* [PATCH for-next 1/5] RDMA/hns: Optimize the base address table config for MTR
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
@ 2021-05-21 9:29 ` Weihang Li
2021-05-21 9:29 ` [PATCH for-next 2/5] RDMA/hns: Refactor root BT allocation " Weihang Li
` (4 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Weihang Li @ 2021-05-21 9:29 UTC (permalink / raw)
To: dledford, jgg; +Cc: leon, linux-rdma, linuxarm, Xi Wang, Weihang Li
From: Xi Wang <wangxi11@huawei.com>
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.
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_alloc.c | 51 ++++++++++++-----------------
drivers/infiniband/hw/hns/hns_roce_cq.c | 4 +--
drivers/infiniband/hw/hns/hns_roce_device.h | 15 ++++++---
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 26 ++++++++++++---
drivers/infiniband/hw/hns/hns_roce_mr.c | 14 ++++----
drivers/infiniband/hw/hns/hns_roce_qp.c | 2 +-
drivers/infiniband/hw/hns/hns_roce_srq.c | 10 +++---
7 files changed, 68 insertions(+), 54 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_alloc.c b/drivers/infiniband/hw/hns/hns_roce_alloc.c
index 5d389ed..51374b68 100644
--- a/drivers/infiniband/hw/hns/hns_roce_alloc.c
+++ b/drivers/infiniband/hw/hns/hns_roce_alloc.c
@@ -208,10 +208,10 @@ 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);
}
@@ -252,50 +252,41 @@ 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;
- if (total >= buf_cnt)
- goto done;
- }
- idx++;
+ bufs[total++] = rdma_block_iter_dma_address(&biter);
+ if (total >= buf_cnt)
+ goto done;
}
done:
diff --git a/drivers/infiniband/hw/hns/hns_roce_cq.c b/drivers/infiniband/hw/hns/hns_roce_cq.c
index 488d86b..a5a74b6 100644
--- a/drivers/infiniband/hw/hns/hns_roce_cq.c
+++ b/drivers/infiniband/hw/hns/hns_roce_cq.c
@@ -202,13 +202,13 @@ 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);
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index 111cab5..02fffb7 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -1060,14 +1060,18 @@ 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)
@@ -1204,9 +1208,10 @@ 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,
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index d5e72e5..0c12d87 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -2018,6 +2018,8 @@ 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,
@@ -2025,8 +2027,12 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
HEM_TYPE_MTPT);
/* QP */
- caps->qpc_timer_ba_pg_sz = 0;
+ 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,
@@ -2039,6 +2045,12 @@ 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);
@@ -2053,6 +2065,12 @@ 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,
@@ -6161,14 +6179,14 @@ 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);
diff --git a/drivers/infiniband/hw/hns/hns_roce_mr.c b/drivers/infiniband/hw/hns/hns_roce_mr.c
index 79b3c30..8e6b1ae 100644
--- a/drivers/infiniband/hw/hns/hns_roce_mr.c
+++ b/drivers/infiniband/hw/hns/hns_roce_mr.c
@@ -122,7 +122,7 @@ 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);
@@ -737,11 +737,11 @@ 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,
@@ -753,8 +753,8 @@ 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;
}
@@ -799,7 +799,7 @@ 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;
}
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index c6e120e..9203cf1 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -761,7 +761,7 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
goto err_inline;
}
ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
- HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
+ PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
udata, addr);
if (ret) {
ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
diff --git a/drivers/infiniband/hw/hns/hns_roce_srq.c b/drivers/infiniband/hw/hns/hns_roce_srq.c
index 546d182..c842210 100644
--- a/drivers/infiniband/hw/hns/hns_roce_srq.c
+++ b/drivers/infiniband/hw/hns/hns_roce_srq.c
@@ -167,14 +167,14 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
srq->idx_que.entry_shift = ilog2(HNS_ROCE_IDX_QUE_ENTRY_SZ);
- buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+ buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + PAGE_SHIFT;
buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
srq->idx_que.entry_shift);
buf_attr.region[0].hopnum = hr_dev->caps.idx_hop_num;
buf_attr.region_count = 1;
