 4);
2615 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2616 if ((i + 1) % 8 == 0)
2623 * Inline data segments may not cross a 64 byte boundary. If
2624 * our UD header is bigger than the space available up to the
2625 * next 64 byte boundary in the WQE, use two inline data
2626 * segments to hold the UD header.
2628 spc = MLX4_INLINE_ALIGN -
2629 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2630 if (header_size <= spc) {
2631 inl->byte_count = cpu_to_be32(1U << 31 | header_size);
2632 memcpy(inl + 1, sqp->header_buf, header_size);
2635 inl->byte_count = cpu_to_be32(1U << 31 | spc);
2636 memcpy(inl + 1, sqp->header_buf, spc);
2638 inl = (void *) (inl + 1) + spc;
2639 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2641 * Need a barrier here to make sure all the data is
2642 * visible before the byte_count field is set.
2643 * Otherwise the HCA prefetcher could grab the 64-byte
2644 * chunk with this inline segment and get a valid (!=
2645 * 0xffffffff) byte count but stale data, and end up
2646 * generating a packet with bad headers.
2648 * The first inline segment's byte_count field doesn't
2649 * need a barrier, because it comes after a
2650 * control/MLX segment and therefore is at an offset
2654 inl->byte_count = cpu_to_be32(1U << 31 | (header_size - spc));
2659 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2663 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2666 struct mlx4_ib_cq *cq;
2668 cur = wq->head - wq->tail;
2669 if (likely(cur + nreq < wq->max_post))
2673 spin_lock(&cq->lock);
2674 cur = wq->head - wq->tail;
2675 spin_unlock(&cq->lock);
2677 return cur + nreq >= wq->max_post;
2680 static __be32 convert_access(int acc)
2682 return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) |
2683 (acc & IB_ACCESS_REMOTE_WRITE ? cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2684 (acc & IB_ACCESS_REMOTE_READ ? cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) |
2685 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) |
2686 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2689 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2691 struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2694 for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2695 mfrpl->mapped_page_list[i] =
2696 cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2697 MLX4_MTT_FLAG_PRESENT);
2699 fseg->flags = convert_access(wr->wr.fast_reg.access_flags);
2700 fseg->mem_key = cpu_to_be32(wr->wr.fast_reg.rkey);
2701 fseg->buf_list = cpu_to_be64(mfrpl->map);
2702 fseg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2703 fseg->reg_len = cpu_to_be64(wr->wr.fast_reg.length);
2704 fseg->offset = 0; /* XXX -- is this just for ZBVA? */
2705 fseg->page_size = cpu_to_be32(wr->wr.fast_reg.page_shift);
2706 fseg->reserved[0] = 0;
2707 fseg->reserved[1] = 0;
2710 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2712 iseg->mem_key = cpu_to_be32(rkey);
2714 iseg->reserved1 = 0;
2715 iseg->reserved2 = 0;
2716 iseg->reserved3[0] = 0;
2717 iseg->reserved3[1] = 0;
2720 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2721 u64 remote_addr, u32 rkey)
2723 rseg->raddr = cpu_to_be64(remote_addr);
2724 rseg->rkey = cpu_to_be32(rkey);
2728 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2730 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2731 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2732 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add);
2733 } else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2734 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2735 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2737 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2743 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2744 struct ib_send_wr *wr)
2746 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2747 aseg->swap_add_mask = cpu_to_be64(wr->wr.atomic.swap_mask);
2748 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add);
2749 aseg->compare_mask = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2752 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2753 struct ib_send_wr *wr)
2755 memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2756 dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2757 dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2758 dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2759 memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2762 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2763 struct mlx4_wqe_datagram_seg *dseg,
2764 struct ib_send_wr *wr, enum ib_qp_type qpt)
2766 union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2767 struct mlx4_av sqp_av = {0};
2768 int port = *((u8 *) &av->ib.port_pd) & 0x3;
2770 /* force loopback */
2771 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2772 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2773 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2774 cpu_to_be32(0xf0000000);
2776 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2777 /* This function used only for sending on QP1 proxies */
2778 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2779 /* Use QKEY from the QP context, which is set by master */
2780 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2783 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2785 struct mlx4_wqe_inline_seg *inl = wqe;
2786 struct mlx4_ib_tunnel_header hdr;
2787 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2791 memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2792 hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2793 hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2794 hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2796 spc = MLX4_INLINE_ALIGN -
2797 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2798 if (sizeof (hdr) <= spc) {
2799 memcpy(inl + 1, &hdr, sizeof (hdr));
2801 inl->byte_count = cpu_to_be32(1U << 31 | sizeof (hdr));
2804 memcpy(inl + 1, &hdr, spc);
2806 inl->byte_count = cpu_to_be32(1U << 31 | spc);
2808 inl = (void *) (inl + 1) + spc;
2809 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2811 inl->byte_count = cpu_to_be32(1U << 31 | (sizeof (hdr) - spc));
2816 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2819 static void set_mlx_icrc_seg(void *dseg)
2822 struct mlx4_wqe_inline_seg *iseg = dseg;
2827 * Need a barrier here before writing the byte_count field to
2828 * make sure that all the data is visible before the
2829 * byte_count field is set. Otherwise, if the segment begins
2830 * a new cacheline, the HCA prefetcher could grab the 64-byte
2831 * chunk and get a valid (!= * 0xffffffff) byte count but
2832 * stale data, and end up sending the wrong data.
