* [PATCH v3 for-rc 1/2] RDMA/hns: Fix ext_sge num error when post send
2022-11-08 13:38 [PATCH v3 for-rc 0/2] Fix sge_num bug Haoyue Xu
@ 2022-11-08 13:38 ` Haoyue Xu
2022-11-08 13:38 ` [PATCH v3 for-rc 2/2] RDMA/hns: Fix the problem of sge nums Haoyue Xu
` (2 subsequent siblings)
3 siblings, 0 replies; 8+ messages in thread
From: Haoyue Xu @ 2022-11-08 13:38 UTC (permalink / raw)
To: jgg, leon; +Cc: linux-rdma, linuxarm, linux-kernel, xuhaoyue1
From: Luoyouming <luoyouming@huawei.com>
In the HNS ROCE driver, The sge is divided into standard sge and extended sge.
There are 2 standard sge in RC/XRC, and the UD standard sge is 0.
In the scenario of RC SQ inline, if the data does not exceed 32bytes,
the standard sge will be used. If it exceeds, only the extended sge will
be used to fill the data. Currently, when filling the extended sge,
max_gs is directly used as the number of the extended sge,
which did not subtract the number of standard sge.
There is a logical error. The new algorithm subtracts the number of
standard sge from max_gs to get the actual number of extended sge.
Fixes: 30b707886aeb ("RDMA/hns: Support inline data in extented sge space for RC")
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
---
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 1435fe2ea176..0937db738be7 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -187,20 +187,29 @@ static void set_atomic_seg(const struct ib_send_wr *wr,
hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
}
+static unsigned int get_std_sge_num(struct hns_roce_qp *qp)
+{
+ if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
+ return 0;
+
+ return HNS_ROCE_SGE_IN_WQE;
+}
+
static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
const struct ib_send_wr *wr,
unsigned int *sge_idx, u32 msg_len)
{
struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
- unsigned int ext_sge_sz = qp->sq.max_gs * HNS_ROCE_SGE_SIZE;
unsigned int left_len_in_pg;
unsigned int idx = *sge_idx;
+ unsigned int std_sge_num;
unsigned int i = 0;
unsigned int len;
void *addr;
void *dseg;
- if (msg_len > ext_sge_sz) {
+ std_sge_num = get_std_sge_num(qp);
+ if (msg_len > (qp->sq.max_gs - std_sge_num) * HNS_ROCE_SGE_SIZE) {
ibdev_err(ibdev,
"no enough extended sge space for inline data.\n");
return -EINVAL;
--
2.30.0
^ permalink raw reply related [flat|nested] 8+ messages in thread* [PATCH v3 for-rc 2/2] RDMA/hns: Fix the problem of sge nums
2022-11-08 13:38 [PATCH v3 for-rc 0/2] Fix sge_num bug Haoyue Xu
2022-11-08 13:38 ` [PATCH v3 for-rc 1/2] RDMA/hns: Fix ext_sge num error when post send Haoyue Xu
@ 2022-11-08 13:38 ` Haoyue Xu
2022-11-09 18:07 ` [PATCH v3 for-rc 0/2] Fix sge_num bug Leon Romanovsky
2022-11-19 0:23 ` Jason Gunthorpe
3 siblings, 0 replies; 8+ messages in thread
From: Haoyue Xu @ 2022-11-08 13:38 UTC (permalink / raw)
To: jgg, leon; +Cc: linux-rdma, linuxarm, linux-kernel, xuhaoyue1
From: Luoyouming <luoyouming@huawei.com>
The user usually configures the number of sge through the max_send_sge
parameter when creating qp, and configures the maximum size of inline data
that can be sent through max_inline_data. Inline uses sge to fill data to
send. Expect the following:
1) When the sge space cannot hold inline data, the sge space needs to be
expanded to accommodate all inline data
2) When the sge space is enough to accommodate inline data, the upper limit
of inline data can be increased so that users can send larger inline data
Currently case one is not implemented. When the inline data is larger than
the sge space, an error of insufficient sge space occurs.
