[mirror_ubuntu-eoan-kernel.git] / drivers / infiniband / hw / hns / hns_roce_srq.c

// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
/*
 * Copyright (c) 2018 Hisilicon Limited.
 */

#include <rdma/ib_umem.h>
#include <rdma/hns-abi.h>
#include "hns_roce_device.h"
#include "hns_roce_cmd.h"
#include "hns_roce_hem.h"

void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)
{
	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	struct hns_roce_srq *srq;

	xa_lock(&srq_table->xa);
	srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
	if (srq)
		atomic_inc(&srq->refcount);
	xa_unlock(&srq_table->xa);

	if (!srq) {
		dev_warn(hr_dev->dev, "Async event for bogus SRQ %08x\n", srqn);
		return;
	}

	srq->event(srq, event_type);

	if (atomic_dec_and_test(&srq->refcount))
		complete(&srq->free);
}
EXPORT_SYMBOL_GPL(hns_roce_srq_event);

static void hns_roce_ib_srq_event(struct hns_roce_srq *srq,
				  enum hns_roce_event event_type)
{
	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
	struct ib_srq *ibsrq = &srq->ibsrq;
	struct ib_event event;

	if (ibsrq->event_handler) {
		event.device      = ibsrq->device;
		event.element.srq = ibsrq;
		switch (event_type) {
		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
			break;
		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
			event.event = IB_EVENT_SRQ_ERR;
			break;
		default:
			dev_err(hr_dev->dev,
			   "hns_roce:Unexpected event type 0x%x on SRQ %06lx\n",
			   event_type, srq->srqn);
			return;
		}

		ibsrq->event_handler(&event, ibsrq->srq_context);
	}
}

static int hns_roce_sw2hw_srq(struct hns_roce_dev *dev,
			      struct hns_roce_cmd_mailbox *mailbox,
			      unsigned long srq_num)
{
	return hns_roce_cmd_mbox(dev, mailbox->dma, 0, srq_num, 0,
				 HNS_ROCE_CMD_SW2HW_SRQ,
				 HNS_ROCE_CMD_TIMEOUT_MSECS);
}

static int hns_roce_hw2sw_srq(struct hns_roce_dev *dev,
			     struct hns_roce_cmd_mailbox *mailbox,
			     unsigned long srq_num)
{
	return hns_roce_cmd_mbox(dev, 0, mailbox ? mailbox->dma : 0, srq_num,
				 mailbox ? 0 : 1, HNS_ROCE_CMD_HW2SW_SRQ,
				 HNS_ROCE_CMD_TIMEOUT_MSECS);
}

int hns_roce_srq_alloc(struct hns_roce_dev *hr_dev, u32 pdn, u32 cqn, u16 xrcd,
		       struct hns_roce_mtt *hr_mtt, u64 db_rec_addr,
		       struct hns_roce_srq *srq)
{
	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	struct hns_roce_cmd_mailbox *mailbox;
	dma_addr_t dma_handle_wqe;
	dma_addr_t dma_handle_idx;
	u64 *mtts_wqe;
	u64 *mtts_idx;
	int ret;

	/* Get the physical address of srq buf */
	mtts_wqe = hns_roce_table_find(hr_dev,
				       &hr_dev->mr_table.mtt_srqwqe_table,
				       srq->mtt.first_seg,
				       &dma_handle_wqe);
	if (!mtts_wqe) {
		dev_err(hr_dev->dev,
			"SRQ alloc.Failed to find srq buf addr.\n");
		return -EINVAL;
	}

	/* Get physical address of idx que buf */
	mtts_idx = hns_roce_table_find(hr_dev, &hr_dev->mr_table.mtt_idx_table,
				       srq->idx_que.mtt.first_seg,
				       &dma_handle_idx);
	if (!mtts_idx) {
		dev_err(hr_dev->dev,
			"SRQ alloc.Failed to find idx que buf addr.\n");
		return -EINVAL;
	}

	ret = hns_roce_bitmap_alloc(&srq_table->bitmap, &srq->srqn);
	if (ret == -1) {
		dev_err(hr_dev->dev, "SRQ alloc.Failed to alloc index.\n");
		return -ENOMEM;
	}

	ret = hns_roce_table_get(hr_dev, &srq_table->table, srq->srqn);
	if (ret)
		goto err_out;

	ret = xa_err(xa_store(&srq_table->xa, srq->srqn, srq, GFP_KERNEL));
	if (ret)
		goto err_put;