ret = hns_roce_mtr_create(hr_dev, &idx_que->mtr, &buf_attr,
- hr_dev->caps.idx_ba_pg_sz + HNS_HW_PAGE_SHIFT,
+ hr_dev->caps.idx_ba_pg_sz + PAGE_SHIFT,
udata, addr);
if (ret) {
ibdev_err(ibdev,
@@ -222,15 +222,15 @@ static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev,
HNS_ROCE_SGE_SIZE *
srq->max_gs)));
- buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + HNS_HW_PAGE_SHIFT;
+ buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + PAGE_SHIFT;
buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
srq->wqe_shift);
buf_attr.region[0].hopnum = hr_dev->caps.srqwqe_hop_num;
buf_attr.region_count = 1;
ret = hns_roce_mtr_create(hr_dev, &srq->buf_mtr, &buf_attr,
- hr_dev->caps.srqwqe_ba_pg_sz +
- HNS_HW_PAGE_SHIFT, udata, addr);
+ hr_dev->caps.srqwqe_ba_pg_sz + PAGE_SHIFT,
+ udata, addr);
if (ret)
ibdev_err(ibdev,
"failed to alloc SRQ buf mtr, ret = %d.\n", ret);
--
2.7.4
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH for-next 2/5] RDMA/hns: Refactor root BT allocation for MTR
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
2021-05-21 9:29 ` [PATCH for-next 1/5] RDMA/hns: Optimize the base address table config for MTR Weihang Li
@ 2021-05-21 9:29 ` Weihang Li
2021-05-21 9:29 ` [PATCH for-next 3/5] RDMA/hns: Fix wrong timer context buffer page size Weihang Li
` (3 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Weihang Li @ 2021-05-21 9:29 UTC (permalink / raw)
To: dledford, jgg; +Cc: leon, linux-rdma, linuxarm, Xi Wang, Weihang Li
From: Xi Wang <wangxi11@huawei.com>
Split the hem_list_alloc_root_bt() into serval small functions to make the
code flow more clear.
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_hem.c | 230 ++++++++++++++++++++-----------
1 file changed, 146 insertions(+), 84 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c b/drivers/infiniband/hw/hns/hns_roce_hem.c
index cfd2e1b..63776da 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
@@ -1053,7 +1053,7 @@ void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
}
-struct roce_hem_item {
+struct hns_roce_hem_item {
struct list_head list; /* link all hems in the same bt level */
struct list_head sibling; /* link all hems in last hop for mtt */
void *addr;
@@ -1063,12 +1063,18 @@ struct roce_hem_item {
int end; /* end buf offset in this hem */
};
-static struct roce_hem_item *hem_list_alloc_item(struct hns_roce_dev *hr_dev,
- int start, int end,
- int count, bool exist_bt,
- int bt_level)
+/* All HEM items are linked in a tree structure */
+struct hns_roce_hem_head {
+ struct list_head branch[HNS_ROCE_MAX_BT_REGION];
+ struct list_head root;
+ struct list_head leaf;
+};
+
+static struct hns_roce_hem_item *
+hem_list_alloc_item(struct hns_roce_dev *hr_dev, int start, int end, int count,
+ bool exist_bt, int bt_level)
{
- struct roce_hem_item *hem;
+ struct hns_roce_hem_item *hem;
hem = kzalloc(sizeof(*hem), GFP_KERNEL);
if (!hem)
@@ -1093,7 +1099,7 @@ static struct roce_hem_item *hem_list_alloc_item(struct hns_roce_dev *hr_dev,
}
static void hem_list_free_item(struct hns_roce_dev *hr_dev,
- struct roce_hem_item *hem, bool exist_bt)
+ struct hns_roce_hem_item *hem, bool exist_bt)
{
if (exist_bt)
dma_free_coherent(hr_dev->dev, hem->count * BA_BYTE_LEN,
@@ -1104,7 +1110,7 @@ static void hem_list_free_item(struct hns_roce_dev *hr_dev,
static void hem_list_free_all(struct hns_roce_dev *hr_dev,
struct list_head *head, bool exist_bt)
{
- struct roce_hem_item *hem, *temp_hem;
+ struct hns_roce_hem_item *hem, *temp_hem;
list_for_each_entry_safe(hem, temp_hem, head, list) {
list_del(&hem->list);
@@ -1120,24 +1126,24 @@ static void hem_list_link_bt(struct hns_roce_dev *hr_dev, void *base_addr,
/* assign L0 table address to hem from root bt */
static void hem_list_assign_bt(struct hns_roce_dev *hr_dev,
- struct roce_hem_item *hem, void *cpu_addr,
+ struct hns_roce_hem_item *hem, void *cpu_addr,
u64 phy_addr)
{
hem->addr = cpu_addr;
hem->dma_addr = (dma_addr_t)phy_addr;
}
-static inline bool hem_list_page_is_in_range(struct roce_hem_item *hem,
+static inline bool hem_list_page_is_in_range(struct hns_roce_hem_item *hem,
int offset)
{
return (hem->start <= offset && offset <= hem->end);
}
-static struct roce_hem_item *hem_list_search_item(struct list_head *ba_list,
- int page_offset)
+static struct hns_roce_hem_item *hem_list_search_item(struct list_head *ba_list,
+ int page_offset)
{
- struct roce_hem_item *hem, *temp_hem;
