2 * Copyright (c) 2016 Hisilicon Limited.
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/platform_device.h>
35 #include <rdma/ib_addr.h>
36 #include <rdma/ib_umem.h>
37 #include "hns_roce_common.h"
38 #include "hns_roce_device.h"
39 #include "hns_roce_hem.h"
40 #include <rdma/hns-abi.h>
42 #define SQP_NUM (2 * HNS_ROCE_MAX_PORTS)
44 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
46 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
47 struct device *dev = hr_dev->dev;
48 struct hns_roce_qp *qp;
50 spin_lock(&qp_table->lock);
52 qp = __hns_roce_qp_lookup(hr_dev, qpn);
54 atomic_inc(&qp->refcount);
56 spin_unlock(&qp_table->lock);
59 dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
63 qp->event(qp, (enum hns_roce_event)event_type);
65 if (atomic_dec_and_test(&qp->refcount))
68 EXPORT_SYMBOL_GPL(hns_roce_qp_event);
70 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
71 enum hns_roce_event type)
73 struct ib_event event;
74 struct ib_qp *ibqp = &hr_qp->ibqp;
76 if (ibqp->event_handler) {
77 event.device = ibqp->device;
78 event.element.qp = ibqp;
80 case HNS_ROCE_EVENT_TYPE_PATH_MIG:
81 event.event = IB_EVENT_PATH_MIG;
83 case HNS_ROCE_EVENT_TYPE_COMM_EST:
84 event.event = IB_EVENT_COMM_EST;
86 case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
87 event.event = IB_EVENT_SQ_DRAINED;
89 case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
90 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
92 case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
93 event.event = IB_EVENT_QP_FATAL;
95 case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
96 event.event = IB_EVENT_PATH_MIG_ERR;
98 case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
99 event.event = IB_EVENT_QP_REQ_ERR;
101 case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
102 event.event = IB_EVENT_QP_ACCESS_ERR;
105 dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
109 ibqp->event_handler(&event, ibqp->qp_context);
113 static int hns_roce_reserve_range_qp(struct hns_roce_dev *hr_dev, int cnt,
114 int align, unsigned long *base)
116 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
118 return hns_roce_bitmap_alloc_range(&qp_table->bitmap, cnt, align, base);
121 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
125 return HNS_ROCE_QP_STATE_RST;
127 return HNS_ROCE_QP_STATE_INIT;
129 return HNS_ROCE_QP_STATE_RTR;
131 return HNS_ROCE_QP_STATE_RTS;
133 return HNS_ROCE_QP_STATE_SQD;
135 return HNS_ROCE_QP_STATE_ERR;
137 return HNS_ROCE_QP_NUM_STATE;
140 EXPORT_SYMBOL_GPL(to_hns_roce_state);
142 static int hns_roce_gsi_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
143 struct hns_roce_qp *hr_qp)
145 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
153 spin_lock_irq(&qp_table->lock);
154 ret = radix_tree_insert(&hr_dev->qp_table_tree,
155 hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
156 spin_unlock_irq(&qp_table->lock);
158 dev_err(hr_dev->dev, "QPC radix_tree_insert failed\n");
162 atomic_set(&hr_qp->refcount, 1);
163 init_completion(&hr_qp->free);
172 static int hns_roce_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
173 struct hns_roce_qp *hr_qp)
175 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
176 struct device *dev = hr_dev->dev;
184 /* Alloc memory for QPC */
185 ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
187 dev_err(dev, "QPC table get failed\n");
191 /* Alloc memory for IRRL */
