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/pci.h>
35 #include <linux/platform_device.h>
36 #include <rdma/ib_addr.h>
37 #include <rdma/ib_umem.h>
38 #include <rdma/uverbs_ioctl.h>
39 #include "hns_roce_common.h"
40 #include "hns_roce_device.h"
41 #include "hns_roce_hem.h"
42 #include <rdma/hns-abi.h>
44 #define SQP_NUM (2 * HNS_ROCE_MAX_PORTS)
46 void hns_roce_qp_event(struct hns_roce_dev *hr_dev, u32 qpn, int event_type)
48 struct device *dev = hr_dev->dev;
49 struct hns_roce_qp *qp;
51 xa_lock(&hr_dev->qp_table_xa);
52 qp = __hns_roce_qp_lookup(hr_dev, qpn);
54 atomic_inc(&qp->refcount);
55 xa_unlock(&hr_dev->qp_table_xa);
58 dev_warn(dev, "Async event for bogus QP %08x\n", qpn);
62 qp->event(qp, (enum hns_roce_event)event_type);
64 if (atomic_dec_and_test(&qp->refcount))
68 static void hns_roce_ib_qp_event(struct hns_roce_qp *hr_qp,
69 enum hns_roce_event type)
71 struct ib_event event;
72 struct ib_qp *ibqp = &hr_qp->ibqp;
74 if (ibqp->event_handler) {
75 event.device = ibqp->device;
76 event.element.qp = ibqp;
78 case HNS_ROCE_EVENT_TYPE_PATH_MIG:
79 event.event = IB_EVENT_PATH_MIG;
81 case HNS_ROCE_EVENT_TYPE_COMM_EST:
82 event.event = IB_EVENT_COMM_EST;
84 case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
85 event.event = IB_EVENT_SQ_DRAINED;
87 case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
88 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
90 case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
91 event.event = IB_EVENT_QP_FATAL;
93 case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
94 event.event = IB_EVENT_PATH_MIG_ERR;
96 case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
97 event.event = IB_EVENT_QP_REQ_ERR;
99 case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
100 event.event = IB_EVENT_QP_ACCESS_ERR;
103 dev_dbg(ibqp->device->dev.parent, "roce_ib: Unexpected event type %d on QP %06lx\n",
107 ibqp->event_handler(&event, ibqp->qp_context);
111 static int hns_roce_reserve_range_qp(struct hns_roce_dev *hr_dev, int cnt,
112 int align, unsigned long *base)
114 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
116 return hns_roce_bitmap_alloc_range(&qp_table->bitmap, cnt, align,
122 enum hns_roce_qp_state to_hns_roce_state(enum ib_qp_state state)
126 return HNS_ROCE_QP_STATE_RST;
128 return HNS_ROCE_QP_STATE_INIT;
130 return HNS_ROCE_QP_STATE_RTR;
132 return HNS_ROCE_QP_STATE_RTS;
134 return HNS_ROCE_QP_STATE_SQD;
136 return HNS_ROCE_QP_STATE_ERR;
138 return HNS_ROCE_QP_NUM_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 xarray *xa = &hr_dev->qp_table_xa;
152 atomic_set(&hr_qp->refcount, 1);
153 init_completion(&hr_qp->free);
155 ret = xa_err(xa_store_irq(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1),
158 dev_err(hr_dev->dev, "QPC xa_store failed\n");
163 static int hns_roce_qp_alloc(struct hns_roce_dev *hr_dev, unsigned long qpn,
164 struct hns_roce_qp *hr_qp)
166 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
167 struct device *dev = hr_dev->dev;
175 /* Alloc memory for QPC */
176 ret = hns_roce_table_get(hr_dev, &qp_table->qp_table, hr_qp->qpn);
178 dev_err(dev, "QPC table get failed\n");
182 /* Alloc memory for IRRL */
183 ret = hns_roce_table_get(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
185 dev_err(dev, "IRRL table get failed\n");
189 if (hr_dev->caps.trrl_entry_sz) {
190 /* Alloc memory for TRRL */
191 ret = hns_roce_table_get(hr_dev, &qp_table->trrl_table,
194 dev_err(dev, "TRRL table get failed\n");
199 if (hr_dev->caps.sccc_entry_sz) {
200 /* Alloc memory for SCC CTX */
201 ret = hns_roce_table_get(hr_dev, &qp_table->sccc_table,
204 dev_err(dev, "SCC CTX table get failed\n");
209 ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
216 if (hr_dev->caps.sccc_entry_sz)
217 hns_roce_table_put(hr_dev, &qp_table->sccc_table,
221 if (hr_dev->caps.trrl_entry_sz)
222 hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn);
225 hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
228 hns_roce_table_put(hr_dev, &qp_table->qp_table, hr_qp->qpn);
234 void hns_roce_qp_remove(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
236 struct xarray *xa = &hr_dev->qp_table_xa;
239 xa_lock_irqsave(xa, flags);
240 __xa_erase(xa, hr_qp->qpn & (hr_dev->caps.num_qps - 1));
241 xa_unlock_irqrestore(xa, flags);