	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
	if (IS_ERR(mailbox)) {
		ret = PTR_ERR(mailbox);
		goto err_xa;
	}

	hr_dev->hw->write_srqc(hr_dev, srq, pdn, xrcd, cqn, mailbox->buf,
			       mtts_wqe, mtts_idx, dma_handle_wqe,
			       dma_handle_idx);

	ret = hns_roce_sw2hw_srq(hr_dev, mailbox, srq->srqn);
	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
	if (ret)
		goto err_xa;

	atomic_set(&srq->refcount, 1);
	init_completion(&srq->free);
	return ret;

err_xa:
	xa_erase(&srq_table->xa, srq->srqn);

err_put:
	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);

err_out:
	hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
	return ret;
}

void hns_roce_srq_free(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
{
	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	int ret;

	ret = hns_roce_hw2sw_srq(hr_dev, NULL, srq->srqn);
	if (ret)
		dev_err(hr_dev->dev, "HW2SW_SRQ failed (%d) for CQN %06lx\n",
			ret, srq->srqn);

	xa_erase(&srq_table->xa, srq->srqn);

	if (atomic_dec_and_test(&srq->refcount))
		complete(&srq->free);
	wait_for_completion(&srq->free);

	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
	hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
}

static int hns_roce_create_idx_que(struct ib_pd *pd, struct hns_roce_srq *srq,
				   u32 page_shift)
{
	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
	struct hns_roce_idx_que *idx_que = &srq->idx_que;
	u32 bitmap_num;
	int i;

	bitmap_num = HNS_ROCE_ALOGN_UP(srq->max, 8 * sizeof(u64));

	idx_que->bitmap = kcalloc(1, bitmap_num / 8, GFP_KERNEL);
	if (!idx_que->bitmap)
		return -ENOMEM;

	bitmap_num = bitmap_num / (8 * sizeof(u64));

	idx_que->buf_size = srq->idx_que.buf_size;

	if (hns_roce_buf_alloc(hr_dev, idx_que->buf_size, (1 << page_shift) * 2,
			       &idx_que->idx_buf, page_shift)) {
		kfree(idx_que->bitmap);
		return -ENOMEM;
	}

	for (i = 0; i < bitmap_num; i++)
		idx_que->bitmap[i] = ~(0UL);

	return 0;
}

struct ib_srq *hns_roce_create_srq(struct ib_pd *pd,
				   struct ib_srq_init_attr *srq_init_attr,
				   struct ib_udata *udata)
{
	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
	struct hns_roce_ib_create_srq_resp resp = {};
	struct hns_roce_srq *srq;
	int srq_desc_size;
	int srq_buf_size;
	u32 page_shift;
	int ret = 0;
	u32 npages;
	u32 cqn;

	/* Check the actual SRQ wqe and SRQ sge num */
	if (srq_init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs ||
	    srq_init_attr->attr.max_sge > hr_dev->caps.max_srq_sges)
		return ERR_PTR(-EINVAL);

	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
	if (!srq)
		return ERR_PTR(-ENOMEM);

	mutex_init(&srq->mutex);
	spin_lock_init(&srq->lock);

	srq->max = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
	srq->max_gs = srq_init_attr->attr.max_sge;

	srq_desc_size = max(16, 16 * srq->max_gs);

	srq->wqe_shift = ilog2(srq_desc_size);

	srq_buf_size = srq->max * srq_desc_size;

	srq->idx_que.entry_sz = HNS_ROCE_IDX_QUE_ENTRY_SZ;
	srq->idx_que.buf_size = srq->max * srq->idx_que.entry_sz;
	srq->mtt.mtt_type = MTT_TYPE_SRQWQE;
	srq->idx_que.mtt.mtt_type = MTT_TYPE_IDX;

	if (udata) {
		struct hns_roce_ib_create_srq  ucmd;

