Commit f03903b6 authored by Xi Wang's avatar Xi Wang Committed by Zheng Zengkai
Browse files

RDMA/hns: Optimize the base address table config for MTR 

mainline inclusion
from mainline-v5.14-rc1
commit 7b0006db
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/I5X0C3
CVE: NA

Reference: https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git/commit/?id=7b0006db6800da4c05883584befa46502d85dede

---------------------------------------------------------------------

The base address table is allocated by dma allocator, and the size is
always aligned to PAGE_SIZE. If a fixed size is used to allocate the
table, the number of base address entries stored in the table will be
smaller than that can actually stored.

Link: https://lore.kernel.org/r/1621589395-2435-2-git-send-email-liweihang@huawei.com


Signed-off-by: default avatarXi Wang <wangxi11@huawei.com>
Signed-off-by: default avatarWeihang Li <liweihang@huawei.com>
Signed-off-by: default avatarJason Gunthorpe <jgg@nvidia.com>
Reviewed-by: default avatarYangyang Li <liyangyang20@huawei.com>
Reviewed-by: default avatarHaoyue Xu <xuhaoyue1@hisilicon.com>
Reviewed-by: default avatarYueHaibing <yuehaibing@huawei.com>
Signed-off-by: default avatarZheng Zengkai <zhengzengkai@huawei.com>
parent 85566b9e

drivers/infiniband/hw/hns/hns_roce_alloc.c

+21 −30
Original line number Diff line number Diff line
@@ -86,10 +86,10 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, 


	/* Calc the trunk size and num by required size and page_shift */

	if (flags & HNS_ROCE_BUF_DIRECT) {

		buf->trunk_shift = ilog2(ALIGN(size, PAGE_SIZE));

		buf->trunk_shift = order_base_2(ALIGN(size, PAGE_SIZE));

		ntrunk = 1;

	} else {

		buf->trunk_shift = ilog2(ALIGN(page_size, PAGE_SIZE));

		buf->trunk_shift = order_base_2(ALIGN(page_size, PAGE_SIZE));

		ntrunk = DIV_ROUND_UP(size, 1 << buf->trunk_shift);

	}
@@ -130,51 +130,42 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, 
}



int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,

			   int buf_cnt, int start, struct hns_roce_buf *buf)

			   int buf_cnt, struct hns_roce_buf *buf,

			   unsigned int page_shift)

{

	int i, end;

	int total;



	end = start + buf_cnt;

	if (end > buf->npages) {

		dev_err(hr_dev->dev,

			"failed to check kmem bufs, end %d + %d total %u!\n",

			start, buf_cnt, buf->npages);

	unsigned int offset, max_size;

	int total = 0;

	int i;



	if (page_shift > buf->trunk_shift) {

		dev_err(hr_dev->dev, "failed to check kmem buf shift %u > %u\n",

			page_shift, buf->trunk_shift);

		return -EINVAL;

	}



	total = 0;

	for (i = start; i < end; i++)

		bufs[total++] = hns_roce_buf_page(buf, i);

	offset = 0;

	max_size = buf->ntrunks << buf->trunk_shift;

	for (i = 0; i < buf_cnt && offset < max_size; i++) {

		bufs[total++] = hns_roce_buf_dma_addr(buf, offset);

		offset += (1 << page_shift);

	}



	return total;

}



int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,

			   int buf_cnt, int start, struct ib_umem *umem,

			   int buf_cnt, struct ib_umem *umem,

			   unsigned int page_shift)

{

	struct ib_block_iter biter;

	int total = 0;

	int idx = 0;

	u64 addr;



	if (page_shift < HNS_HW_PAGE_SHIFT) {

		dev_err(hr_dev->dev, "failed to check umem page shift %u!\n",

			page_shift);

		return -EINVAL;

	}



	/* convert system page cnt to hw page cnt */

	rdma_umem_for_each_dma_block(umem, &biter, 1 << page_shift) {

		addr = rdma_block_iter_dma_address(&biter);

		if (idx >= start) {

			bufs[total++] = addr;

		bufs[total++] = rdma_block_iter_dma_address(&biter);

		if (total >= buf_cnt)

			goto done;

	}

		idx++;

	}



done:

	return total;

drivers/infiniband/hw/hns/hns_roce_cq.c

+2 −2
Original line number Diff line number Diff line
@@ -203,13 +203,13 @@ static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, 
	struct hns_roce_buf_attr buf_attr = {};

	int ret;



	buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;

	buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + PAGE_SHIFT;

	buf_attr.region[0].size = hr_cq->cq_depth * hr_cq->cqe_size;

	buf_attr.region[0].hopnum = hr_dev->caps.cqe_hop_num;

	buf_attr.region_count = 1;



	ret = hns_roce_mtr_create(hr_dev, &hr_cq->mtr, &buf_attr,

				  hr_dev->caps.cqe_ba_pg_sz + HNS_HW_PAGE_SHIFT,

				  hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT,

				  udata, addr);

	if (ret)

		ibdev_err(ibdev, "Failed to alloc CQ mtr, ret = %d\n", ret);

drivers/infiniband/hw/hns/hns_roce_device.h

+10 −5
Original line number Diff line number Diff line
@@ -1029,14 +1029,18 @@ static inline void *hns_roce_buf_offset(struct hns_roce_buf *buf, 
			(offset & ((1 << buf->trunk_shift) - 1));