244 void hns_roce_qp_free(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
246 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
248 if (atomic_dec_and_test(&hr_qp->refcount))
249 complete(&hr_qp->free);
250 wait_for_completion(&hr_qp->free);
252 if ((hr_qp->ibqp.qp_type) != IB_QPT_GSI) {
253 if (hr_dev->caps.trrl_entry_sz)
254 hns_roce_table_put(hr_dev, &qp_table->trrl_table,
256 hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn);
260 void hns_roce_release_range_qp(struct hns_roce_dev *hr_dev, int base_qpn,
263 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
265 if (base_qpn < hr_dev->caps.reserved_qps)
268 hns_roce_bitmap_free_range(&qp_table->bitmap, base_qpn, cnt, BITMAP_RR);
271 static int hns_roce_set_rq_size(struct hns_roce_dev *hr_dev,
272 struct ib_qp_cap *cap, bool is_user, int has_rq,
273 struct hns_roce_qp *hr_qp)
275 struct device *dev = hr_dev->dev;
278 /* Check the validity of QP support capacity */
279 if (cap->max_recv_wr > hr_dev->caps.max_wqes ||
280 cap->max_recv_sge > hr_dev->caps.max_rq_sg) {
281 dev_err(dev, "RQ WR or sge error!max_recv_wr=%d max_recv_sge=%d\n",
282 cap->max_recv_wr, cap->max_recv_sge);
286 /* If srq exist, set zero for relative number of rq */
288 hr_qp->rq.wqe_cnt = 0;
289 hr_qp->rq.max_gs = 0;
290 cap->max_recv_wr = 0;
291 cap->max_recv_sge = 0;
293 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) {
294 dev_err(dev, "user space no need config max_recv_wr max_recv_sge\n");
298 if (hr_dev->caps.min_wqes)
299 max_cnt = max(cap->max_recv_wr, hr_dev->caps.min_wqes);
301 max_cnt = cap->max_recv_wr;
303 hr_qp->rq.wqe_cnt = roundup_pow_of_two(max_cnt);
305 if ((u32)hr_qp->rq.wqe_cnt > hr_dev->caps.max_wqes) {
306 dev_err(dev, "while setting rq size, rq.wqe_cnt too large\n");
310 max_cnt = max(1U, cap->max_recv_sge);
311 hr_qp->rq.max_gs = roundup_pow_of_two(max_cnt);
312 if (hr_dev->caps.max_rq_sg <= 2)
313 hr_qp->rq.wqe_shift =
314 ilog2(hr_dev->caps.max_rq_desc_sz);
316 hr_qp->rq.wqe_shift =
317 ilog2(hr_dev->caps.max_rq_desc_sz
321 cap->max_recv_wr = hr_qp->rq.max_post = hr_qp->rq.wqe_cnt;
322 cap->max_recv_sge = hr_qp->rq.max_gs;
327 static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
328 struct ib_qp_cap *cap,
329 struct hns_roce_ib_create_qp *ucmd)
331 u32 roundup_sq_stride = roundup_pow_of_two(hr_dev->caps.max_sq_desc_sz);
332 u8 max_sq_stride = ilog2(roundup_sq_stride);
334 /* Sanity check SQ size before proceeding */
335 if ((u32)(1 << ucmd->log_sq_bb_count) > hr_dev->caps.max_wqes ||
336 ucmd->log_sq_stride > max_sq_stride ||
337 ucmd->log_sq_stride < HNS_ROCE_IB_MIN_SQ_STRIDE) {
338 ibdev_err(&hr_dev->ib_dev, "check SQ size error!\n");
342 if (cap->max_send_sge > hr_dev->caps.max_sq_sg) {
343 ibdev_err(&hr_dev->ib_dev, "SQ sge error! max_send_sge=%d\n",
351 static int hns_roce_set_user_sq_size(struct hns_roce_dev *hr_dev,
352 struct ib_qp_cap *cap,
353 struct hns_roce_qp *hr_qp,
354 struct hns_roce_ib_create_qp *ucmd)
361 ret = check_sq_size_with_integrity(hr_dev, cap, ucmd);
363 ibdev_err(&hr_dev->ib_dev, "Sanity check sq size failed\n");
367 hr_qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
368 hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
370 max_cnt = max(1U, cap->max_send_sge);
371 if (hr_dev->caps.max_sq_sg <= 2)
372 hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
374 hr_qp->sq.max_gs = max_cnt;
376 if (hr_qp->sq.max_gs > 2)
377 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
378 (hr_qp->sq.max_gs - 2));
380 if ((hr_qp->sq.max_gs > 2) && (hr_dev->pci_dev->revision == 0x20)) {
381 if (hr_qp->sge.sge_cnt > hr_dev->caps.max_extend_sg) {
383 "The extended sge cnt error! sge_cnt=%d\n",
389 hr_qp->sge.sge_shift = 4;
390 ex_sge_num = hr_qp->sge.sge_cnt;
392 /* Get buf size, SQ and RQ are aligned to page_szie */
393 if (hr_dev->caps.max_sq_sg <= 2) {
394 hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
395 hr_qp->rq.wqe_shift), PAGE_SIZE) +
396 HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
397 hr_qp->sq.wqe_shift), PAGE_SIZE);
399 hr_qp->sq.offset = 0;
400 hr_qp->rq.offset = HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
401 hr_qp->sq.wqe_shift), PAGE_SIZE);
403 page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
404 hr_qp->sge.sge_cnt = ex_sge_num ?