		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
			ret = -EFAULT;
			goto err_srq;
		}

		srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
					srq_buf_size, 0, 0);
		if (IS_ERR(srq->umem)) {
			ret = PTR_ERR(srq->umem);
			goto err_srq;
		}

		if (hr_dev->caps.srqwqe_buf_pg_sz) {
			npages = (ib_umem_page_count(srq->umem) +
				  (1 << hr_dev->caps.srqwqe_buf_pg_sz) - 1) /
				  (1 << hr_dev->caps.srqwqe_buf_pg_sz);
			page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
			ret = hns_roce_mtt_init(hr_dev, npages,
						page_shift,
						&srq->mtt);
		} else
			ret = hns_roce_mtt_init(hr_dev,
						ib_umem_page_count(srq->umem),
						srq->umem->page_shift,
						&srq->mtt);
		if (ret)
			goto err_buf;

		ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->mtt, srq->umem);
		if (ret)
			goto err_srq_mtt;

		/* config index queue BA */
		srq->idx_que.umem = ib_umem_get(pd->uobject->context,
						ucmd.que_addr,
						srq->idx_que.buf_size, 0, 0);
		if (IS_ERR(srq->idx_que.umem)) {
			dev_err(hr_dev->dev,
				"ib_umem_get error for index queue\n");
			ret = PTR_ERR(srq->idx_que.umem);
			goto err_srq_mtt;
		}

		if (hr_dev->caps.idx_buf_pg_sz) {
			npages = (ib_umem_page_count(srq->idx_que.umem) +
				  (1 << hr_dev->caps.idx_buf_pg_sz) - 1) /
				  (1 << hr_dev->caps.idx_buf_pg_sz);
			page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
			ret = hns_roce_mtt_init(hr_dev, npages,
						page_shift, &srq->idx_que.mtt);
		} else {
			ret = hns_roce_mtt_init(hr_dev,
				       ib_umem_page_count(srq->idx_que.umem),
				       srq->idx_que.umem->page_shift,
				       &srq->idx_que.mtt);
		}

		if (ret) {
			dev_err(hr_dev->dev,
				"hns_roce_mtt_init error for idx que\n");
			goto err_idx_mtt;
		}

		ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->idx_que.mtt,
						 srq->idx_que.umem);
		if (ret) {
			dev_err(hr_dev->dev,
			      "hns_roce_ib_umem_write_mtt error for idx que\n");
			goto err_idx_buf;
		}
	} else {
		page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
		if (hns_roce_buf_alloc(hr_dev, srq_buf_size,
				      (1 << page_shift) * 2,
				      &srq->buf, page_shift)) {
			ret = -ENOMEM;
			goto err_srq;
		}

		srq->head = 0;
		srq->tail = srq->max - 1;

		ret = hns_roce_mtt_init(hr_dev, srq->buf.npages,
					srq->buf.page_shift, &srq->mtt);
		if (ret)
			goto err_buf;

		ret = hns_roce_buf_write_mtt(hr_dev, &srq->mtt, &srq->buf);
		if (ret)
			goto err_srq_mtt;

		page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
		ret = hns_roce_create_idx_que(pd, srq, page_shift);
		if (ret) {
			dev_err(hr_dev->dev, "Create idx queue fail(%d)!\n",
				ret);
			goto err_srq_mtt;
		}

		/* Init mtt table for idx_que */
		ret = hns_roce_mtt_init(hr_dev, srq->idx_que.idx_buf.npages,
					srq->idx_que.idx_buf.page_shift,
					&srq->idx_que.mtt);
		if (ret)
			goto err_create_idx;