192 ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
194 dev_err(dev, "IRRL table get failed\n");
198 if (hr_dev->caps.trrl_entry_sz) {
199 /* Alloc memory for TRRL */
200 ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
203 dev_err(dev, "TRRL table get failed\n");
208 spin_lock_irq(&qp_table->lock);
209 ret = radix_tree_insert(&hr_dev->qp_table_tree,
210 hr_qp->qpn & (hr_dev->caps.num_qps - 1), hr_qp);
211 spin_unlock_irq(&qp_table->lock);
213 dev_err(dev, "QPC radix_tree_insert failed\n");
217 atomic_set(&hr_qp->refcount, 1);
218 init_completion(&hr_qp->free);
223 if (hr_dev->caps.trrl_entry_sz)
224 hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
227 hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
230 hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
236 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
238 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
241 spin_lock_irqsave(&qp_table->lock, flags);
242 radix_tree_delete(&hr_dev->qp_table_tree,
243 hr_qp->qpn & (hr_dev->caps.num_qps - 1));
244 spin_unlock_irqrestore(&qp_table->lock, flags);
246 EXPORT_SYMBOL_GPL(hns_roce_qp_remove);
248 void hns_roce_qp_free(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
250 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
252 if (atomic_dec_and_test(&hr_qp->refcount))
253 complete(&hr_qp->free);
254 wait_for_completion(&hr_qp->free);
256 if ((hr_qp->ibqp.qp_type) != IB_QPT_GSI) {
257 if (hr_dev->caps.trrl_entry_sz)
258 hns_roce_table_put(hr_dev, &qp_table->trrl_table,
260 hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
261 hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
264 EXPORT_SYMBOL_GPL(hns_roce_qp_free);
266 void hns_roce_release_range_qp(struct hns_roce_dev *hr_dev, int base_qpn,
269 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
271 if (base_qpn < SQP_NUM)
274 hns_roce_bitmap_free_range(&qp_table->bitmap, base_qpn, cnt, BITMAP_RR);
276 EXPORT_SYMBOL_GPL(hns_roce_release_range_qp);
278 static int hns_roce_set_rq_size(struct hns_roce_dev *hr_dev,
279 struct ib_qp_cap *cap, int is_user, int has_srq,
280 struct hns_roce_qp *hr_qp)
282 struct device *dev = hr_dev->dev;
285 /* Check the validity of QP support capacity */
286 if (cap->max_recv_wr > hr_dev->caps.max_wqes ||
287 cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
288 dev_err(dev, "RQ WR or sge error!max_recv_wr=%d max_recv_sge=%d\n",
289 cap->max_recv_wr, cap->max_recv_sge);
293 /* If srq exit, set zero for relative number of rq */
295 if (cap->max_recv_wr) {
296 dev_dbg(dev, "srq no need config max_recv_wr\n");
300 hr_qp->rq.wqe_cnt = hr_qp->rq.max_gs = 0;
302 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) {
303 dev_err(dev, "user space no need config max_recv_wr max_recv_sge\n");
307 if (hr_dev->caps.min_wqes)
308 max_cnt = max(cap->max_recv_wr, hr_dev->caps.min_wqes);
310 max_cnt = cap->max_recv_wr;
312 hr_qp->rq.wqe_cnt = roundup_pow_of_two(max_cnt);
314 if ((u32)hr_qp->rq.wqe_cnt > hr_dev->caps.max_wqes) {
315 dev_err(dev, "while setting rq size, rq.wqe_cnt too large\n");
319 max_cnt = max(1U, cap->max_recv_sge);
320 hr_qp->rq.max_gs = roundup_pow_of_two(max_cnt);
321 if (hr_dev->caps.max_rq_sg <= 2)
322 hr_qp->rq.wqe_shift =
323 ilog2(hr_dev->caps.max_rq_desc_sz);
325 hr_qp->rq.wqe_shift =
326 ilog2(hr_dev->caps.max_rq_desc_sz
330 cap->max_recv_wr = hr_qp->rq.max_post = hr_qp->rq.wqe_cnt;