405 max(page_size / (1 << hr_qp->sge.sge_shift), ex_sge_num) : 0;
406 hr_qp->buff_size = HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt <<
407 hr_qp->rq.wqe_shift), page_size) +
408 HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
409 hr_qp->sge.sge_shift), page_size) +
410 HNS_ROCE_ALOGN_UP((hr_qp->sq.wqe_cnt <<
411 hr_qp->sq.wqe_shift), page_size);
413 hr_qp->sq.offset = 0;
415 hr_qp->sge.offset = HNS_ROCE_ALOGN_UP(
416 (hr_qp->sq.wqe_cnt <<
417 hr_qp->sq.wqe_shift),
419 hr_qp->rq.offset = hr_qp->sge.offset +
420 HNS_ROCE_ALOGN_UP((hr_qp->sge.sge_cnt <<
421 hr_qp->sge.sge_shift),
424 hr_qp->rq.offset = HNS_ROCE_ALOGN_UP(
425 (hr_qp->sq.wqe_cnt <<
426 hr_qp->sq.wqe_shift),
434 static int split_wqe_buf_region(struct hns_roce_dev *hr_dev,
435 struct hns_roce_qp *hr_qp,
436 struct hns_roce_buf_region *regions,
437 int region_max, int page_shift)
439 int page_size = 1 << page_shift;
445 if (hr_qp->buff_size < 1 || region_max < 1)
448 if (hr_qp->sge.sge_cnt > 0)
449 is_extend_sge = true;
451 is_extend_sge = false;
455 buf_size = hr_qp->sge.offset - hr_qp->sq.offset;
457 buf_size = hr_qp->rq.offset - hr_qp->sq.offset;
459 if (buf_size > 0 && region_cnt < region_max) {
460 buf_cnt = DIV_ROUND_UP(buf_size, page_size);
461 hns_roce_init_buf_region(®ions[region_cnt],
462 hr_dev->caps.wqe_sq_hop_num,
463 hr_qp->sq.offset / page_size,
470 buf_size = hr_qp->rq.offset - hr_qp->sge.offset;
471 if (buf_size > 0 && region_cnt < region_max) {
472 buf_cnt = DIV_ROUND_UP(buf_size, page_size);
473 hns_roce_init_buf_region(®ions[region_cnt],
474 hr_dev->caps.wqe_sge_hop_num,
475 hr_qp->sge.offset / page_size,
482 buf_size = hr_qp->buff_size - hr_qp->rq.offset;
484 buf_cnt = DIV_ROUND_UP(buf_size, page_size);
485 hns_roce_init_buf_region(®ions[region_cnt],
486 hr_dev->caps.wqe_rq_hop_num,
487 hr_qp->rq.offset / page_size,
495 static int calc_wqe_bt_page_shift(struct hns_roce_dev *hr_dev,
496 struct hns_roce_buf_region *regions,
503 bt_pg_shift = PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz;
505 /* all root ba entries must in one bt page */
507 ba_num = (1 << bt_pg_shift) / BA_BYTE_LEN;
508 ret = hns_roce_hem_list_calc_root_ba(regions, region_cnt,
514 } while (ret > ba_num);
516 return bt_pg_shift - PAGE_SHIFT;
519 static int set_extend_sge_param(struct hns_roce_dev *hr_dev,
520 struct hns_roce_qp *hr_qp)
522 struct device *dev = hr_dev->dev;
524 if (hr_qp->sq.max_gs > 2) {
525 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
526 (hr_qp->sq.max_gs - 2));
527 hr_qp->sge.sge_shift = 4;
530 /* ud sqwqe's sge use extend sge */
531 if (hr_dev->caps.max_sq_sg > 2 && hr_qp->ibqp.qp_type == IB_QPT_GSI) {
532 hr_qp->sge.sge_cnt = roundup_pow_of_two(hr_qp->sq.wqe_cnt *
534 hr_qp->sge.sge_shift = 4;
537 if ((hr_qp->sq.max_gs > 2) && hr_dev->pci_dev->revision == 0x20) {
538 if (hr_qp->sge.sge_cnt > hr_dev->caps.max_extend_sg) {
539 dev_err(dev, "The extended sge cnt error! sge_cnt=%d\n",
548 static int hns_roce_set_kernel_sq_size(struct hns_roce_dev *hr_dev,
549 struct ib_qp_cap *cap,
550 struct hns_roce_qp *hr_qp)
552 struct device *dev = hr_dev->dev;
558 if (cap->max_send_wr > hr_dev->caps.max_wqes ||
559 cap->max_send_sge > hr_dev->caps.max_sq_sg ||
560 cap->max_inline_data > hr_dev->caps.max_sq_inline) {
561 dev_err(dev, "SQ WR or sge or inline data error!\n");
565 hr_qp->sq.wqe_shift = ilog2(hr_dev->caps.max_sq_desc_sz);
567 if (hr_dev->caps.min_wqes)
568 max_cnt = max(cap->max_send_wr, hr_dev->caps.min_wqes);
570 max_cnt = cap->max_send_wr;