		/* Write buffer address into the mtt table */
		ret = hns_roce_buf_write_mtt(hr_dev, &srq->idx_que.mtt,
					     &srq->idx_que.idx_buf);
		if (ret)
			goto err_idx_buf;

		srq->wrid = kvmalloc_array(srq->max, sizeof(u64), GFP_KERNEL);
		if (!srq->wrid) {
			ret = -ENOMEM;
			goto err_idx_buf;
		}
	}

	cqn = ib_srq_has_cq(srq_init_attr->srq_type) ?
	      to_hr_cq(srq_init_attr->ext.cq)->cqn : 0;

	srq->db_reg_l = hr_dev->reg_base + SRQ_DB_REG;

	ret = hns_roce_srq_alloc(hr_dev, to_hr_pd(pd)->pdn, cqn, 0,
				 &srq->mtt, 0, srq);
	if (ret)
		goto err_wrid;

	srq->event = hns_roce_ib_srq_event;
	srq->ibsrq.ext.xrc.srq_num = srq->srqn;
	resp.srqn = srq->srqn;

	if (udata) {
		if (ib_copy_to_udata(udata, &resp,
				     min(udata->outlen, sizeof(resp)))) {
			ret = -EFAULT;
			goto err_srqc_alloc;
		}
	}

	return &srq->ibsrq;

err_srqc_alloc:
	hns_roce_srq_free(hr_dev, srq);

err_wrid:
	kvfree(srq->wrid);

err_idx_buf:
	hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);

err_idx_mtt:
	if (udata)
		ib_umem_release(srq->idx_que.umem);

err_create_idx:
	hns_roce_buf_free(hr_dev, srq->idx_que.buf_size,
			  &srq->idx_que.idx_buf);
	kfree(srq->idx_que.bitmap);

err_srq_mtt:
	hns_roce_mtt_cleanup(hr_dev, &srq->mtt);

err_buf:
	if (udata)
		ib_umem_release(srq->umem);
	else
		hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);

err_srq:
	kfree(srq);
	return ERR_PTR(ret);
}

int hns_roce_destroy_srq(struct ib_srq *ibsrq)
{
	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
	struct hns_roce_srq *srq = to_hr_srq(ibsrq);

	hns_roce_srq_free(hr_dev, srq);
	hns_roce_mtt_cleanup(hr_dev, &srq->mtt);

	if (ibsrq->uobject) {
		hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
		ib_umem_release(srq->idx_que.umem);
		ib_umem_release(srq->umem);
	} else {
		kvfree(srq->wrid);
		hns_roce_buf_free(hr_dev, srq->max << srq->wqe_shift,
				  &srq->buf);
	}

	kfree(srq);

	return 0;
}

int hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
{
	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;

	xa_init(&srq_table->xa);

	return hns_roce_bitmap_init(&srq_table->bitmap, hr_dev->caps.num_srqs,
				    hr_dev->caps.num_srqs - 1,
				    hr_dev->caps.reserved_srqs, 0);
}