- struct roce_hem_item *found = NULL;
+ struct hns_roce_hem_item *hem, *temp_hem;
+ struct hns_roce_hem_item *found = NULL;
list_for_each_entry_safe(hem, temp_hem, ba_list, list) {
if (hem_list_page_is_in_range(hem, page_offset)) {
@@ -1227,9 +1233,9 @@ static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
int offset, struct list_head *mid_bt,
struct list_head *btm_bt)
{
- struct roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
+ struct hns_roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
struct list_head temp_list[HNS_ROCE_MAX_BT_LEVEL];
- struct roce_hem_item *cur, *pre;
+ struct hns_roce_hem_item *cur, *pre;
const int hopnum = r->hopnum;
int start_aligned;
int distance;
@@ -1307,56 +1313,96 @@ static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
return ret;
}
-static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
- struct hns_roce_hem_list *hem_list, int unit,
- const struct hns_roce_buf_region *regions,
- int region_cnt)
+static struct hns_roce_hem_item *
+alloc_root_hem(struct hns_roce_dev *hr_dev, int unit, int *max_ba_num,
+ const struct hns_roce_buf_region *regions, int region_cnt)
{
- struct list_head temp_list[HNS_ROCE_MAX_BT_REGION];
- struct roce_hem_item *hem, *temp_hem, *root_hem;
const struct hns_roce_buf_region *r;
- struct list_head temp_root;
- struct list_head temp_btm;
- void *cpu_base;
- u64 phy_base;
- int ret = 0;
+ struct hns_roce_hem_item *hem;
int ba_num;
int offset;
- int total;
- int step;
- int i;
-
- r = ®ions[0];
- root_hem = hem_list_search_item(&hem_list->root_bt, r->offset);
- if (root_hem)
- return 0;
ba_num = hns_roce_hem_list_calc_root_ba(regions, region_cnt, unit);
if (ba_num < 1)
- return -ENOMEM;
+ return ERR_PTR(-ENOMEM);
if (ba_num > unit)
- return -ENOBUFS;
+ return ERR_PTR(-ENOBUFS);
- ba_num = min_t(int, ba_num, unit);
- INIT_LIST_HEAD(&temp_root);
- offset = r->offset;
+ offset = regions[0].offset;
/* indicate to last region */
r = ®ions[region_cnt - 1];
- root_hem = hem_list_alloc_item(hr_dev, offset, r->offset + r->count - 1,
- ba_num, true, 0);
- if (!root_hem)
+ hem = hem_list_alloc_item(hr_dev, offset, r->offset + r->count - 1,
+ ba_num, true, 0);
+ if (!hem)
+ return ERR_PTR(-ENOMEM);
+
+ *max_ba_num = ba_num;
+
+ return hem;
+}
+
+static int alloc_fake_root_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
+ u64 phy_base, const struct hns_roce_buf_region *r,
+ struct list_head *branch_head,
+ struct list_head *leaf_head)
+{
+ struct hns_roce_hem_item *hem;
+
+ hem = hem_list_alloc_item(hr_dev, r->offset, r->offset + r->count - 1,
+ r->count, false, 0);
+ if (!hem)
return -ENOMEM;
- list_add(&root_hem->list, &temp_root);
- hem_list->root_ba = root_hem->dma_addr;
+ hem_list_assign_bt(hr_dev, hem, cpu_base, phy_base);
+ list_add(&hem->list, branch_head);
+ list_add(&hem->sibling, leaf_head);
- INIT_LIST_HEAD(&temp_btm);
- for (i = 0; i < region_cnt; i++)
- INIT_LIST_HEAD(&temp_list[i]);
+ return r->count;
+}
+
+static int setup_middle_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
+ int unit, const struct hns_roce_buf_region *r,
+ const struct list_head *branch_head)
+{
+ struct hns_roce_hem_item *hem, *temp_hem;
+ int total = 0;
+ int offset;
+ int step;
+
+ step = hem_list_calc_ba_range(r->hopnum, 1, unit);
+ if (step < 1)
+ return -EINVAL;
+
+ /* if exist mid bt, link L1 to L0 */
+ list_for_each_entry_safe(hem, temp_hem, branch_head, list) {
+ offset = (hem->start - r->offset) / step * BA_BYTE_LEN;
+ hem_list_link_bt(hr_dev, cpu_base + offset, hem->dma_addr);
+ total++;
+ }
+
+ return total;
+}
+
+static int
+setup_root_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem_list *hem_list,
+ int unit, int max_ba_num, struct hns_roce_hem_head *head,
+ const struct hns_roce_buf_region *regions, int region_cnt)
+{
+ const struct hns_roce_buf_region *r;
+ struct hns_roce_hem_item *root_hem;
+ void *cpu_base;
+ u64 phy_base;
+ int i, total;
+ int ret;
+
+ root_hem = list_first_entry(&head->root,
+ struct hns_roce_hem_item, list);
+ if (!root_hem)
+ return -ENOMEM;
total = 0;
- for (i = 0; i < region_cnt && total < ba_num; i++) {
+ for (i = 0; i < region_cnt && total < max_ba_num; i++) {
r = ®ions[i];
if (!r->count)
continue;
@@ -1368,48 +1414,64 @@ static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
/* if hopnum is 0 or 1, cut a new fake hem from the root bt
* which's address share to all regions.