572 hr_qp->sq.wqe_cnt = roundup_pow_of_two(max_cnt);
573 if ((u32)hr_qp->sq.wqe_cnt > hr_dev->caps.max_wqes) {
574 dev_err(dev, "while setting kernel sq size, sq.wqe_cnt too large\n");
578 /* Get data_seg numbers */
579 max_cnt = max(1U, cap->max_send_sge);
580 if (hr_dev->caps.max_sq_sg <= 2)
581 hr_qp->sq.max_gs = roundup_pow_of_two(max_cnt);
583 hr_qp->sq.max_gs = max_cnt;
585 ret = set_extend_sge_param(hr_dev, hr_qp);
587 dev_err(dev, "set extend sge parameters fail\n");
591 /* Get buf size, SQ and RQ are aligned to PAGE_SIZE */
592 page_size = 1 << (hr_dev->caps.mtt_buf_pg_sz + PAGE_SHIFT);
593 hr_qp->sq.offset = 0;
594 size = HNS_ROCE_ALOGN_UP(hr_qp->sq.wqe_cnt << hr_qp->sq.wqe_shift,
597 if (hr_dev->caps.max_sq_sg > 2 && hr_qp->sge.sge_cnt) {
598 hr_qp->sge.sge_cnt = max(page_size/(1 << hr_qp->sge.sge_shift),
599 (u32)hr_qp->sge.sge_cnt);
600 hr_qp->sge.offset = size;
601 size += HNS_ROCE_ALOGN_UP(hr_qp->sge.sge_cnt <<
602 hr_qp->sge.sge_shift, page_size);
605 hr_qp->rq.offset = size;
606 size += HNS_ROCE_ALOGN_UP((hr_qp->rq.wqe_cnt << hr_qp->rq.wqe_shift),
608 hr_qp->buff_size = size;
610 /* Get wr and sge number which send */
611 cap->max_send_wr = hr_qp->sq.max_post = hr_qp->sq.wqe_cnt;
612 cap->max_send_sge = hr_qp->sq.max_gs;
614 /* We don't support inline sends for kernel QPs (yet) */
615 cap->max_inline_data = 0;
620 static int hns_roce_qp_has_sq(struct ib_qp_init_attr *attr)
622 if (attr->qp_type == IB_QPT_XRC_TGT || !attr->cap.max_send_wr)
628 static int hns_roce_qp_has_rq(struct ib_qp_init_attr *attr)
630 if (attr->qp_type == IB_QPT_XRC_INI ||
631 attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
632 !attr->cap.max_recv_wr)
638 static int alloc_rq_inline_buf(struct hns_roce_qp *hr_qp,
639 struct ib_qp_init_attr *init_attr)
641 u32 max_recv_sge = init_attr->cap.max_recv_sge;
642 struct hns_roce_rinl_wqe *wqe_list;
643 u32 wqe_cnt = hr_qp->rq.wqe_cnt;
646 /* allocate recv inline buf */
647 wqe_list = kcalloc(wqe_cnt, sizeof(struct hns_roce_rinl_wqe),
653 /* Allocate a continuous buffer for all inline sge we need */
654 wqe_list[0].sg_list = kcalloc(wqe_cnt, (max_recv_sge *
655 sizeof(struct hns_roce_rinl_sge)),
657 if (!wqe_list[0].sg_list)
660 /* Assign buffers of sg_list to each inline wqe */
661 for (i = 1; i < wqe_cnt; i++)
662 wqe_list[i].sg_list = &wqe_list[0].sg_list[i * max_recv_sge];
664 hr_qp->rq_inl_buf.wqe_list = wqe_list;
665 hr_qp->rq_inl_buf.wqe_cnt = wqe_cnt;
676 static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
678 kfree(hr_qp->rq_inl_buf.wqe_list[0].sg_list);
679 kfree(hr_qp->rq_inl_buf.wqe_list);
682 static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
684 struct ib_qp_init_attr *init_attr,
685 struct ib_udata *udata, unsigned long sqpn,
686 struct hns_roce_qp *hr_qp)
688 dma_addr_t *buf_list[ARRAY_SIZE(hr_qp->regions)] = { NULL };
689 struct device *dev = hr_dev->dev;
690 struct hns_roce_ib_create_qp ucmd;
691 struct hns_roce_ib_create_qp_resp resp = {};
692 struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(
693 udata, struct hns_roce_ucontext, ibucontext);
694 struct hns_roce_buf_region *r;
695 unsigned long qpn = 0;
701 mutex_init(&hr_qp->mutex);
702 spin_lock_init(&hr_qp->sq.lock);
703 spin_lock_init(&hr_qp->rq.lock);
705 hr_qp->state = IB_QPS_RESET;
707 hr_qp->ibqp.qp_type = init_attr->qp_type;
709 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
710 hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
712 hr_qp->sq_signal_bits = IB_SIGNAL_REQ_WR;
714 ret = hns_roce_set_rq_size(hr_dev, &init_attr->cap, udata,