This part of the code needs to be reimplemented according to the expected
rules. The calculation method of sge num is modified to take the maximum
value of max_send_sge and the sge for max_inline_data to solve this problem.
Fixes: 05201e01be93 ("RDMA/hns: Refactor process of setting extended sge")
Fixes: 30b707886aeb ("RDMA/hns: Support inline data in extented sge space for RC")
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
---
drivers/infiniband/hw/hns/hns_roce_device.h | 3 +
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 12 +--
drivers/infiniband/hw/hns/hns_roce_main.c | 18 +++-
drivers/infiniband/hw/hns/hns_roce_qp.c | 107 ++++++++++++++++----
include/uapi/rdma/hns-abi.h | 15 +++
5 files changed, 125 insertions(+), 30 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index 723e55a7de8d..f701cc86896b 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -202,6 +202,7 @@ struct hns_roce_ucontext {
struct list_head page_list;
struct mutex page_mutex;
struct hns_user_mmap_entry *db_mmap_entry;
+ u32 config;
};
struct hns_roce_pd {
@@ -334,6 +335,7 @@ struct hns_roce_wq {
u32 head;
u32 tail;
void __iomem *db_reg;
+ u32 ext_sge_cnt;
};
struct hns_roce_sge {
@@ -635,6 +637,7 @@ struct hns_roce_qp {
struct list_head rq_node; /* all recv qps are on a list */
struct list_head sq_node; /* all send qps are on a list */
struct hns_user_mmap_entry *dwqe_mmap_entry;
+ u32 config;
};
struct hns_roce_ib_iboe {
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 0937db738be7..65875b4cff13 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -187,14 +187,6 @@ static void set_atomic_seg(const struct ib_send_wr *wr,
hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
}
-static unsigned int get_std_sge_num(struct hns_roce_qp *qp)
-{
- if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
- return 0;
-
- return HNS_ROCE_SGE_IN_WQE;
-}
-
static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
const struct ib_send_wr *wr,
unsigned int *sge_idx, u32 msg_len)
@@ -202,14 +194,12 @@ static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
unsigned int left_len_in_pg;
unsigned int idx = *sge_idx;
- unsigned int std_sge_num;
unsigned int i = 0;
unsigned int len;
void *addr;
void *dseg;
- std_sge_num = get_std_sge_num(qp);
- if (msg_len > (qp->sq.max_gs - std_sge_num) * HNS_ROCE_SGE_SIZE) {
+ if (msg_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE) {
ibdev_err(ibdev,
"no enough extended sge space for inline data.\n");
return -EINVAL;
diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
index dcf89689a4c6..8ba68ac12388 100644
--- a/drivers/infiniband/hw/hns/hns_roce_main.c
+++ b/drivers/infiniband/hw/hns/hns_roce_main.c
@@ -354,10 +354,11 @@ static int hns_roce_alloc_uar_entry(struct ib_ucontext *uctx)
static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
struct ib_udata *udata)
{
- int ret;
struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
- struct hns_roce_ib_alloc_ucontext_resp resp = {};
struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
+ struct hns_roce_ib_alloc_ucontext_resp resp = {};
+ struct hns_roce_ib_alloc_ucontext ucmd = {};
+ int ret;
if (!hr_dev->active)
return -EAGAIN;
@@ -365,6 +366,19 @@ static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx,
resp.qp_tab_size = hr_dev->caps.num_qps;
resp.srq_tab_size = hr_dev->caps.num_srqs;