void hns_roce_cleanup_srq_table(struct hns_roce_dev *hr_dev)
{
	hns_roce_bitmap_cleanup(&hr_dev->srq_table.bitmap);
}
Commit	Line	Data
5c1f167a LO	1	// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
	2	/*
	3	* Copyright (c) 2018 Hisilicon Limited.
	4	*/
	5
	6	#include <rdma/ib_umem.h>
	7	#include <rdma/hns-abi.h>
	8	#include "hns_roce_device.h"
	9	#include "hns_roce_cmd.h"
	10	#include "hns_roce_hem.h"
	11
81fce629 LO	12	void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)
	13	{
	14	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	15	struct hns_roce_srq *srq;
	16
	17	xa_lock(&srq_table->xa);
	18	srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
	19	if (srq)
	20	atomic_inc(&srq->refcount);
	21	xa_unlock(&srq_table->xa);
	22
	23	if (!srq) {
	24	dev_warn(hr_dev->dev, "Async event for bogus SRQ %08x\n", srqn);
	25	return;
	26	}
	27
	28	srq->event(srq, event_type);
	29
	30	if (atomic_dec_and_test(&srq->refcount))
	31	complete(&srq->free);
	32	}
	33	EXPORT_SYMBOL_GPL(hns_roce_srq_event);
	34
c7bcb134 LO	35	static void hns_roce_ib_srq_event(struct hns_roce_srq *srq,
	36	enum hns_roce_event event_type)
	37	{
	38	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
	39	struct ib_srq *ibsrq = &srq->ibsrq;
	40	struct ib_event event;
	41
	42	if (ibsrq->event_handler) {
	43	event.device = ibsrq->device;
	44	event.element.srq = ibsrq;
	45	switch (event_type) {
	46	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
	47	event.event = IB_EVENT_SRQ_LIMIT_REACHED;
	48	break;
	49	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
	50	event.event = IB_EVENT_SRQ_ERR;
	51	break;
	52	default:
	53	dev_err(hr_dev->dev,
	54	"hns_roce:Unexpected event type 0x%x on SRQ %06lx\n",
	55	event_type, srq->srqn);
	56	return;
	57	}
	58
	59	ibsrq->event_handler(&event, ibsrq->srq_context);
	60	}
	61	}
	62
	63	static int hns_roce_sw2hw_srq(struct hns_roce_dev *dev,
	64	struct hns_roce_cmd_mailbox *mailbox,
	65	unsigned long srq_num)
	66	{
	67	return hns_roce_cmd_mbox(dev, mailbox->dma, 0, srq_num, 0,
	68	HNS_ROCE_CMD_SW2HW_SRQ,
	69	HNS_ROCE_CMD_TIMEOUT_MSECS);
	70	}
	71
	72	static int hns_roce_hw2sw_srq(struct hns_roce_dev *dev,
	73	struct hns_roce_cmd_mailbox *mailbox,
	74	unsigned long srq_num)
	75	{
	76	return hns_roce_cmd_mbox(dev, 0, mailbox ? mailbox->dma : 0, srq_num,
	77	mailbox ? 0 : 1, HNS_ROCE_CMD_HW2SW_SRQ,
	78	HNS_ROCE_CMD_TIMEOUT_MSECS);
	79	}
	80
	81	int hns_roce_srq_alloc(struct hns_roce_dev *hr_dev, u32 pdn, u32 cqn, u16 xrcd,
	82	struct hns_roce_mtt *hr_mtt, u64 db_rec_addr,
	83	struct hns_roce_srq *srq)
	84	{
	85	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	86	struct hns_roce_cmd_mailbox *mailbox;
	87	dma_addr_t dma_handle_wqe;
	88	dma_addr_t dma_handle_idx;
	89	u64 *mtts_wqe;
	90	u64 *mtts_idx;
	91	int ret;
	92
	93	/* Get the physical address of srq buf */
	94	mtts_wqe = hns_roce_table_find(hr_dev,
	95	&hr_dev->mr_table.mtt_srqwqe_table,
	96	srq->mtt.first_seg,
	97	&dma_handle_wqe);
	98	if (!mtts_wqe) {
99	dev_err(hr_dev->dev,
100	"SRQ alloc.Failed to find srq buf addr.\n");
101	return -EINVAL;
102	}
103
104	/* Get physical address of idx que buf */
105	mtts_idx = hns_roce_table_find(hr_dev, &hr_dev->mr_table.mtt_idx_table,
106	srq->idx_que.mtt.first_seg,
107	&dma_handle_idx);
108	if (!mtts_idx) {
109	dev_err(hr_dev->dev,