715 hns_roce_qp_has_rq(init_attr), hr_qp);
717 dev_err(dev, "hns_roce_set_rq_size failed\n");
721 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
722 hns_roce_qp_has_rq(init_attr)) {
723 ret = alloc_rq_inline_buf(hr_qp, init_attr);
725 dev_err(dev, "allocate receive inline buffer failed\n");
730 page_shift = PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
732 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
733 dev_err(dev, "ib_copy_from_udata error for create qp\n");
735 goto err_alloc_rq_inline_buf;
738 ret = hns_roce_set_user_sq_size(hr_dev, &init_attr->cap, hr_qp,
741 dev_err(dev, "hns_roce_set_user_sq_size error for create qp\n");
742 goto err_alloc_rq_inline_buf;
745 hr_qp->umem = ib_umem_get(udata, ucmd.buf_addr,
746 hr_qp->buff_size, 0, 0);
747 if (IS_ERR(hr_qp->umem)) {
748 dev_err(dev, "ib_umem_get error for create qp\n");
749 ret = PTR_ERR(hr_qp->umem);
750 goto err_alloc_rq_inline_buf;
752 hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
753 hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
755 ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
758 dev_err(dev, "alloc buf_list error for create qp\n");
762 for (i = 0; i < hr_qp->region_cnt; i++) {
763 r = &hr_qp->regions[i];
764 buf_count = hns_roce_get_umem_bufs(hr_dev,
765 buf_list[i], r->count, r->offset,
766 hr_qp->umem, page_shift);
767 if (buf_count != r->count) {
769 "get umem buf err, expect %d,ret %d.\n",
770 r->count, buf_count);
776 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
777 (udata->inlen >= sizeof(ucmd)) &&
778 (udata->outlen >= sizeof(resp)) &&
779 hns_roce_qp_has_sq(init_attr)) {
780 ret = hns_roce_db_map_user(uctx, udata, ucmd.sdb_addr,
783 dev_err(dev, "sq record doorbell map failed!\n");
787 /* indicate kernel supports sq record db */
788 resp.cap_flags |= HNS_ROCE_SUPPORT_SQ_RECORD_DB;
792 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
793 (udata->outlen >= sizeof(resp)) &&
794 hns_roce_qp_has_rq(init_attr)) {
795 ret = hns_roce_db_map_user(uctx, udata, ucmd.db_addr,
798 dev_err(dev, "rq record doorbell map failed!\n");
802 /* indicate kernel supports rq record db */
803 resp.cap_flags |= HNS_ROCE_SUPPORT_RQ_RECORD_DB;
807 if (init_attr->create_flags &
808 IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
809 dev_err(dev, "init_attr->create_flags error!\n");
811 goto err_alloc_rq_inline_buf;
814 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) {
815 dev_err(dev, "init_attr->create_flags error!\n");
817 goto err_alloc_rq_inline_buf;
821 ret = hns_roce_set_kernel_sq_size(hr_dev, &init_attr->cap,
824 dev_err(dev, "hns_roce_set_kernel_sq_size error!\n");
825 goto err_alloc_rq_inline_buf;
828 /* QP doorbell register address */
829 hr_qp->sq.db_reg_l = hr_dev->reg_base + hr_dev->sdb_offset +
830 DB_REG_OFFSET * hr_dev->priv_uar.index;
831 hr_qp->rq.db_reg_l = hr_dev->reg_base + hr_dev->odb_offset +
832 DB_REG_OFFSET * hr_dev->priv_uar.index;
834 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
835 hns_roce_qp_has_rq(init_attr)) {
836 ret = hns_roce_alloc_db(hr_dev, &hr_qp->rdb, 0);
838 dev_err(dev, "rq record doorbell alloc failed!\n");
839 goto err_alloc_rq_inline_buf;
841 *hr_qp->rdb.db_record = 0;
845 /* Allocate QP buf */
846 if (hns_roce_buf_alloc(hr_dev, hr_qp->buff_size,
847 (1 << page_shift) * 2,
848 &hr_qp->hr_buf, page_shift)) {
849 dev_err(dev, "hns_roce_buf_alloc error!\n");
853 hr_qp->region_cnt = split_wqe_buf_region(hr_dev, hr_qp,
854 hr_qp->regions, ARRAY_SIZE(hr_qp->regions),