331 cap->max_recv_sge = hr_qp->rq.max_gs;
336 static int hns_roce_set_user_sq_size(struct hns_roce_dev *hr_dev,
337 struct ib_qp_cap *cap,
338 struct hns_roce_qp *hr_qp,
339 struct hns_roce_ib_create_qp *ucmd)
341 u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
342 u8 max_sq_stride = ilog2(roundup_sq_stride);
346 /* Sanity check SQ size before proceeding */
347 if ((u32)(1 << ucmd->log_sq_bb_count) > hr_dev->caps.max_wqes ||
348 ucmd->log_sq_stride > max_sq_stride ||
349 ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
350 dev_err(hr_dev->dev, "check SQ size error!\n");
354 if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
355 dev_err(hr_dev->dev, "SQ sge error! max_send_sge=%d\n",
360 hr_qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
361 hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
363 max_cnt = max(1U, cap->max_send_sge);
364 if (hr_dev->caps.max_sq_sg <= 2)
365 hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
367 hr_qp->sq.max_gs = max_cnt;
369 if (hr_qp->sq.max_gs > 2)
370 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
371 (hr_qp->sq.max_gs - 2));
372 hr_qp->sge.sge_shift = 4;
374 /* Get buf size, SQ and RQ are aligned to page_szie */
375 if (hr_dev->caps.max_sq_sg <= 2) {
376 hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
377 hr_qp->rq.wqe_shift), PAGE_SIZE) +
378 HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
379 hr_qp->sq.wqe_shift), PAGE_SIZE);
381 hr_qp->sq.offset = 0;
382 hr_qp->rq.offset = HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
383 hr_qp->sq.wqe_shift), PAGE_SIZE);
385 page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
386 hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
387 hr_qp->rq.wqe_shift), page_size) +
388 HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
389 hr_qp->sge.sge_shift), page_size) +
390 HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
391 hr_qp->sq.wqe_shift), page_size);
393 hr_qp->sq.offset = 0;
394 if (hr_qp->sge.sge_cnt) {
395 hr_qp->sge.offset = HNS_ROCE_ALOGN_UP(
396 (hr_qp->sq.wqe_cnt <<
397 hr_qp->sq.wqe_shift),
399 hr_qp->rq.offset = hr_qp->sge.offset +
400 HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
401 hr_qp->sge.sge_shift),
404 hr_qp->rq.offset = HNS_ROCE_ALOGN_UP(
405 (hr_qp->sq.wqe_cnt <<
406 hr_qp->sq.wqe_shift),
414 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
415 struct ib_qp_cap *cap,
416 struct hns_roce_qp *hr_qp)
418 struct device *dev = hr_dev->dev;
423 if (cap->max_send_wr > hr_dev->caps.max_wqes ||
424 cap->max_send_sge > hr_dev->caps.max_sq_sg ||
425 cap->max_inline_data > hr_dev->caps.max_sq_inline) {
426 dev_err(dev, "SQ WR or sge or inline data error!\n");
430 hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
431 hr_qp->sq_max_wqes_per_wr = 1;
432 hr_qp->sq_spare_wqes = 0;
434 if (hr_dev->caps.min_wqes)
435 max_cnt = max(cap->max_send_wr, hr_dev->caps.min_wqes);
437 max_cnt = cap->max_send_wr;
439 hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
440 if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
441 dev_err(dev, "while setting kernel sq size, sq.wqe_cnt too large\n");
445 /* Get data_seg numbers */
446 max_cnt = max(1U, cap->max_send_sge);
447 if (hr_dev->caps.max_sq_sg <= 2)
448 hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
450 hr_qp->sq.max_gs = max_cnt;
452 if (hr_qp->sq.max_gs > 2) {
453 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