}



static inline dma_addr_t hns_roce_buf_page(struct hns_roce_buf *buf, u32 idx)

static inline dma_addr_t hns_roce_buf_dma_addr(struct hns_roce_buf *buf,

					       unsigned int offset)

{

	unsigned int offset = idx << buf->page_shift;



	return buf->trunk_list[offset >> buf->trunk_shift].map +

			(offset & ((1 << buf->trunk_shift) - 1));

}



static inline dma_addr_t hns_roce_buf_page(struct hns_roce_buf *buf, u32 idx)

{

	return hns_roce_buf_dma_addr(buf, idx << buf->page_shift);

}



#define hr_hw_page_align(x)		ALIGN(x, 1 << HNS_HW_PAGE_SHIFT)



static inline u64 to_hr_hw_page_addr(u64 addr)
@@ -1152,9 +1156,10 @@ struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, 
					u32 page_shift, u32 flags);



int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,

			   int buf_cnt, int start, struct hns_roce_buf *buf);

			   int buf_cnt, struct hns_roce_buf *buf,

			   unsigned int page_shift);

int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,

			   int buf_cnt, int start, struct ib_umem *umem,

			   int buf_cnt, struct ib_umem *umem,

			   unsigned int page_shift);



int hns_roce_create_srq(struct ib_srq *srq,

drivers/infiniband/hw/hns/hns_roce_hw_v2.c

+22 −4
Original line number Diff line number Diff line
@@ -2083,6 +2083,8 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev) 
	caps->llm_buf_pg_sz = 0;



	/* MR */

	caps->mpt_ba_pg_sz = 0;

	caps->mpt_buf_pg_sz = 0;

	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;

	caps->pbl_buf_pg_sz = 0;

	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
@@ -2090,8 +2092,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev) 
		   HEM_TYPE_MTPT);



	/* QP */

	caps->qpc_ba_pg_sz = 0;

	caps->qpc_buf_pg_sz = 0;

	caps->qpc_timer_ba_pg_sz = 0;

	caps->qpc_timer_buf_pg_sz = 0;

	caps->sccc_ba_pg_sz = 0;

	caps->sccc_buf_pg_sz = 0;

	caps->mtt_ba_pg_sz = 0;

	caps->mtt_buf_pg_sz = 0;

	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
@@ -2104,6 +2110,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev) 
			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);



	/* CQ */

	caps->cqc_ba_pg_sz = 0;

	caps->cqc_buf_pg_sz = 0;

	caps->cqc_timer_ba_pg_sz = 0;

	caps->cqc_timer_buf_pg_sz = 0;

	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;

	caps->cqe_buf_pg_sz = 0;

	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,

		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,

		   HEM_TYPE_CQC);
@@ -2112,6 +2124,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev) 


	/* SRQ */

	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {

		caps->srqc_ba_pg_sz = 0;

		caps->srqc_buf_pg_sz = 0;

		caps->srqwqe_ba_pg_sz = 0;

		caps->srqwqe_buf_pg_sz = 0;

		caps->idx_ba_pg_sz = 0;

		caps->idx_buf_pg_sz = 0;

		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,

			   caps->srqc_hop_num, caps->srqc_bt_num,

			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
@@ -6365,14 +6383,14 @@ static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq) 
	else

		eq->hop_num = hr_dev->caps.eqe_hop_num;



	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;

	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;

	buf_attr.region[0].size = eq->entries * eq->eqe_size;

	buf_attr.region[0].hopnum = eq->hop_num;

	buf_attr.region_count = 1;



	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,

				  hr_dev->caps.eqe_ba_pg_sz +

				  HNS_HW_PAGE_SHIFT, NULL, 0);

				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,

				  0);

	if (err)

		dev_err(hr_dev->dev, "Failed to alloc EQE mtr, err %d\n", err);

drivers/infiniband/hw/hns/hns_roce_mr.c

+7 −7
Original line number Diff line number Diff line
@@ -104,7 +104,7 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, 
	buf_attr.mtt_only = is_fast;



	err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,

				  hr_dev->caps.pbl_ba_pg_sz + HNS_HW_PAGE_SHIFT,

				  hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,

				  udata, start);

	if (err)

		ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
@@ -705,11 +705,11 @@ static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 
		return -ENOMEM;



	if (mtr->umem)

		npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count, 0,

		npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count,

					       mtr->umem, page_shift);

	else

		npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count, 0,

					       mtr->kmem);

		npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count,

					       mtr->kmem, page_shift);



	if (npage != page_count) {

		ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage,
@@ -721,8 +721,8 @@ static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 
	if (mtr->hem_cfg.is_direct && npage > 1) {

		ret = mtr_check_direct_pages(pages, npage, page_shift);

		if (ret) {

			ibdev_err(ibdev, "failed to check %s mtr, idx = %d.\n",

				  mtr->umem ? "user" : "kernel", ret);

			ibdev_err(ibdev, "failed to check %s page: %d / %d.\n",

				  mtr->umem ? "umtr" : "kmtr", ret, npage);

			ret = -ENOBUFS;

			goto err_alloc_list;

		}
@@ -767,7 +767,7 @@ int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 
		if (r->offset + r->count > page_cnt) {

			ret = -EINVAL;

			ibdev_err(ibdev,

				  "failed to check mtr%u end %u + %u, max %u.\n",

				  "failed to check mtr%u count %u + %u > %u.\n",

				  i, r->offset, r->count, page_cnt);

			return ret;

		}