856 ret = hns_roce_alloc_buf_list(hr_qp->regions, buf_list,
859 dev_err(dev, "alloc buf_list error for create qp!\n");
863 for (i = 0; i < hr_qp->region_cnt; i++) {
864 r = &hr_qp->regions[i];
865 buf_count = hns_roce_get_kmem_bufs(hr_dev,
866 buf_list[i], r->count, r->offset,
868 if (buf_count != r->count) {
870 "get kmem buf err, expect %d,ret %d.\n",
871 r->count, buf_count);
877 hr_qp->sq.wrid = kcalloc(hr_qp->sq.wqe_cnt, sizeof(u64),
879 if (ZERO_OR_NULL_PTR(hr_qp->sq.wrid)) {
884 if (hr_qp->rq.wqe_cnt) {
885 hr_qp->rq.wrid = kcalloc(hr_qp->rq.wqe_cnt, sizeof(u64),
887 if (ZERO_OR_NULL_PTR(hr_qp->rq.wrid)) {
898 ret = hns_roce_reserve_range_qp(hr_dev, 1, 1, &qpn);
900 dev_err(dev, "hns_roce_reserve_range_qp alloc qpn error\n");
905 hr_qp->wqe_bt_pg_shift = calc_wqe_bt_page_shift(hr_dev, hr_qp->regions,
907 hns_roce_mtr_init(&hr_qp->mtr, PAGE_SHIFT + hr_qp->wqe_bt_pg_shift,
909 ret = hns_roce_mtr_attach(hr_dev, &hr_qp->mtr, buf_list,
910 hr_qp->regions, hr_qp->region_cnt);
912 dev_err(dev, "mtr attach error for create qp\n");
916 if (init_attr->qp_type == IB_QPT_GSI &&
917 hr_dev->hw_rev == HNS_ROCE_HW_VER1) {
918 /* In v1 engine, GSI QP context in RoCE engine's register */
919 ret = hns_roce_gsi_qp_alloc(hr_dev, qpn, hr_qp);
921 dev_err(dev, "hns_roce_qp_alloc failed!\n");
925 ret = hns_roce_qp_alloc(hr_dev, qpn, hr_qp);
927 dev_err(dev, "hns_roce_qp_alloc failed!\n");
933 hr_qp->doorbell_qpn = 1;
935 hr_qp->doorbell_qpn = (u32)hr_qp->qpn;
938 ret = ib_copy_to_udata(udata, &resp,
939 min(udata->outlen, sizeof(resp)));
944 if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL) {
945 ret = hr_dev->hw->qp_flow_control_init(hr_dev, hr_qp);
950 hr_qp->event = hns_roce_ib_qp_event;
951 hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
956 if (init_attr->qp_type == IB_QPT_GSI &&
957 hr_dev->hw_rev == HNS_ROCE_HW_VER1)
958 hns_roce_qp_remove(hr_dev, hr_qp);
960 hns_roce_qp_free(hr_dev, hr_qp);
964 hns_roce_release_range_qp(hr_dev, qpn, 1);
967 hns_roce_mtr_cleanup(hr_dev, &hr_qp->mtr);
971 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB) &&
972 (udata->outlen >= sizeof(resp)) &&
973 hns_roce_qp_has_rq(init_attr))
974 hns_roce_db_unmap_user(uctx, &hr_qp->rdb);
976 if (hr_qp->rq.wqe_cnt)
977 kfree(hr_qp->rq.wrid);
982 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SQ_RECORD_DB) &&
983 (udata->inlen >= sizeof(ucmd)) &&
984 (udata->outlen >= sizeof(resp)) &&
985 hns_roce_qp_has_sq(init_attr))
986 hns_roce_db_unmap_user(uctx, &hr_qp->sdb);
990 kfree(hr_qp->sq.wrid);
993 hns_roce_free_buf_list(buf_list, hr_qp->region_cnt);
997 hns_roce_buf_free(hr_dev, hr_qp->buff_size, &hr_qp->hr_buf);
998 ib_umem_release(hr_qp->umem);
1001 if (!udata && hns_roce_qp_has_rq(init_attr) &&
1002 (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RECORD_DB))
1003 hns_roce_free_db(hr_dev, &hr_qp->rdb);
1005 err_alloc_rq_inline_buf:
1006 if ((hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE) &&
1007 hns_roce_qp_has_rq(init_attr))
1008 free_rq_inline_buf(hr_qp);
1014 struct ib_qp *hns_roce_create_qp(struct ib_pd *pd,
1015 struct ib_qp_init_attr *init_attr,
1016 struct ib_udata *udata)
1018 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
1019 struct ib_device *ibdev = &hr_dev->ib_dev;
1020 struct hns_roce_sqp *hr_sqp;
1021 struct hns_roce_qp *hr_qp;
1024 switch (init_attr->qp_type) {
1026 hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
1028 return ERR_PTR(-ENOMEM);
1030 ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata, 0,