+ ret = ib_copy_from_udata(&ucmd, udata,
+ min(udata->inlen, sizeof(ucmd)));
+ if (ret)
+ return ret;
+
+ if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
+ context->config = ucmd.config & HNS_ROCE_EXSGE_FLAGS;
+
+ if (context->config & HNS_ROCE_EXSGE_FLAGS) {
+ resp.config |= HNS_ROCE_RSP_EXSGE_FLAGS;
+ resp.max_inline_data = hr_dev->caps.max_sq_inline;
+ }
+
ret = hns_roce_uar_alloc(hr_dev, &context->uar);
if (ret)
goto error_fail_uar_alloc;
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index f0bd82a18069..0ae335fb205c 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -476,38 +476,109 @@ static int set_rq_size(struct hns_roce_dev *hr_dev, struct ib_qp_cap *cap,
return 0;
}
-static u32 get_wqe_ext_sge_cnt(struct hns_roce_qp *qp)
+static u32 get_max_inline_data(struct hns_roce_dev *hr_dev,
+ struct ib_qp_cap *cap)
{
- /* GSI/UD QP only has extended sge */
- if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_UD)
- return qp->sq.max_gs;
-
- if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE)
- return qp->sq.max_gs - HNS_ROCE_SGE_IN_WQE;
+ if (cap->max_inline_data) {
+ cap->max_inline_data = roundup_pow_of_two(cap->max_inline_data);
+ return min(cap->max_inline_data,
+ hr_dev->caps.max_sq_inline);
+ }
return 0;
}
+static void update_inline_data(struct hns_roce_qp *hr_qp,
+ struct ib_qp_cap *cap)
+{
+ u32 sge_num = hr_qp->sq.ext_sge_cnt;
+
+ if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
+ if (!(hr_qp->ibqp.qp_type == IB_QPT_GSI ||
+ hr_qp->ibqp.qp_type == IB_QPT_UD))
+ sge_num = max((u32)HNS_ROCE_SGE_IN_WQE, sge_num);
+
+ cap->max_inline_data = max(cap->max_inline_data,
+ sge_num * HNS_ROCE_SGE_SIZE);
+ }
+
+ hr_qp->max_inline_data = cap->max_inline_data;
+}
+
+static u32 get_sge_num_from_max_send_sge(bool is_ud_or_gsi,
+ u32 max_send_sge)
+{
+ unsigned int std_sge_num;
+ unsigned int min_sge;
+
+ std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
+ min_sge = is_ud_or_gsi ? 1 : 0;
+ return max_send_sge > std_sge_num ? (max_send_sge - std_sge_num) :
+ min_sge;
+}
+
+static unsigned int get_sge_num_from_max_inl_data(bool is_ud_or_gsi,
+ u32 max_inline_data)
+{
+ unsigned int inline_sge;
+
+ inline_sge = roundup_pow_of_two(max_inline_data) / HNS_ROCE_SGE_SIZE;
+
+ /*
+ * if max_inline_data less than
+ * HNS_ROCE_SGE_IN_WQE * HNS_ROCE_SGE_SIZE,
+ * In addition to ud's mode, no need to extend sge.
+ */
+ if (!is_ud_or_gsi && inline_sge <= HNS_ROCE_SGE_IN_WQE)
+ inline_sge = 0;
+
+ return inline_sge;
+}
+
static void set_ext_sge_param(struct hns_roce_dev *hr_dev, u32 sq_wqe_cnt,
struct hns_roce_qp *hr_qp, struct ib_qp_cap *cap)
{
+ bool is_ud_or_gsi = (hr_qp->ibqp.qp_type == IB_QPT_GSI ||
+ hr_qp->ibqp.qp_type == IB_QPT_UD);
+ unsigned int std_sge_num;
+ u32 inline_ext_sge = 0;
+ u32 ext_wqe_sge_cnt;
u32 total_sge_cnt;
- u32 wqe_sge_cnt;
+
+ cap->max_inline_data = get_max_inline_data(hr_dev, cap);
hr_qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT;
+ std_sge_num = is_ud_or_gsi ? 0 : HNS_ROCE_SGE_IN_WQE;
+ ext_wqe_sge_cnt = get_sge_num_from_max_send_sge(is_ud_or_gsi,
+ cap->max_send_sge);
- hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
+ if (hr_qp->config & HNS_ROCE_EXSGE_FLAGS) {
+ inline_ext_sge = max(ext_wqe_sge_cnt,
+ get_sge_num_from_max_inl_data(is_ud_or_gsi,
+ cap->max_inline_data));
+ hr_qp->sq.ext_sge_cnt = inline_ext_sge ?