454 (hr_qp->sq.max_gs - 2));
455 hr_qp->sge.sge_shift = 4;
458 /* ud sqwqe's sge use extend sge */
459 if (hr_dev->caps.max_sq_sg > 2 && hr_qp->ibqp.qp_type == IB_QPT_GSI) {
460 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
462 hr_qp->sge.sge_shift = 4;
465 /* Get buf size, SQ and RQ are aligned to PAGE_SIZE */
466 page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
467 hr_qp->sq.offset = 0;
468 size = HNS_ROCE_ALOGN_UP(hr_qp->sq.wqe_cnt << hr_qp->sq.wqe_shift,
471 if (hr_dev->caps.max_sq_sg > 2 && hr_qp->sge.sge_cnt) {
472 hr_qp->sge.offset = size;
473 size += HNS_ROCE_ALOGN_UP(hr_qp->sge.sge_cnt <<
474 hr_qp->sge.sge_shift, page_size);
477 hr_qp->rq.offset = size;
478 size += HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt << hr_qp->rq.wqe_shift),
480 hr_qp->buff_size = size;
482 /* Get wr and sge number which send */
483 cap->max_send_wr = hr_qp->sq.max_post = hr_qp->sq.wqe_cnt;
484 cap->max_send_sge = hr_qp->sq.max_gs;
486 /* We don't support inline sends for kernel QPs (yet) */
487 cap->max_inline_data = 0;
492 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
494 if (attr->qp_type == IB_QPT_XRC_INI ||
495 attr->qp_type == IB_QPT_XRC_TGT || attr->srq)
501 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
503 struct ib_qp_init_attr *init_attr,
504 struct ib_udata *udata, unsigned long sqpn,
505 struct hns_roce_qp *hr_qp)
507 struct device *dev = hr_dev->dev;
508 struct hns_roce_ib_create_qp ucmd;
509 struct hns_roce_ib_create_qp_resp resp = {};
510 unsigned long qpn = 0;
516 mutex_init(&hr_qp->mutex);
517 spin_lock_init(&hr_qp->sq.lock);
518 spin_lock_init(&hr_qp->rq.lock);
520 hr_qp->state = IB_QPS_RESET;
522 hr_qp->ibqp.qp_type = init_attr->qp_type;
524 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
525 hr_qp->sq_signal_bits = cpu_to_le32(IB_SIGNAL_ALL_WR);
527 hr_qp->sq_signal_bits = cpu_to_le32(IB_SIGNAL_REQ_WR);
529 ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, !!ib_pd->uobject,
530 !!init_attr->srq, hr_qp);
532 dev_err(dev, "hns_roce_set_rq_size failed\n");
536 if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) {
537 /* allocate recv inline buf */
538 hr_qp->rq_inl_buf.wqe_list = kcalloc(hr_qp->rq.wqe_cnt,
539 sizeof(struct hns_roce_rinl_wqe),
541 if (!hr_qp->rq_inl_buf.wqe_list) {
546 hr_qp->rq_inl_buf.wqe_cnt = hr_qp->rq.wqe_cnt;
548 /* Firstly, allocate a list of sge space buffer */
549 hr_qp->rq_inl_buf.wqe_list[0].sg_list =
550 kcalloc(hr_qp->rq_inl_buf.wqe_cnt,
551 init_attr->cap.max_recv_sge *
552 sizeof(struct hns_roce_rinl_sge),
554 if (!hr_qp->rq_inl_buf.wqe_list[0].sg_list) {
559 for (i = 1; i < hr_qp->rq_inl_buf.wqe_cnt; i++)
560 /* Secondly, reallocate the buffer */
561 hr_qp->rq_inl_buf.wqe_list[i].sg_list =
562 &hr_qp->rq_inl_buf.wqe_list[0].sg_list[i *
563 init_attr->cap.max_recv_sge];
566 if (ib_pd->uobject) {
567 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
568 dev_err(dev, "ib_copy_from_udata error for create qp\n");
570 goto err_rq_sge_list;
573 ret = hns_roce_set_user_sq_size(hr_dev, &init_attr->cap, hr_qp,
576 dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
577 goto err_rq_sge_list;
580 hr_qp->umem = ib_umem_get(ib_pd->uobject->context,
581 ucmd.buf_addr, hr_qp->buff_size, 0,
583 if (IS_ERR(hr_qp->umem)) {
584 dev_err(dev, "ib_umem_get error for create qp\n");
585 ret = PTR_ERR(hr_qp->umem);
586 goto err_rq_sge_list;
589 hr_qp->mtt.mtt_type = MTT_TYPE_WQE;