1033 ibdev_err(ibdev, "Create RC QP 0x%06lx failed(%d)\n",
1036 return ERR_PTR(ret);
1039 hr_qp->ibqp.qp_num = hr_qp->qpn;
1044 /* Userspace is not allowed to create special QPs: */
1046 ibdev_err(ibdev, "not support usr space GSI\n");
1047 return ERR_PTR(-EINVAL);
1050 hr_sqp = kzalloc(sizeof(*hr_sqp), GFP_KERNEL);
1052 return ERR_PTR(-ENOMEM);
1054 hr_qp = &hr_sqp->hr_qp;
1055 hr_qp->port = init_attr->port_num - 1;
1056 hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
1058 /* when hw version is v1, the sqpn is allocated */
1059 if (hr_dev->caps.max_sq_sg <= 2)
1060 hr_qp->ibqp.qp_num = HNS_ROCE_MAX_PORTS +
1061 hr_dev->iboe.phy_port[hr_qp->port];
1063 hr_qp->ibqp.qp_num = 1;
1065 ret = hns_roce_create_qp_common(hr_dev, pd, init_attr, udata,
1066 hr_qp->ibqp.qp_num, hr_qp);
1068 ibdev_err(ibdev, "Create GSI QP failed!\n");
1070 return ERR_PTR(ret);
1076 ibdev_err(ibdev, "not support QP type %d\n",
1077 init_attr->qp_type);
1078 return ERR_PTR(-EINVAL);
1082 return &hr_qp->ibqp;
1085 int to_hr_qp_type(int qp_type)
1089 if (qp_type == IB_QPT_RC)
1090 transport_type = SERV_TYPE_RC;
1091 else if (qp_type == IB_QPT_UC)
1092 transport_type = SERV_TYPE_UC;
1093 else if (qp_type == IB_QPT_UD)
1094 transport_type = SERV_TYPE_UD;
1095 else if (qp_type == IB_QPT_GSI)
1096 transport_type = SERV_TYPE_UD;
1098 transport_type = -1;
1100 return transport_type;
1103 static int check_mtu_validate(struct hns_roce_dev *hr_dev,
1104 struct hns_roce_qp *hr_qp,
1105 struct ib_qp_attr *attr, int attr_mask)
1107 enum ib_mtu active_mtu;
1110 p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1111 active_mtu = iboe_get_mtu(hr_dev->iboe.netdevs[p]->mtu);
1113 if ((hr_dev->caps.max_mtu >= IB_MTU_2048 &&
1114 attr->path_mtu > hr_dev->caps.max_mtu) ||
1115 attr->path_mtu < IB_MTU_256 || attr->path_mtu > active_mtu) {
1116 ibdev_err(&hr_dev->ib_dev,
1117 "attr path_mtu(%d)invalid while modify qp",
1125 static int hns_roce_check_qp_attr(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1128 struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1129 struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1132 if ((attr_mask & IB_QP_PORT) &&
1133 (attr->port_num == 0 || attr->port_num > hr_dev->caps.num_ports)) {
1134 ibdev_err(&hr_dev->ib_dev,
1135 "attr port_num invalid.attr->port_num=%d\n",
1140 if (attr_mask & IB_QP_PKEY_INDEX) {
1141 p = attr_mask & IB_QP_PORT ? (attr->port_num - 1) : hr_qp->port;
1142 if (attr->pkey_index >= hr_dev->caps.pkey_table_len[p]) {
1143 ibdev_err(&hr_dev->ib_dev,
1144 "attr pkey_index invalid.attr->pkey_index=%d\n",
1150 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1151 attr->max_rd_atomic > hr_dev->caps.max_qp_init_rdma) {
1152 ibdev_err(&hr_dev->ib_dev,
1153 "attr max_rd_atomic invalid.attr->max_rd_atomic=%d\n",
1154 attr->max_rd_atomic);
1158 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1159 attr->max_dest_rd_atomic > hr_dev->caps.max_qp_dest_rdma) {
1160 ibdev_err(&hr_dev->ib_dev,
1161 "attr max_dest_rd_atomic invalid.attr->max_dest_rd_atomic=%d\n",
1162 attr->max_dest_rd_atomic);
1166 if (attr_mask & IB_QP_PATH_MTU)
1167 return check_mtu_validate(hr_dev, hr_qp, attr, attr_mask);
1172 int hns_roce_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1173 int attr_mask, struct ib_udata *udata)
1175 struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
1176 struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
1177 enum ib_qp_state cur_state, new_state;
1180 mutex_lock(&hr_qp->mutex);
1182 cur_state = attr_mask & IB_QP_CUR_STATE ?