2836 iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2839 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2841 dseg->lkey = cpu_to_be32(sg->lkey);
2842 dseg->addr = cpu_to_be64(sg->addr);
2845 * Need a barrier here before writing the byte_count field to
2846 * make sure that all the data is visible before the
2847 * byte_count field is set. Otherwise, if the segment begins
2848 * a new cacheline, the HCA prefetcher could grab the 64-byte
2849 * chunk and get a valid (!= * 0xffffffff) byte count but
2850 * stale data, and end up sending the wrong data.
2854 dseg->byte_count = cpu_to_be32(sg->length);
2857 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2859 dseg->byte_count = cpu_to_be32(sg->length);
2860 dseg->lkey = cpu_to_be32(sg->lkey);
2861 dseg->addr = cpu_to_be64(sg->addr);
2864 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2865 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2866 __be32 *lso_hdr_sz, __be32 *blh)
2868 unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2870 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2871 *blh = cpu_to_be32(1 << 6);
2873 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2874 wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2877 memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2879 *lso_hdr_sz = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2881 *lso_seg_len = halign;
2885 static __be32 send_ieth(struct ib_send_wr *wr)
2887 switch (wr->opcode) {
2888 case IB_WR_SEND_WITH_IMM:
2889 case IB_WR_RDMA_WRITE_WITH_IMM:
2890 return wr->ex.imm_data;
2892 case IB_WR_SEND_WITH_INV:
2893 return cpu_to_be32(wr->ex.invalidate_rkey);
2900 static void add_zero_len_inline(void *wqe)
2902 struct mlx4_wqe_inline_seg *inl = wqe;
2904 inl->byte_count = cpu_to_be32(1 << 31);
2907 static int lay_inline_data(struct mlx4_ib_qp *qp, struct ib_send_wr *wr,
2910 struct mlx4_wqe_inline_seg *seg;
2920 off = ((unsigned long)wqe) & (unsigned long)(MLX4_INLINE_ALIGN - 1);
2924 for (i = 0; i < wr->num_sge; ++i) {
2925 addr = (void *) (unsigned long)(wr->sg_list[i].addr);
2926 len = wr->sg_list[i].length;
2929 if (inl > qp->max_inline_data) {
2934 while (len >= MLX4_INLINE_ALIGN - off) {
2935 to_copy = MLX4_INLINE_ALIGN - off;
2936 memcpy(wqe, addr, to_copy);
2941 wmb(); /* see comment below */
2942 seg->byte_count = htonl(MLX4_INLINE_SEG | seg_len);
2950 memcpy(wqe, addr, len);
2959 * Need a barrier here to make sure
2960 * all the data is visible before the
2961 * byte_count field is set. Otherwise
2962 * the HCA prefetcher could grab the
2963 * 64-byte chunk with this inline
2964 * segment and get a valid (!=
2965 * 0xffffffff) byte count but stale
2966 * data, and end up sending the wrong
2970 seg->byte_count = htonl(MLX4_INLINE_SEG | seg_len);
2973 *sz = (inl + num_seg * sizeof *seg + 15) / 16;
2979 * Avoid using memcpy() to copy to BlueFlame page, since memcpy()
2980 * implementations may use move-string-buffer assembler instructions,
2981 * which do not guarantee order of copying.