590 if (hr_dev->caps.mtt_buf_pg_sz) {
591 npages = (ib_umem_page_count(hr_qp->umem) +
592 (1 << hr_dev->caps.mtt_buf_pg_sz) - 1) /
593 (1 << hr_dev->caps.mtt_buf_pg_sz);
594 page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
595 ret = hns_roce_mtt_init(hr_dev, npages,
599 ret = hns_roce_mtt_init(hr_dev,
600 ib_umem_page_count(hr_qp->umem),
601 hr_qp->umem->page_shift,
605 dev_err(dev, "hns_roce_mtt_init error for create qp\n");
609 ret = hns_roce_ib_umem_write_mtt(hr_dev, &hr_qp->mtt,
612 dev_err(dev, "hns_roce_ib_umem_write_mtt error for create qp\n");
616 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
617 (udata->outlen >= sizeof(resp)) &&
618 hns_roce_qp_has_rq(init_attr)) {
619 ret = hns_roce_db_map_user(
620 to_hr_ucontext(ib_pd->uobject->context),
621 ucmd.db_addr, &hr_qp->rdb);
623 dev_err(dev, "rq record doorbell map failed!\n");
628 if (init_attr->create_flags &
629 IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
630 dev_err(dev, "init_attr->create_flags error!\n");
632 goto err_rq_sge_list;
635 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
636 dev_err(dev, "init_attr->create_flags error!\n");
638 goto err_rq_sge_list;
642 ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
645 dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
646 goto err_rq_sge_list;
649 /* QP doorbell register address */
650 hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
651 DB_REG_OFFSET * hr_dev->priv_uar.index;
652 hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
653 DB_REG_OFFSET * hr_dev->priv_uar.index;
655 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
656 hns_roce_qp_has_rq(init_attr)) {
657 ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
659 dev_err(dev, "rq record doorbell alloc failed!\n");
660 goto err_rq_sge_list;
662 *hr_qp->rdb.db_record = 0;
665 /* Allocate QP buf */
666 page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
667 if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size,
668 (1 << page_shift) * 2,
669 &hr_qp->hr_buf, page_shift)) {
670 dev_err(dev, "hns_roce_buf_alloc error!\n");
675 hr_qp->mtt.mtt_type = MTT_TYPE_WQE;
677 ret = hns_roce_mtt_init(hr_dev, hr_qp->hr_buf.npages,
678 hr_qp->hr_buf.page_shift, &hr_qp->mtt);
680 dev_err(dev, "hns_roce_mtt_init error for kernel create qp\n");
684 ret = hns_roce_buf_write_mtt(hr_dev, &hr_qp->mtt,
687 dev_err(dev, "hns_roce_buf_write_mtt error for kernel create qp\n");
691 hr_qp->sq.wrid = kmalloc_array(hr_qp->sq.wqe_cnt, sizeof(u64),
693 hr_qp->rq.wrid = kmalloc_array(hr_qp->rq.wqe_cnt, sizeof(u64),
695 if (!hr_qp->sq.wrid || !hr_qp->rq.wrid) {
705 ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
707 dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
712 if (init_attr->qp_type == IB_QPT_GSI &&
713 hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
714 /* In v1 engine, GSI QP context in RoCE engine's register */
715 ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
717 dev_err(dev, "hns_roce_qp_alloc failed!\n");
721 ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
723 dev_err(dev, "hns_roce_qp_alloc failed!\n");
729 hr_qp->doorbell_qpn = 1;
731 hr_qp->doorbell_qpn = cpu_to_le64(hr_qp->qpn);
733 if (ib_pd->uobject && (udata->outlen >= sizeof(resp)) &&
734 (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB)) {
736 /* indicate kernel supports record db */
737 resp.cap_flags |= HNS_ROCE_SUPPORT_RQ_RECORD_DB;