*/
- if (hem_list_is_bottom_bt(r->hopnum, 0)) {
- hem = hem_list_alloc_item(hr_dev, r->offset,
- r->offset + r->count - 1,
- r->count, false, 0);
- if (!hem) {
- ret = -ENOMEM;
- goto err_exit;
- }
- hem_list_assign_bt(hr_dev, hem, cpu_base, phy_base);
- list_add(&hem->list, &temp_list[i]);
- list_add(&hem->sibling, &temp_btm);
- total += r->count;
- } else {
- step = hem_list_calc_ba_range(r->hopnum, 1, unit);
- if (step < 1) {
- ret = -EINVAL;
- goto err_exit;
- }
- /* if exist mid bt, link L1 to L0 */
- list_for_each_entry_safe(hem, temp_hem,
- &hem_list->mid_bt[i][1], list) {
- offset = (hem->start - r->offset) / step *
- BA_BYTE_LEN;
- hem_list_link_bt(hr_dev, cpu_base + offset,
- hem->dma_addr);
- total++;
- }
- }
+ if (hem_list_is_bottom_bt(r->hopnum, 0))
+ ret = alloc_fake_root_bt(hr_dev, cpu_base, phy_base, r,
+ &head->branch[i], &head->leaf);
+ else
+ ret = setup_middle_bt(hr_dev, cpu_base, unit, r,
+ &hem_list->mid_bt[i][1]);
+
+ if (ret < 0)
+ return ret;
+
+ total += ret;
}
- list_splice(&temp_btm, &hem_list->btm_bt);
- list_splice(&temp_root, &hem_list->root_bt);
+ list_splice(&head->leaf, &hem_list->btm_bt);
+ list_splice(&head->root, &hem_list->root_bt);
for (i = 0; i < region_cnt; i++)
- list_splice(&temp_list[i], &hem_list->mid_bt[i][0]);
+ list_splice(&head->branch[i], &hem_list->mid_bt[i][0]);
return 0;
+}
-err_exit:
+static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
+ struct hns_roce_hem_list *hem_list, int unit,
+ const struct hns_roce_buf_region *regions,
+ int region_cnt)
+{
+ struct hns_roce_hem_item *root_hem;
+ struct hns_roce_hem_head head;
+ int max_ba_num;
+ int ret;
+ int i;
+
+ root_hem = hem_list_search_item(&hem_list->root_bt, regions[0].offset);
+ if (root_hem)
+ return 0;
+
+ max_ba_num = 0;
+ root_hem = alloc_root_hem(hr_dev, unit, &max_ba_num, regions,
+ region_cnt);
+ if (IS_ERR(root_hem))
+ return PTR_ERR(root_hem);
+
+ /* List head for storing all allocated HEM items */
+ INIT_LIST_HEAD(&head.root);
+ INIT_LIST_HEAD(&head.leaf);
for (i = 0; i < region_cnt; i++)
- hem_list_free_all(hr_dev, &temp_list[i], false);
+ INIT_LIST_HEAD(&head.branch[i]);
- hem_list_free_all(hr_dev, &temp_root, true);
+ hem_list->root_ba = root_hem->dma_addr;
+ list_add(&root_hem->list, &head.root);
+ ret = setup_root_hem(hr_dev, hem_list, unit, max_ba_num, &head, regions,
+ region_cnt);
+ if (ret) {
+ for (i = 0; i < region_cnt; i++)
+ hem_list_free_all(hr_dev, &head.branch[i], false);
+
+ hem_list_free_all(hr_dev, &head.root, true);
+ }
return ret;
}
@@ -1495,7 +1557,7 @@ void *hns_roce_hem_list_find_mtt(struct hns_roce_dev *hr_dev,
int offset, int *mtt_cnt, u64 *phy_addr)
{
struct list_head *head = &hem_list->btm_bt;
- struct roce_hem_item *hem, *temp_hem;
+ struct hns_roce_hem_item *hem, *temp_hem;
void *cpu_base = NULL;
u64 phy_base = 0;
int nr = 0;
--
2.7.4
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH for-next 3/5] RDMA/hns: Fix wrong timer context buffer page size
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
2021-05-21 9:29 ` [PATCH for-next 1/5] RDMA/hns: Optimize the base address table config for MTR Weihang Li
2021-05-21 9:29 ` [PATCH for-next 2/5] RDMA/hns: Refactor root BT allocation " Weihang Li
@ 2021-05-21 9:29 ` Weihang Li
2021-05-21 9:29 ` [PATCH for-next 4/5] RDMA/hns: Use refcount_t APIs for HEM Weihang Li
` (2 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Weihang Li @ 2021-05-21 9:29 UTC (permalink / raw)
To: dledford, jgg; +Cc: leon, linux-rdma, linuxarm, Xi Wang, Weihang Li
From: Xi Wang <wangxi11@huawei.com>
The HEM page size for QPC timer and CQC timer is always 4K and there's no