110	"SRQ alloc.Failed to find idx que buf addr.\n");
111	return -EINVAL;
112	}
113
114	ret = hns_roce_bitmap_alloc(&srq_table->bitmap, &srq->srqn);
115	if (ret == -1) {
116	dev_err(hr_dev->dev, "SRQ alloc.Failed to alloc index.\n");
117	return -ENOMEM;
118	}
119
120	ret = hns_roce_table_get(hr_dev, &srq_table->table, srq->srqn);
121	if (ret)
122	goto err_out;
123
124	ret = xa_err(xa_store(&srq_table->xa, srq->srqn, srq, GFP_KERNEL));
125	if (ret)
126	goto err_put;
127
128	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
129	if (IS_ERR(mailbox)) {
130	ret = PTR_ERR(mailbox);
131	goto err_xa;
132	}
133
134	hr_dev->hw->write_srqc(hr_dev, srq, pdn, xrcd, cqn, mailbox->buf,
135	mtts_wqe, mtts_idx, dma_handle_wqe,
136	dma_handle_idx);
137
138	ret = hns_roce_sw2hw_srq(hr_dev, mailbox, srq->srqn);
139	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
140	if (ret)
141	goto err_xa;
142
143	atomic_set(&srq->refcount, 1);
144	init_completion(&srq->free);
145	return ret;
146
147	err_xa:
148	xa_erase(&srq_table->xa, srq->srqn);
149
150	err_put:
151	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
152
153	err_out:
154	hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
155	return ret;
156	}
157
158	void hns_roce_srq_free(struct hns_roce_dev hr_dev, struct hns_roce_srq srq)
159	{
160	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
161	int ret;
162
163	ret = hns_roce_hw2sw_srq(hr_dev, NULL, srq->srqn);
164	if (ret)
165	dev_err(hr_dev->dev, "HW2SW_SRQ failed (%d) for CQN %06lx\n",
166	ret, srq->srqn);
167
168	xa_erase(&srq_table->xa, srq->srqn);
169
170	if (atomic_dec_and_test(&srq->refcount))
171	complete(&srq->free);
172	wait_for_completion(&srq->free);
173
174	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
175	hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
176	}
177
178	static int hns_roce_create_idx_que(struct ib_pd pd, struct hns_roce_srq srq,
179	u32 page_shift)
180	{
181	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
182	struct hns_roce_idx_que *idx_que = &srq->idx_que;
183	u32 bitmap_num;
184	int i;
185
186	bitmap_num = HNS_ROCE_ALOGN_UP(srq->max, 8 * sizeof(u64));
187
188	idx_que->bitmap = kcalloc(1, bitmap_num / 8, GFP_KERNEL);
189	if (!idx_que->bitmap)
190	return -ENOMEM;
191
192	bitmap_num = bitmap_num / (8 * sizeof(u64));
193
194	idx_que->buf_size = srq->idx_que.buf_size;
195
196	if (hns_roce_buf_alloc(hr_dev, idx_que->buf_size, (1 << page_shift) * 2,
197	&idx_que->idx_buf, page_shift)) {
198	kfree(idx_que->bitmap);
199	return -ENOMEM;
200	}
201
202	for (i = 0; i < bitmap_num; i++)
203	idx_que->bitmap[i] = ~(0UL);
204
205	return 0;
206	}
207
208	struct ib_srq hns_roce_create_srq(struct ib_pd pd,
209	struct ib_srq_init_attr *srq_init_attr,
210	struct ib_udata *udata)
211	{
212	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
9d9d4ff7	213	struct hns_roce_ib_create_srq_resp resp = {};
c7bcb134 LO	214	struct hns_roce_srq *srq;
	215	int srq_desc_size;
	216	int srq_buf_size;
	217	u32 page_shift;
	218	int ret = 0;
	219	u32 npages;
	220	u32 cqn;
	221
	222	/* Check the actual SRQ wqe and SRQ sge num */
	223	if (srq_init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs \|\|
	224	srq_init_attr->attr.max_sge > hr_dev->caps.max_srq_sges)
	225	return ERR_PTR(-EINVAL);
	226