738 ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
744 hr_qp->event = hns_roce_ib_qp_event;
749 if (init_attr->qp_type == IB_QPT_GSI &&
750 hr_dev->hw_rev == HNS_ROCE_HW_VER1)
751 hns_roce_qp_remove(hr_dev, hr_qp);
753 hns_roce_qp_free(hr_dev, hr_qp);
757 hns_roce_release_range_qp(hr_dev, qpn, 1);
760 if (ib_pd->uobject) {
761 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
762 (udata->outlen >= sizeof(resp)) &&
763 hns_roce_qp_has_rq(init_attr))
764 hns_roce_db_unmap_user(
765 to_hr_ucontext(ib_pd->uobject->context),
768 kfree(hr_qp->sq.wrid);
769 kfree(hr_qp->rq.wrid);
773 hns_roce_mtt_cleanup(hr_dev, &hr_qp->mtt);
777 ib_umem_release(hr_qp->umem);
779 hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
782 if (!ib_pd->uobject && hns_roce_qp_has_rq(init_attr) &&
783 (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB))
784 hns_roce_free_db(hr_dev, &hr_qp->rdb);
787 if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE)
788 kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
791 if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE)
792 kfree(hr_qp->rq_inl_buf.wqe_list);
798 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
799 struct ib_qp_init_attr *init_attr,
800 struct ib_udata *udata)
802 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
803 struct device *dev = hr_dev->dev;
804 struct hns_roce_sqp *hr_sqp;
805 struct hns_roce_qp *hr_qp;
808 switch (init_attr->qp_type) {
810 hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
812 return ERR_PTR(-ENOMEM);
814 ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
817 dev_err(dev, "Create RC QP failed\n");
822 hr_qp->ibqp.qp_num = hr_qp->qpn;
827 /* Userspace is not allowed to create special QPs: */
829 dev_err(dev, "not support usr space GSI\n");
830 return ERR_PTR(-EINVAL);
833 hr_sqp = kzalloc(sizeof(*hr_sqp), GFP_KERNEL);
835 return ERR_PTR(-ENOMEM);
837 hr_qp = &hr_sqp->hr_qp;
838 hr_qp->port = init_attr->port_num - 1;
839 hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
841 /* when hw version is v1, the sqpn is allocated */
842 if (hr_dev->caps.max_sq_sg <= 2)
843 hr_qp->ibqp.qp_num = HNS_ROCE_MAX_PORTS +
844 hr_dev->iboe.phy_port[hr_qp->port];
846 hr_qp->ibqp.qp_num = 1;
848 ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
849 hr_qp->ibqp.qp_num, hr_qp);
851 dev_err(dev, "Create GSI QP failed!\n");
859 dev_err(dev, "not support QP type %d\n", init_attr->qp_type);
860 return ERR_PTR(-EINVAL);
866 EXPORT_SYMBOL_GPL(hns_roce_create_qp);
868 int to_hr_qp_type(int qp_type)
872 if (qp_type == IB_QPT_RC)
873 transport_type = SERV_TYPE_RC;
874 else if (qp_type == IB_QPT_UC)
875 transport_type = SERV_TYPE_UC;
876 else if (qp_type == IB_QPT_UD)
877 transport_type = SERV_TYPE_UD;
878 else if (qp_type == IB_QPT_GSI)
879 transport_type = SERV_TYPE_UD;
883 return transport_type;
885 EXPORT_SYMBOL_GPL(to_hr_qp_type);
887 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
888 int attr_mask, struct ib_udata *udata)
890 struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
891 struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
892 enum ib_qp_state cur_state, new_state;
893 struct device *dev = hr_dev->dev;
896 enum ib_mtu active_mtu;
898 mutex_lock(&hr_qp->mutex);
900 cur_state = attr_mask & IB_QP_CUR_STATE ?
901 attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
902 new_state = attr_mask & IB_QP_STATE ?