need to calculate a different size by the hns driver, otherwise the ROCEE
may access an invalid address.
Fixes: 719d13415f59 ("RDMA/hns: Remove duplicated hem page size config code")
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 6 ------
1 file changed, 6 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 0c12d87..17cf51a 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -2057,12 +2057,6 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev)
calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
- if (caps->cqc_timer_entry_sz)
- calc_pg_sz(caps->num_cqc_timer, caps->cqc_timer_entry_sz,
- caps->cqc_timer_hop_num, caps->cqc_timer_bt_num,
- &caps->cqc_timer_buf_pg_sz,
- &caps->cqc_timer_ba_pg_sz, HEM_TYPE_CQC_TIMER);
-
/* SRQ */
if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
caps->srqc_ba_pg_sz = 0;
--
2.7.4
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH for-next 4/5] RDMA/hns: Use refcount_t APIs for HEM
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
` (2 preceding siblings ...)
2021-05-21 9:29 ` [PATCH for-next 3/5] RDMA/hns: Fix wrong timer context buffer page size Weihang Li
@ 2021-05-21 9:29 ` Weihang Li
2021-05-21 9:29 ` [PATCH for-next 5/5] RDMA/hns: Clean the hardware related code " Weihang Li
2021-05-28 23:20 ` [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Weihang Li @ 2021-05-21 9:29 UTC (permalink / raw)
To: dledford, jgg; +Cc: leon, linux-rdma, linuxarm, Weihang Li
refcount_t is better than integer for reference counting, it will WARN on
overflow/underflow and avoid use-after-free risks.
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_hem.c | 32 +++++++++++++++-----------------
drivers/infiniband/hw/hns/hns_roce_hem.h | 4 ++--
2 files changed, 17 insertions(+), 19 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c b/drivers/infiniband/hw/hns/hns_roce_hem.c
index 63776da..fe6940b 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
@@ -271,7 +271,6 @@ static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
if (!hem)
return NULL;
- hem->refcount = 0;
INIT_LIST_HEAD(&hem->chunk_list);
order = get_order(hem_alloc_size);
@@ -618,7 +617,7 @@ static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
mutex_lock(&table->mutex);
if (table->hem[index.buf]) {
- ++table->hem[index.buf]->refcount;
+ refcount_inc(&table->hem[index.buf]->refcount);
goto out;
}
@@ -637,7 +636,7 @@ static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
}
}
- ++table->hem[index.buf]->refcount;
+ refcount_set(&table->hem[index.buf]->refcount, 1);
goto out;
err_alloc:
@@ -663,7 +662,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,
mutex_lock(&table->mutex);
if (table->hem[i]) {
- ++table->hem[i]->refcount;
+ refcount_inc(&table->hem[i]->refcount);
goto out;
}
@@ -686,7 +685,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,
goto out;
}
- ++table->hem[i]->refcount;
+ refcount_set(&table->hem[i]->refcount, 1);
out:
mutex_unlock(&table->mutex);
return ret;
@@ -753,11 +752,11 @@ static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
return;
}
- mutex_lock(&table->mutex);
- if (check_refcount && (--table->hem[index.buf]->refcount > 0)) {
- mutex_unlock(&table->mutex);
+ if (!check_refcount)
+ mutex_lock(&table->mutex);
+ else if (!refcount_dec_and_mutex_lock(&table->hem[index.buf]->refcount,
+ &table->mutex))
return;
- }
clear_mhop_hem(hr_dev, table, obj, &mhop, &index);
free_mhop_hem(hr_dev, table, &mhop, &index);
@@ -779,16 +778,15 @@ void hns_roce_table_put(struct hns_roce_dev *hr_dev,
i = (obj & (table->num_obj - 1)) /
(table->table_chunk_size / table->obj_size);
- mutex_lock(&table->mutex);
+ if (!refcount_dec_and_mutex_lock(&table->hem[i]->refcount,