	227	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
	228	if (!srq)
	229	return ERR_PTR(-ENOMEM);
	230
	231	mutex_init(&srq->mutex);
	232	spin_lock_init(&srq->lock);
	233
	234	srq->max = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
	235	srq->max_gs = srq_init_attr->attr.max_sge;
	236
	237	srq_desc_size = max(16, 16 * srq->max_gs);
	238
	239	srq->wqe_shift = ilog2(srq_desc_size);
	240
	241	srq_buf_size = srq->max * srq_desc_size;
	242
	243	srq->idx_que.entry_sz = HNS_ROCE_IDX_QUE_ENTRY_SZ;
	244	srq->idx_que.buf_size = srq->max * srq->idx_que.entry_sz;
	245	srq->mtt.mtt_type = MTT_TYPE_SRQWQE;
	246	srq->idx_que.mtt.mtt_type = MTT_TYPE_IDX;
	247
	248	if (udata) {
	249	struct hns_roce_ib_create_srq ucmd;
	250
	251	if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
	252	ret = -EFAULT;
	253	goto err_srq;
	254	}
	255
	256	srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
	257	srq_buf_size, 0, 0);
	258	if (IS_ERR(srq->umem)) {
	259	ret = PTR_ERR(srq->umem);
	260	goto err_srq;
	261	}
	262
	263	if (hr_dev->caps.srqwqe_buf_pg_sz) {
	264	npages = (ib_umem_page_count(srq->umem) +
	265	(1 << hr_dev->caps.srqwqe_buf_pg_sz) - 1) /
	266	(1 << hr_dev->caps.srqwqe_buf_pg_sz);
	267	page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
	268	ret = hns_roce_mtt_init(hr_dev, npages,
	269	page_shift,
	270	&srq->mtt);
	271	} else
	272	ret = hns_roce_mtt_init(hr_dev,
	273	ib_umem_page_count(srq->umem),
	274	srq->umem->page_shift,
	275	&srq->mtt);
	276	if (ret)
	277	goto err_buf;
278
279	ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->mtt, srq->umem);
280	if (ret)
281	goto err_srq_mtt;
282
283	/* config index queue BA */
284	srq->idx_que.umem = ib_umem_get(pd->uobject->context,
285	ucmd.que_addr,
286	srq->idx_que.buf_size, 0, 0);
287	if (IS_ERR(srq->idx_que.umem)) {
288	dev_err(hr_dev->dev,
289	"ib_umem_get error for index queue\n");
e9dfa53a	290	ret = PTR_ERR(srq->idx_que.umem);
c7bcb134 LO	291	goto err_srq_mtt;
	292	}
	293
	294	if (hr_dev->caps.idx_buf_pg_sz) {
	295	npages = (ib_umem_page_count(srq->idx_que.umem) +
	296	(1 << hr_dev->caps.idx_buf_pg_sz) - 1) /
	297	(1 << hr_dev->caps.idx_buf_pg_sz);
	298	page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
	299	ret = hns_roce_mtt_init(hr_dev, npages,
	300	page_shift, &srq->idx_que.mtt);
	301	} else {
	302	ret = hns_roce_mtt_init(hr_dev,
	303	ib_umem_page_count(srq->idx_que.umem),
	304	srq->idx_que.umem->page_shift,
	305	&srq->idx_que.mtt);
	306	}
	307
	308	if (ret) {
	309	dev_err(hr_dev->dev,
	310	"hns_roce_mtt_init error for idx que\n");
	311	goto err_idx_mtt;
	312	}
	313
	314	ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->idx_que.mtt,
	315	srq->idx_que.umem);
	316	if (ret) {
	317	dev_err(hr_dev->dev,
	318	"hns_roce_ib_umem_write_mtt error for idx que\n");
	319	goto err_idx_buf;
	320	}
	321	} else {
	322	page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
	323	if (hns_roce_buf_alloc(hr_dev, srq_buf_size,
	324	(1 << page_shift) * 2,
	325	&srq->buf, page_shift)) {
	326	ret = -ENOMEM;
	327	goto err_srq;
	328	}
	329
	330	srq->head = 0;
	331	srq->tail = srq->max - 1;
	332
	333	ret = hns_roce_mtt_init(hr_dev, srq->buf.npages,
	334	srq->buf.page_shift, &srq->mtt);
	335	if (ret)
	336	goto err_buf;
	337
	338	ret = hns_roce_buf_write_mtt(hr_dev, &srq->mtt, &srq->buf);
	339	if (ret)