2983 static void mlx4_bf_copy(unsigned long *dst, unsigned long *src,
2986 __iowrite64_copy(dst, src, bytecnt / 8);
2989 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2990 struct ib_send_wr **bad_wr)
2992 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2994 struct mlx4_wqe_ctrl_seg *uninitialized_var(ctrl);
2995 struct mlx4_wqe_data_seg *dseg;
2996 unsigned long flags;
3000 int uninitialized_var(stamp);
3001 int uninitialized_var(size);
3002 unsigned uninitialized_var(seglen);
3005 __be32 uninitialized_var(lso_hdr_sz);
3009 spin_lock_irqsave(&qp->sq.lock, flags);
3011 ind = qp->sq_next_wqe;
3013 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3017 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
3023 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3029 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3030 *((u32 *) (&ctrl->vlan_tag)) = 0;
3031 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3034 (wr->send_flags & IB_SEND_SIGNALED ?
3035 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3036 (wr->send_flags & IB_SEND_SOLICITED ?
3037 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3038 ((wr->send_flags & IB_SEND_IP_CSUM) ?
3039 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3040 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3043 ctrl->imm = send_ieth(wr);
3045 wqe += sizeof *ctrl;
3046 size = sizeof *ctrl / 16;
3048 switch (qp->mlx4_ib_qp_type) {
3049 case MLX4_IB_QPT_RC:
3050 case MLX4_IB_QPT_UC:
3051 switch (wr->opcode) {
3052 case IB_WR_ATOMIC_CMP_AND_SWP:
3053 case IB_WR_ATOMIC_FETCH_AND_ADD:
3054 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3055 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
3056 wr->wr.atomic.rkey);
3057 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3059 set_atomic_seg(wqe, wr);
3060 wqe += sizeof (struct mlx4_wqe_atomic_seg);
3062 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3063 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3067 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3068 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
3069 wr->wr.atomic.rkey);
3070 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3072 set_masked_atomic_seg(wqe, wr);
3073 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg);
3075 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3076 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3080 case IB_WR_RDMA_READ:
3081 case IB_WR_RDMA_WRITE:
3082 case IB_WR_RDMA_WRITE_WITH_IMM:
3083 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
3085 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3086 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3089 case IB_WR_LOCAL_INV:
3090 ctrl->srcrb_flags |=
3091 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3092 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3093 wqe += sizeof (struct mlx4_wqe_local_inval_seg);
3094 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3097 case IB_WR_FAST_REG_MR:
3098 ctrl->srcrb_flags |=
3099 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3100 set_fmr_seg(wqe, wr);
3101 wqe += sizeof (struct mlx4_wqe_fmr_seg);
3102 size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
3106 /* No extra segments required for sends */
3111 case MLX4_IB_QPT_TUN_SMI_OWNER:
3112 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
3113 if (unlikely(err)) {
3118 size += seglen / 16;
3120 case MLX4_IB_QPT_TUN_SMI:
3121 case MLX4_IB_QPT_TUN_GSI:
3122 /* this is a UD qp used in MAD responses to slaves. */
3123 set_datagram_seg(wqe, wr);
3124 /* set the forced-loopback bit in the data seg av */
3125 *(__be32 *) wqe |= cpu_to_be32(0x80000000);
3126 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3127 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3129 case MLX4_IB_QPT_UD:
3130 set_datagram_seg(wqe, wr);
3131 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3132 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3134 if (wr->opcode == IB_WR_LSO) {
3135 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
3136 if (unlikely(err)) {
3140 lso_wqe = (__be32 *) wqe;
3142 size += seglen / 16;
3146 case MLX4_IB_QPT_PROXY_SMI_OWNER:
3147 if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) {
3152 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
3153 if (unlikely(err)) {
3158 size += seglen / 16;
3159 /* to start tunnel header on a cache-line boundary */
3160 add_zero_len_inline(wqe);
3163 build_tunnel_header(wr, wqe, &seglen);
3165 size += seglen / 16;
3167 case MLX4_IB_QPT_PROXY_SMI:
3168 /* don't allow QP0 sends on guests */
3172 case MLX4_IB_QPT_PROXY_GSI:
3173 /* If we are tunneling special qps, this is a UD qp.