+ &table->mutex))
+ return;
- if (--table->hem[i]->refcount == 0) {
- /* Clear HEM base address */
- if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
- dev_warn(dev, "Clear HEM base address failed.\n");
+ if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
+ dev_warn(dev, "failed to clear HEM base address.\n");
- hns_roce_free_hem(hr_dev, table->hem[i]);
- table->hem[i] = NULL;
- }
+ hns_roce_free_hem(hr_dev, table->hem[i]);
+ table->hem[i] = NULL;
mutex_unlock(&table->mutex);
}
diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.h b/drivers/infiniband/hw/hns/hns_roce_hem.h
index 13fdeb3..ffa65e8 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.h
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.h
@@ -88,8 +88,8 @@ struct hns_roce_hem_chunk {
};
struct hns_roce_hem {
- struct list_head chunk_list;
- int refcount;
+ struct list_head chunk_list;
+ refcount_t refcount;
};
struct hns_roce_hem_iter {
--
2.7.4
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH for-next 5/5] RDMA/hns: Clean the hardware related code for HEM
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
` (3 preceding siblings ...)
2021-05-21 9:29 ` [PATCH for-next 4/5] RDMA/hns: Use refcount_t APIs for HEM Weihang Li
@ 2021-05-21 9:29 ` Weihang Li
2021-05-28 23:20 ` [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Weihang Li @ 2021-05-21 9:29 UTC (permalink / raw)
To: dledford, jgg; +Cc: leon, linux-rdma, linuxarm, Xi Wang, Weihang Li
From: Xi Wang <wangxi11@huawei.com>
Move the HIP06 related code to the hw v1 source file for HEM.
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_device.h | 2 -
drivers/infiniband/hw/hns/hns_roce_hem.c | 82 +----------------------------
drivers/infiniband/hw/hns/hns_roce_hem.h | 9 +---
drivers/infiniband/hw/hns/hns_roce_hw_v1.c | 77 +++++++++++++++++++++++++++
drivers/infiniband/hw/hns/hns_roce_hw_v1.h | 5 ++
5 files changed, 85 insertions(+), 90 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index 02fffb7..c6cacd2 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -47,8 +47,6 @@
#define HNS_ROCE_IB_MIN_SQ_STRIDE 6
-#define HNS_ROCE_BA_SIZE (32 * 4096)
-
#define BA_BYTE_LEN 8
/* Hardware specification only for v1 engine */
diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c b/drivers/infiniband/hw/hns/hns_roce_hem.c
index fe6940b..7fdeedd 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
@@ -36,9 +36,6 @@
#include "hns_roce_hem.h"
#include "hns_roce_common.h"
-#define DMA_ADDR_T_SHIFT 12
-#define BT_BA_SHIFT 32
-
#define HEM_INDEX_BUF BIT(0)
#define HEM_INDEX_L0 BIT(1)
#define HEM_INDEX_L1 BIT(2)
@@ -337,81 +334,6 @@ void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
kfree(hem);
}
-static int hns_roce_set_hem(struct hns_roce_dev *hr_dev,
- struct hns_roce_hem_table *table, unsigned long obj)
-{
- spinlock_t *lock = &hr_dev->bt_cmd_lock;
- struct device *dev = hr_dev->dev;
- struct hns_roce_hem_iter iter;
- void __iomem *bt_cmd;
- __le32 bt_cmd_val[2];
- __le32 bt_cmd_h = 0;
- unsigned long flags;
- __le32 bt_cmd_l;
- int ret = 0;
- u64 bt_ba;
- long end;
-
- /* Find the HEM(Hardware Entry Memory) entry */
- unsigned long i = (obj & (table->num_obj - 1)) /
- (table->table_chunk_size / table->obj_size);
-
- switch (table->type) {
- case HEM_TYPE_QPC:
- case HEM_TYPE_MTPT:
- case HEM_TYPE_CQC:
- case HEM_TYPE_SRQC:
- roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
- ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, table->type);
- break;
- default:
- return ret;
- }
-
- roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
- ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
- roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
- roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
-
- /* Currently iter only a chunk */
- for (hns_roce_hem_first(table->hem[i], &iter);
- !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
- bt_ba = hns_roce_hem_addr(&iter) >> DMA_ADDR_T_SHIFT;
-
- spin_lock_irqsave(lock, flags);
-
- bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
-
- end = HW_SYNC_TIMEOUT_MSECS;
- while (end > 0) {
- if (!(readl(bt_cmd) >> BT_CMD_SYNC_SHIFT))
- break;
-
- mdelay(HW_SYNC_SLEEP_TIME_INTERVAL);
- end -= HW_SYNC_SLEEP_TIME_INTERVAL;
- }
-
- if (end <= 0) {
- dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
- spin_unlock_irqrestore(lock, flags);
- return -EBUSY;
- }
-
- bt_cmd_l = cpu_to_le32(bt_ba);
- roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
- ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
- bt_ba >> BT_BA_SHIFT);
-
- bt_cmd_val[0] = bt_cmd_l;
- bt_cmd_val[1] = bt_cmd_h;
- hns_roce_write64_k(bt_cmd_val,
- hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
- spin_unlock_irqrestore(lock, flags);
- }
-
- return ret;
-}
-
static int calc_hem_config(struct hns_roce_dev *hr_dev,
struct hns_roce_hem_table *table, unsigned long obj,
struct hns_roce_hem_mhop *mhop,
@@ -677,7 +599,7 @@ int hns_roce_table_get(struct hns_roce_dev *hr_dev,
}
/* Set HEM base address(128K/page, pa) to Hardware */
- if (hns_roce_set_hem(hr_dev, table, obj)) {
+ if (hr_dev->hw->set_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT)) {
hns_roce_free_hem(hr_dev, table->hem[i]);
table->hem[i] = NULL;
ret = -ENODEV;
@@ -782,7 +704,7 @@ void hns_roce_table_put(struct hns_roce_dev *hr_dev,
&table->mutex))
return;
- if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
+ if (hr_dev->hw->clear_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT))
dev_warn(dev, "failed to clear HEM base address.\n");
hns_roce_free_hem(hr_dev, table->hem[i]);
diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.h b/drivers/infiniband/hw/hns/hns_roce_hem.h
index ffa65e8..2d84a6b 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hem.h
+++ b/drivers/infiniband/hw/hns/hns_roce_hem.h
@@ -34,9 +34,7 @@
#ifndef _HNS_ROCE_HEM_H
#define _HNS_ROCE_HEM_H
-#define HW_SYNC_SLEEP_TIME_INTERVAL 20
-#define HW_SYNC_TIMEOUT_MSECS (25 * HW_SYNC_SLEEP_TIME_INTERVAL)
-#define BT_CMD_SYNC_SHIFT 31
+#define HEM_HOP_STEP_DIRECT 0xff
enum {
/* MAP HEM(Hardware Entry Memory) */
@@ -74,11 +72,6 @@ enum {
(type >= HEM_TYPE_MTT && hop_num == 1) || \
(type >= HEM_TYPE_MTT && hop_num == HNS_ROCE_HOP_NUM_0))
-enum {
- HNS_ROCE_HEM_PAGE_SHIFT = 12,
- HNS_ROCE_HEM_PAGE_SIZE = 1 << HNS_ROCE_HEM_PAGE_SHIFT,
-};
-
struct hns_roce_hem_chunk {
struct list_head list;
int npages;
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v1.c b/drivers/infiniband/hw/hns/hns_roce_hw_v1.c
index 620acf6..0c836cc 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v1.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v1.c
@@ -462,6 +462,82 @@ static void hns_roce_set_db_event_mode(struct hns_roce_dev *hr_dev,
roce_write(hr_dev, ROCEE_GLB_CFG_REG, val);
}
+static int hns_roce_v1_set_hem(struct hns_roce_dev *hr_dev,
+ struct hns_roce_hem_table *table, int obj,
+ int step_idx)
+{
+ spinlock_t *lock = &hr_dev->bt_cmd_lock;
+ struct device *dev = hr_dev->dev;
+ struct hns_roce_hem_iter iter;
+ void __iomem *bt_cmd;
+ __le32 bt_cmd_val[2];
+ __le32 bt_cmd_h = 0;
+ unsigned long flags;
+ __le32 bt_cmd_l;
+ int ret = 0;
+ u64 bt_ba;
+ long end;
+
+ /* Find the HEM(Hardware Entry Memory) entry */
+ unsigned long i = (obj & (table->num_obj - 1)) /