+ roundup_pow_of_two(inline_ext_sge) : 0;
- wqe_sge_cnt = get_wqe_ext_sge_cnt(hr_qp);
+ hr_qp->sq.max_gs = max(1U, (hr_qp->sq.ext_sge_cnt + std_sge_num));
+ hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
+
+ ext_wqe_sge_cnt = hr_qp->sq.ext_sge_cnt;
+ } else {
+ hr_qp->sq.max_gs = max(1U, cap->max_send_sge);
+ hr_qp->sq.max_gs = min(hr_qp->sq.max_gs, hr_dev->caps.max_sq_sg);
+ hr_qp->sq.ext_sge_cnt = hr_qp->sq.max_gs;
+ }
/* If the number of extended sge is not zero, they MUST use the
* space of HNS_HW_PAGE_SIZE at least.
*/
- if (wqe_sge_cnt) {
- total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * wqe_sge_cnt);
+ if (ext_wqe_sge_cnt) {
+ total_sge_cnt = roundup_pow_of_two(sq_wqe_cnt * ext_wqe_sge_cnt);
hr_qp->sge.sge_cnt = max(total_sge_cnt,
(u32)HNS_HW_PAGE_SIZE / HNS_ROCE_SGE_SIZE);
}
+
+ update_inline_data(hr_qp, cap);
}
static int check_sq_size_with_integrity(struct hns_roce_dev *hr_dev,
@@ -556,6 +627,7 @@ static int set_user_sq_size(struct hns_roce_dev *hr_dev,
hr_qp->sq.wqe_shift = ucmd->log_sq_stride;
hr_qp->sq.wqe_cnt = cnt;
+ cap->max_send_sge = hr_qp->sq.max_gs;
return 0;
}
@@ -986,13 +1058,9 @@ static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
struct hns_roce_ib_create_qp *ucmd)
{
struct ib_device *ibdev = &hr_dev->ib_dev;
+ struct hns_roce_ucontext *uctx;
int ret;
- if (init_attr->cap.max_inline_data > hr_dev->caps.max_sq_inline)
- init_attr->cap.max_inline_data = hr_dev->caps.max_sq_inline;
-
- hr_qp->max_inline_data = init_attr->cap.max_inline_data;
-
if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
hr_qp->sq_signal_bits = IB_SIGNAL_ALL_WR;
else
@@ -1015,12 +1083,17 @@ static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
return ret;
}
+ uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext,
+ ibucontext);
+ hr_qp->config = uctx->config;
ret = set_user_sq_size(hr_dev, &init_attr->cap, hr_qp, ucmd);
if (ret)
ibdev_err(ibdev,
"failed to set user SQ size, ret = %d.\n",
ret);
} else {
+ if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
+ hr_qp->config = HNS_ROCE_EXSGE_FLAGS;
ret = set_kernel_sq_size(hr_dev, &init_attr->cap, hr_qp);
if (ret)
ibdev_err(ibdev,
diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h
index f6fde06db4b4..745790ce3c26 100644
--- a/include/uapi/rdma/hns-abi.h
+++ b/include/uapi/rdma/hns-abi.h
@@ -85,11 +85,26 @@ struct hns_roce_ib_create_qp_resp {
__aligned_u64 dwqe_mmap_key;
};
+enum {
+ HNS_ROCE_EXSGE_FLAGS = 1 << 0,
+};
+
+enum {
+ HNS_ROCE_RSP_EXSGE_FLAGS = 1 << 0,
+};
+
struct hns_roce_ib_alloc_ucontext_resp {
__u32 qp_tab_size;
__u32 cqe_size;
__u32 srq_tab_size;
__u32 reserved;
+ __u32 config;
+ __u32 max_inline_data;
+};
+
+struct hns_roce_ib_alloc_ucontext {
+ __u32 config;
+ __u32 reserved;
};
struct hns_roce_ib_alloc_pd_resp {
--
2.30.0
^ permalink raw reply related [flat|nested] 8+ messages in thread