3174 * In this case we first add a UD segment targeting
3175 * the tunnel qp, and then add a header with address
3177 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type);
3178 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3179 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3180 build_tunnel_header(wr, wqe, &seglen);
3182 size += seglen / 16;
3185 case MLX4_IB_QPT_SMI:
3186 case MLX4_IB_QPT_GSI:
3187 err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
3188 if (unlikely(err)) {
3193 size += seglen / 16;
3201 * Write data segments in reverse order, so as to
3202 * overwrite cacheline stamp last within each
3203 * cacheline. This avoids issues with WQE
3207 dseg += wr->num_sge - 1;
3209 /* Add one more inline data segment for ICRC for MLX sends */
3210 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3211 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3212 qp->mlx4_ib_qp_type &
3213 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3214 set_mlx_icrc_seg(dseg + 1);
3215 size += sizeof (struct mlx4_wqe_data_seg) / 16;
3218 if (wr->send_flags & IB_SEND_INLINE && wr->num_sge) {
3220 err = lay_inline_data(qp, wr, wqe, &sz);
3226 size += wr->num_sge *
3227 (sizeof(struct mlx4_wqe_data_seg) / 16);
3228 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3229 set_data_seg(dseg, wr->sg_list + i);
3233 * Possibly overwrite stamping in cacheline with LSO
3234 * segment only after making sure all data segments
3238 *lso_wqe = lso_hdr_sz;
3239 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
3240 MLX4_WQE_CTRL_FENCE : 0) | size;
3243 * Make sure descriptor is fully written before
3244 * setting ownership bit (because HW can start
3245 * executing as soon as we do).
3249 if (wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3255 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3256 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
3258 stamp = ind + qp->sq_spare_wqes;
3259 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3262 * We can improve latency by not stamping the last
3263 * send queue WQE until after ringing the doorbell, so
3264 * only stamp here if there are still more WQEs to post.
3266 * Same optimization applies to padding with NOP wqe
3267 * in case of WQE shrinking (used to prevent wrap-around
3268 * in the middle of WR).
3271 stamp_send_wqe(qp, stamp, size * 16);
3272 ind = pad_wraparound(qp, ind);
3277 if (nreq == 1 && inl && size > 1 && size < qp->bf.buf_size / 16) {
3278 ctrl->owner_opcode |= htonl((qp->sq_next_wqe & 0xffff) << 8);
3279 /* We set above doorbell_qpn bits to 0 as part of vlan
3280 * tag initialization, so |= should be correct.
3282 *(u32 *) (&ctrl->vlan_tag) |= qp->doorbell_qpn;
3284 * Make sure that descriptor is written to memory
3285 * before writing to BlueFlame page.
3291 mlx4_bf_copy(qp->bf.reg + qp->bf.offset, (unsigned long *) ctrl,
3292 ALIGN(size * 16, 64));
3295 qp->bf.offset ^= qp->bf.buf_size;
3298 qp->sq.head += nreq;
3301 * Make sure that descriptors are written before
3306 writel(qp->doorbell_qpn, qp->bf.uar->map + MLX4_SEND_DOORBELL);
3309 * Make sure doorbells don't leak out of SQ spinlock
3310 * and reach the HCA out of order.
3317 stamp_send_wqe(qp, stamp, size * 16);
3318 ind = pad_wraparound(qp, ind);
3319 qp->sq_next_wqe = ind;
3322 spin_unlock_irqrestore(&qp->sq.lock, flags);
3327 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3328 struct ib_recv_wr **bad_wr)
3330 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3331 struct mlx4_wqe_data_seg *scat;
3332 unsigned long flags;
3339 max_gs = qp->rq.max_gs;
3340 spin_lock_irqsave(&qp->rq.lock, flags);
3342 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3344 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3345 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3351 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3357 scat = get_recv_wqe(qp, ind);
3359 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3360 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3361 ib_dma_sync_single_for_device(ibqp->device,
3362 qp->sqp_proxy_rcv[ind].map,
3363 sizeof (struct mlx4_ib_proxy_sqp_hdr),
3366 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3367 /* use dma lkey from upper layer entry */
3368 scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3369 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3374 for (i = 0; i < wr->num_sge; ++i)
3375 __set_data_seg(scat + i, wr->sg_list + i);
3378 scat[i].byte_count = 0;
3379 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
3383 qp->rq.wrid[ind] = wr->wr_id;
3385 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3390 qp->rq.head += nreq;
3393 * Make sure that descriptors are written before
3398 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3401 spin_unlock_irqrestore(&qp->rq.lock, flags);
3406 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3408 switch (mlx4_state) {
3409 case MLX4_QP_STATE_RST: return IB_QPS_RESET;
3410 case MLX4_QP_STATE_INIT: return IB_QPS_INIT;
3411 case MLX4_QP_STATE_RTR: return IB_QPS_RTR;
3412 case MLX4_QP_STATE_RTS: return IB_QPS_RTS;
3413 case MLX4_QP_STATE_SQ_DRAINING:
3414 case MLX4_QP_STATE_SQD: return IB_QPS_SQD;
3415 case MLX4_QP_STATE_SQER: return IB_QPS_SQE;