1183 attr->cur_qp_state : (enum ib_qp_state)hr_qp->state;
1184 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1186 if (ibqp->uobject &&
1187 (attr_mask & IB_QP_STATE) && new_state == IB_QPS_ERR) {
1188 if (hr_qp->sdb_en == 1) {
1189 hr_qp->sq.head = *(int *)(hr_qp->sdb.virt_addr);
1191 if (hr_qp->rdb_en == 1)
1192 hr_qp->rq.head = *(int *)(hr_qp->rdb.virt_addr);
1194 ibdev_warn(&hr_dev->ib_dev,
1195 "flush cqe is not supported in userspace!\n");
1200 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1202 ibdev_err(&hr_dev->ib_dev, "ib_modify_qp_is_ok failed\n");
1206 ret = hns_roce_check_qp_attr(ibqp, attr, attr_mask);
1210 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1211 if (hr_dev->caps.min_wqes) {
1213 ibdev_err(&hr_dev->ib_dev,
1214 "cur_state=%d new_state=%d\n", cur_state,
1223 ret = hr_dev->hw->modify_qp(ibqp, attr, attr_mask, cur_state,
1227 mutex_unlock(&hr_qp->mutex);
1232 void hns_roce_lock_cqs(struct hns_roce_cq *send_cq, struct hns_roce_cq *recv_cq)
1233 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1235 if (send_cq == recv_cq) {
1236 spin_lock_irq(&send_cq->lock);
1237 __acquire(&recv_cq->lock);
1238 } else if (send_cq->cqn < recv_cq->cqn) {
1239 spin_lock_irq(&send_cq->lock);
1240 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1242 spin_lock_irq(&recv_cq->lock);
1243 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1247 void hns_roce_unlock_cqs(struct hns_roce_cq *send_cq,
1248 struct hns_roce_cq *recv_cq) __releases(&send_cq->lock)
1249 __releases(&recv_cq->lock)
1251 if (send_cq == recv_cq) {
1252 __release(&recv_cq->lock);
1253 spin_unlock_irq(&send_cq->lock);
1254 } else if (send_cq->cqn < recv_cq->cqn) {
1255 spin_unlock(&recv_cq->lock);
1256 spin_unlock_irq(&send_cq->lock);
1258 spin_unlock(&send_cq->lock);
1259 spin_unlock_irq(&recv_cq->lock);
1263 static void *get_wqe(struct hns_roce_qp *hr_qp, int offset)
1266 return hns_roce_buf_offset(&hr_qp->hr_buf, offset);
1269 void *get_recv_wqe(struct hns_roce_qp *hr_qp, int n)
1271 return get_wqe(hr_qp, hr_qp->rq.offset + (n << hr_qp->rq.wqe_shift));
1274 void *get_send_wqe(struct hns_roce_qp *hr_qp, int n)
1276 return get_wqe(hr_qp, hr_qp->sq.offset + (n << hr_qp->sq.wqe_shift));
1279 void *get_send_extend_sge(struct hns_roce_qp *hr_qp, int n)
1281 return hns_roce_buf_offset(&hr_qp->hr_buf, hr_qp->sge.offset +
1282 (n << hr_qp->sge.sge_shift));
1285 bool hns_roce_wq_overflow(struct hns_roce_wq *hr_wq, int nreq,
1286 struct ib_cq *ib_cq)
1288 struct hns_roce_cq *hr_cq;
1291 cur = hr_wq->head - hr_wq->tail;
1292 if (likely(cur + nreq < hr_wq->max_post))
1295 hr_cq = to_hr_cq(ib_cq);
1296 spin_lock(&hr_cq->lock);
1297 cur = hr_wq->head - hr_wq->tail;
1298 spin_unlock(&hr_cq->lock);
1300 return cur + nreq >= hr_wq->max_post;
1303 int hns_roce_init_qp_table(struct hns_roce_dev *hr_dev)
1305 struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;
1306 int reserved_from_top = 0;
1307 int reserved_from_bot;
1310 mutex_init(&qp_table->scc_mutex);
1311 xa_init(&hr_dev->qp_table_xa);
1313 reserved_from_bot = hr_dev->caps.reserved_qps;
1315 ret = hns_roce_bitmap_init(&qp_table->bitmap, hr_dev->caps.num_qps,
1316 hr_dev->caps.num_qps - 1, reserved_from_bot,
1319 dev_err(hr_dev->dev, "qp bitmap init failed!error=%d\n",
1327 void hns_roce_cleanup_qp_table(struct hns_roce_dev *hr_dev)
1329 hns_roce_bitmap_cleanup(&hr_dev->qp_table.bitmap);