903 attr->qp_state : cur_state;
905 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
906 IB_LINK_LAYER_ETHERNET)) {
907 dev_err(dev, "ib_modify_qp_is_ok failed\n");
911 if ((attr_mask & IB_QP_PORT) &&
912 (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
913 dev_err(dev, "attr port_num invalid.attr->port_num=%d\n",
918 if (attr_mask & IB_QP_PKEY_INDEX) {
919 p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
920 if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
921 dev_err(dev, "attr pkey_index invalid.attr->pkey_index=%d\n",
927 if (attr_mask & IB_QP_PATH_MTU) {
928 p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
929 active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
931 if ((hr_dev->caps.max_mtu == IB_MTU_4096 &&
932 attr->path_mtu > IB_MTU_4096) ||
933 (hr_dev->caps.max_mtu == IB_MTU_2048 &&
934 attr->path_mtu > IB_MTU_2048) ||
935 attr->path_mtu < IB_MTU_256 ||
936 attr->path_mtu > active_mtu) {
937 dev_err(dev, "attr path_mtu(%d)invalid while modify qp",
943 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
944 attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
945 dev_err(dev, "attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
946 attr->max_rd_atomic);
950 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
951 attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
952 dev_err(dev, "attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
953 attr->max_dest_rd_atomic);
957 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
962 ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
966 mutex_unlock(&hr_qp->mutex);
971 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
972 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
974 if (send_cq == recv_cq) {
975 spin_lock_irq(&send_cq->lock);
976 __acquire(&recv_cq->lock);
977 } else if (send_cq->cqn < recv_cq->cqn) {
978 spin_lock_irq(&send_cq->lock);
979 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
981 spin_lock_irq(&recv_cq->lock);
982 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
985 EXPORT_SYMBOL_GPL(hns_roce_lock_cqs);
987 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
988 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
989 __releases(&recv_cq->lock)
991 if (send_cq == recv_cq) {
992 __release(&recv_cq->lock);
993 spin_unlock_irq(&send_cq->lock);
994 } else if (send_cq->cqn < recv_cq->cqn) {
995 spin_unlock(&recv_cq->lock);
996 spin_unlock_irq(&send_cq->lock);
998 spin_unlock(&send_cq->lock);
999 spin_unlock_irq(&recv_cq->lock);
1002 EXPORT_SYMBOL_GPL(hns_roce_unlock_cqs);
1004 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1007 return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
1010 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1012 return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1014 EXPORT_SYMBOL_GPL(get_recv_wqe);
1016 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1018 return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1020 EXPORT_SYMBOL_GPL(get_send_wqe);
1022 void *get_send_extend_sge(struct hns_roce_qp *hr_qp, int n)
1024 return hns_roce_buf_offset(&hr_qp->hr_buf, hr_qp->sge.offset +
1025 (n << hr_qp->sge.sge_shift));
1027 EXPORT_SYMBOL_GPL(get_send_extend_sge);
1029 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1030 struct ib_cq *ib_cq)
1032 struct hns_roce_cq *hr_cq;
1035 cur = hr_wq->head - hr_wq->tail;
1036 if (likely(cur + nreq < hr_wq->max_post))
1039 hr_cq = to_hr_cq(ib_cq);
1040 spin_lock(&hr_cq->lock);
1041 cur = hr_wq->head - hr_wq->tail;
1042 spin_unlock(&hr_cq->lock);
1044 return cur + nreq >= hr_wq->max_post;
1046 EXPORT_SYMBOL_GPL(hns_roce_wq_overflow);
1048 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1050 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1051 int reserved_from_top = 0;
1054 spin_lock_init(&qp_table->lock);
1055 INIT_RADIX_TREE(&hr_dev->qp_table_tree, GFP_ATOMIC);
1057 /* A port include two SQP, six port total 12 */
1058 ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1059 hr_dev->caps.num_qps - 1, SQP_NUM,
1062 dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1070 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1072 hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);