+ (table->table_chunk_size / table->obj_size);
+
+ switch (table->type) {
+ case HEM_TYPE_QPC:
+ case HEM_TYPE_MTPT:
+ case HEM_TYPE_CQC:
+ case HEM_TYPE_SRQC:
+ roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
+ ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, table->type);
+ break;
+ default:
+ return ret;
+ }
+
+ roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
+ ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
+ roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
+ roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
+
+ /* Currently iter only a chunk */
+ for (hns_roce_hem_first(table->hem[i], &iter);
+ !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
+ bt_ba = hns_roce_hem_addr(&iter) >> HNS_HW_PAGE_SHIFT;
+
+ spin_lock_irqsave(lock, flags);
+
+ bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
+
+ end = HW_SYNC_TIMEOUT_MSECS;
+ while (end > 0) {
+ if (!(readl(bt_cmd) >> BT_CMD_SYNC_SHIFT))
+ break;
+
+ mdelay(HW_SYNC_SLEEP_TIME_INTERVAL);
+ end -= HW_SYNC_SLEEP_TIME_INTERVAL;
+ }
+
+ if (end <= 0) {
+ dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
+ spin_unlock_irqrestore(lock, flags);
+ return -EBUSY;
+ }
+
+ bt_cmd_l = cpu_to_le32(bt_ba);
+ roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
+ ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
+ upper_32_bits(bt_ba));
+
+ bt_cmd_val[0] = bt_cmd_l;
+ bt_cmd_val[1] = bt_cmd_h;
+ hns_roce_write64_k(bt_cmd_val,
+ hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
+ spin_unlock_irqrestore(lock, flags);
+ }
+
+ return ret;
+}
+
static void hns_roce_set_db_ext_mode(struct hns_roce_dev *hr_dev, u32 sdb_mode,
u32 odb_mode)
{
@@ -4352,6 +4428,7 @@ static const struct hns_roce_hw hns_roce_hw_v1 = {
.set_mtu = hns_roce_v1_set_mtu,
.write_mtpt = hns_roce_v1_write_mtpt,
.write_cqc = hns_roce_v1_write_cqc,
+ .set_hem = hns_roce_v1_set_hem,
.clear_hem = hns_roce_v1_clear_hem,
.modify_qp = hns_roce_v1_modify_qp,
.dereg_mr = hns_roce_v1_dereg_mr,
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v1.h b/drivers/infiniband/hw/hns/hns_roce_hw_v1.h
index 8438323..60fdcba 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v1.h
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v1.h
@@ -1085,6 +1085,11 @@ struct hns_roce_db_table {
struct hns_roce_ext_db *ext_db;
};
+#define HW_SYNC_SLEEP_TIME_INTERVAL 20
+#define HW_SYNC_TIMEOUT_MSECS (25 * HW_SYNC_SLEEP_TIME_INTERVAL)
+#define BT_CMD_SYNC_SHIFT 31
+#define HNS_ROCE_BA_SIZE (32 * 4096)
+
struct hns_roce_bt_table {
struct hns_roce_buf_list qpc_buf;
struct hns_roce_buf_list mtpt_buf;
--
2.7.4
^ permalink raw reply related [flat|nested] 7+ messages in thread* Re: [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration
2021-05-21 9:29 [PATCH for-next 0/5] RDMA/hns: Changes about memory allocation and configuration Weihang Li
` (4 preceding siblings ...)
2021-05-21 9:29 ` [PATCH for-next 5/5] RDMA/hns: Clean the hardware related code " Weihang Li
@ 2021-05-28 23:20 ` Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Jason Gunthorpe @ 2021-05-28 23:20 UTC (permalink / raw)
To: Weihang Li; +Cc: dledford, leon, linux-rdma, linuxarm
On Fri, May 21, 2021 at 05:29:50PM +0800, Weihang Li wrote:
> These series includes several cleanups and fixes for HEM and MTR.
>
> Weihang Li (1):
> RDMA/hns: Use refcount_t APIs for HEM
>
> Xi Wang (4):
> RDMA/hns: Optimize the base address table config for MTR
> RDMA/hns: Refactor root BT allocation for MTR
> RDMA/hns: Fix wrong timer context buffer page size
> RDMA/hns: Clean the hardware related code for HEM
Applied to for-next, thanks
Jason
^ permalink raw reply [flat|nested] 7+ messages in thread