	340	goto err_srq_mtt;
	341
	342	page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
	343	ret = hns_roce_create_idx_que(pd, srq, page_shift);
	344	if (ret) {
	345	dev_err(hr_dev->dev, "Create idx queue fail(%d)!\n",
	346	ret);
	347	goto err_srq_mtt;
	348	}
	349
	350	/* Init mtt table for idx_que */
	351	ret = hns_roce_mtt_init(hr_dev, srq->idx_que.idx_buf.npages,
	352	srq->idx_que.idx_buf.page_shift,
	353	&srq->idx_que.mtt);
	354	if (ret)
355	goto err_create_idx;
356
357	/* Write buffer address into the mtt table */
358	ret = hns_roce_buf_write_mtt(hr_dev, &srq->idx_que.mtt,
359	&srq->idx_que.idx_buf);
360	if (ret)
361	goto err_idx_buf;
362
363	srq->wrid = kvmalloc_array(srq->max, sizeof(u64), GFP_KERNEL);
364	if (!srq->wrid) {
365	ret = -ENOMEM;
366	goto err_idx_buf;
367	}
368	}
369
370	cqn = ib_srq_has_cq(srq_init_attr->srq_type) ?
371	to_hr_cq(srq_init_attr->ext.cq)->cqn : 0;
372
373	srq->db_reg_l = hr_dev->reg_base + SRQ_DB_REG;
374
375	ret = hns_roce_srq_alloc(hr_dev, to_hr_pd(pd)->pdn, cqn, 0,
376	&srq->mtt, 0, srq);
377	if (ret)
378	goto err_wrid;
379
380	srq->event = hns_roce_ib_srq_event;
381	srq->ibsrq.ext.xrc.srq_num = srq->srqn;
9d9d4ff7	382	resp.srqn = srq->srqn;
c7bcb134	383
e00b64f7	384	if (udata) {
9d9d4ff7 LO	385	if (ib_copy_to_udata(udata, &resp,
9d9d4ff7 LO	386	min(udata->outlen, sizeof(resp)))) {
c7bcb134	387	ret = -EFAULT;
9d9d4ff7	388	goto err_srqc_alloc;
c7bcb134 LO	389	}
	390	}
	391
	392	return &srq->ibsrq;
	393
9d9d4ff7 LO	394	err_srqc_alloc:
	395	hns_roce_srq_free(hr_dev, srq);
	396
c7bcb134 LO	397	err_wrid:
	398	kvfree(srq->wrid);
	399
	400	err_idx_buf:
	401	hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
	402
	403	err_idx_mtt:
	404	if (udata)
	405	ib_umem_release(srq->idx_que.umem);
	406
	407	err_create_idx:
	408	hns_roce_buf_free(hr_dev, srq->idx_que.buf_size,
	409	&srq->idx_que.idx_buf);
	410	kfree(srq->idx_que.bitmap);
	411
	412	err_srq_mtt:
	413	hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
	414
	415	err_buf:
	416	if (udata)
	417	ib_umem_release(srq->umem);
	418	else
	419	hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);
	420
	421	err_srq:
	422	kfree(srq);
	423	return ERR_PTR(ret);
	424	}
	425
	426	int hns_roce_destroy_srq(struct ib_srq *ibsrq)
	427	{
	428	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
	429	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
	430
	431	hns_roce_srq_free(hr_dev, srq);
	432	hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
	433
	434	if (ibsrq->uobject) {
	435	hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
	436	ib_umem_release(srq->idx_que.umem);
	437	ib_umem_release(srq->umem);
	438	} else {
	439	kvfree(srq->wrid);
	440	hns_roce_buf_free(hr_dev, srq->max << srq->wqe_shift,
	441	&srq->buf);
	442	}
	443
	444	kfree(srq);
	445
	446	return 0;
	447	}
	448
5c1f167a LO	449	int hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
	450	{
	451	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
	452
	453	xa_init(&srq_table->xa);
	454
	455	return hns_roce_bitmap_init(&srq_table->bitmap, hr_dev->caps.num_srqs,
	456	hr_dev->caps.num_srqs - 1,
	457	hr_dev->caps.reserved_srqs, 0);
	458	}
	459
	460	void hns_roce_cleanup_srq_table(struct hns_roce_dev *hr_dev)
	461	{
	462	hns_roce_bitmap_cleanup(&hr_dev->srq_table.bitmap);
	463	}