3416 case MLX4_QP_STATE_ERR: return IB_QPS_ERR;
3421 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3423 switch (mlx4_mig_state) {
3424 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED;
3425 case MLX4_QP_PM_REARM: return IB_MIG_REARM;
3426 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
3431 static int to_ib_qp_access_flags(int mlx4_flags)
3435 if (mlx4_flags & MLX4_QP_BIT_RRE)
3436 ib_flags |= IB_ACCESS_REMOTE_READ;
3437 if (mlx4_flags & MLX4_QP_BIT_RWE)
3438 ib_flags |= IB_ACCESS_REMOTE_WRITE;
3439 if (mlx4_flags & MLX4_QP_BIT_RAE)
3440 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3445 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3446 struct mlx4_qp_path *path)
3448 struct mlx4_dev *dev = ibdev->dev;
3451 memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3452 ib_ah_attr->port_num = path->sched_queue & 0x40 ? 2 : 1;
3454 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3457 is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3458 IB_LINK_LAYER_ETHERNET;
3460 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3461 ((path->sched_queue & 4) << 1);
3463 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3465 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
3466 ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3467 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
3468 ib_ah_attr->ah_flags = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3469 if (ib_ah_attr->ah_flags) {
3470 ib_ah_attr->grh.sgid_index = path->mgid_index;
3471 ib_ah_attr->grh.hop_limit = path->hop_limit;
3472 ib_ah_attr->grh.traffic_class =
3473 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3474 ib_ah_attr->grh.flow_label =
3475 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3476 memcpy(ib_ah_attr->grh.dgid.raw,
3477 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3481 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3482 struct ib_qp_init_attr *qp_init_attr)
3484 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3485 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3486 struct mlx4_qp_context context;
3490 mutex_lock(&qp->mutex);
3492 if (qp->state == IB_QPS_RESET) {
3493 qp_attr->qp_state = IB_QPS_RESET;
3497 err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3503 mlx4_state = be32_to_cpu(context.flags) >> 28;
3505 qp->state = to_ib_qp_state(mlx4_state);
3506 qp_attr->qp_state = qp->state;
3507 qp_attr->path_mtu = context.mtu_msgmax >> 5;
3508 qp_attr->path_mig_state =
3509 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3510 qp_attr->qkey = be32_to_cpu(context.qkey);
3511 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3512 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff;
3513 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff;
3514 qp_attr->qp_access_flags =
3515 to_ib_qp_access_flags(be32_to_cpu(context.params2));
3517 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3518 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3519 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3520 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3521 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
3524 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3525 if (qp_attr->qp_state == IB_QPS_INIT)
3526 qp_attr->port_num = qp->port;
3528 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3530 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3531 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3533 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3535 qp_attr->max_dest_rd_atomic =
3536 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3537 qp_attr->min_rnr_timer =
3538 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3539 qp_attr->timeout = context.pri_path.ackto >> 3;
3540 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7;
3541 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7;
3542 qp_attr->alt_timeout = context.alt_path.ackto >> 3;
3545 qp_attr->cur_qp_state = qp_attr->qp_state;
3546 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
3547 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
3549 if (!ibqp->uobject) {
3550 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
3551 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3553 qp_attr->cap.max_send_wr = 0;
3554 qp_attr->cap.max_send_sge = 0;
3558 * We don't support inline sends for kernel QPs (yet), and we
3559 * don't know what userspace's value should be.
3561 qp_attr->cap.max_inline_data = 0;
3563 qp_init_attr->cap = qp_attr->cap;
3565 qp_init_attr->create_flags = 0;
3566 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3567 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3569 if (qp->flags & MLX4_IB_QP_LSO)
3570 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3572 if (qp->flags & MLX4_IB_QP_NETIF)
3573 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3575 qp_init_attr->sq_sig_type =
3576 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3577 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3579 qp_init_attr->qpg_type = ibqp->qpg_type;
3580 if (ibqp->qpg_type == IB_QPG_PARENT)
3581 qp_init_attr->cap.qpg_tss_mask_sz = qp->qpg_data->qpg_tss_mask_sz;
3583 qp_init_attr->cap.qpg_tss_mask_sz = 0;
3586 mutex_unlock(&qp->mutex);