From: Wenpeng Liang <liangwenpeng@huawei.com>
To: <dledford@redhat.com>, <jgg@nvidia.com>
Cc: <linux-rdma@vger.kernel.org>, <linuxarm@huawei.com>,
<leon@kernel.org>, <liangwenpeng@huawei.com>,
Xi Wang <wangxi11@huawei.com>
Subject: [PATCH v4 for-next 03/12] RDMA/hns: Configure DCA mode for the userspace QP
Date: Thu, 29 Jul 2021 10:19:14 +0800 [thread overview]
Message-ID: <1627525163-1683-4-git-send-email-liangwenpeng@huawei.com> (raw)
In-Reply-To: <1627525163-1683-1-git-send-email-liangwenpeng@huawei.com>
From: Xi Wang <wangxi11@huawei.com>
If the userspace driver assign a NULL to the field of 'buf_addr' in
'struct hns_roce_ib_create_qp' when creating QP, this means the kernel
driver need setup the QP as DCA mode. So add a QP capability bit in
response to indicate the userspace driver that the DCA mode has been
enabled.
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
---
drivers/infiniband/hw/hns/hns_roce_dca.c | 15 ++++
drivers/infiniband/hw/hns/hns_roce_dca.h | 6 +-
drivers/infiniband/hw/hns/hns_roce_device.h | 5 ++
drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 19 ++++-
drivers/infiniband/hw/hns/hns_roce_qp.c | 106 ++++++++++++++++++++++------
include/uapi/rdma/hns-abi.h | 1 +
6 files changed, 127 insertions(+), 25 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_dca.c b/drivers/infiniband/hw/hns/hns_roce_dca.c
index 92256784a..741e009 100644
--- a/drivers/infiniband/hw/hns/hns_roce_dca.c
+++ b/drivers/infiniband/hw/hns/hns_roce_dca.c
@@ -342,6 +342,21 @@ static void free_dca_mem(struct dca_mem *mem)
spin_unlock(&mem->lock);
}
+void hns_roce_enable_dca(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
+{
+ struct hns_roce_dca_cfg *cfg = &hr_qp->dca_cfg;
+
+ cfg->buf_id = HNS_DCA_INVALID_BUF_ID;
+}
+
+void hns_roce_disable_dca(struct hns_roce_dev *hr_dev,
+ struct hns_roce_qp *hr_qp)
+{
+ struct hns_roce_dca_cfg *cfg = &hr_qp->dca_cfg;
+
+ cfg->buf_id = HNS_DCA_INVALID_BUF_ID;
+}
+
static struct hns_roce_ucontext *
uverbs_attr_to_hr_uctx(struct uverbs_attr_bundle *attrs)
{
diff --git a/drivers/infiniband/hw/hns/hns_roce_dca.h b/drivers/infiniband/hw/hns/hns_roce_dca.h
index a82ed5e..a13a2d6 100644
--- a/drivers/infiniband/hw/hns/hns_roce_dca.h
+++ b/drivers/infiniband/hw/hns/hns_roce_dca.h
@@ -14,16 +14,20 @@ struct hns_dca_page_state {
u32 head : 1; /* This page is the head in a continuous address range. */
};
+#define HNS_DCA_INVALID_BUF_ID 0UL
struct hns_dca_shrink_resp {
u64 free_key; /* free buffer's key which registered by the user */
u32 free_mems; /* free buffer count which no any QP be using */
};
-#define HNS_DCA_INVALID_BUF_ID 0UL
void hns_roce_register_udca(struct hns_roce_dev *hr_dev,
struct hns_roce_ucontext *uctx);
void hns_roce_unregister_udca(struct hns_roce_dev *hr_dev,
struct hns_roce_ucontext *uctx);
+void hns_roce_enable_dca(struct hns_roce_dev *hr_dev,
+ struct hns_roce_qp *hr_qp);
+void hns_roce_disable_dca(struct hns_roce_dev *hr_dev,
+ struct hns_roce_qp *hr_qp);
#endif
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index d18bc22..00f80b3 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -331,6 +331,10 @@ struct hns_roce_mtr {
struct hns_roce_hem_cfg hem_cfg; /* config for hardware addressing */
};
+struct hns_roce_dca_cfg {
+ u32 buf_id;
+};
+
struct hns_roce_mw {
struct ib_mw ibmw;
u32 pdn;
@@ -631,6 +635,7 @@ struct hns_roce_qp {
struct hns_roce_wq sq;
struct hns_roce_mtr mtr;
+ struct hns_roce_dca_cfg dca_cfg;
u32 buff_size;
struct mutex mutex;
diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 594d4ce..ced0c44 100644
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -345,6 +345,12 @@ static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
return 0;
}
+static bool check_dca_attach_enable(struct hns_roce_qp *hr_qp)
+{
+ return hr_qp->en_flags & HNS_ROCE_QP_CAP_DYNAMIC_CTX_ATTACH;
+}
+
+
static int check_send_valid(struct hns_roce_dev *hr_dev,
struct hns_roce_qp *hr_qp)
{
@@ -4371,6 +4377,17 @@ static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> (32 + 16 + 4));
hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
+ /* hip09 reused the IRRL_HEAD fileds in hip08 */
+ if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
+ if (check_dca_attach_enable(hr_qp)) {
+ hr_reg_enable(context, QPC_DCA_MODE);
+ hr_reg_clear(qpc_mask, QPC_DCA_MODE);
+ }
+ } else {
+ /* reset IRRL_HEAD */
+ hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
+ }
+
context->irrl_ba = cpu_to_le32(irrl_ba >> 6);
qpc_mask->irrl_ba = 0;
hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> (32 + 6));
@@ -4489,8 +4506,6 @@ static int modify_qp_rtr_to_rts(struct ib_qp *ibqp,
hr_reg_write(context, QPC_LSN, 0x100);
hr_reg_clear(qpc_mask, QPC_LSN);
- hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
-
return 0;
}
diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c
index b101b7e5..91311a0 100644
--- a/drivers/infiniband/hw/hns/hns_roce_qp.c
+++ b/drivers/infiniband/hw/hns/hns_roce_qp.c
@@ -39,6 +39,7 @@
#include "hns_roce_common.h"
#include "hns_roce_device.h"
#include "hns_roce_hem.h"
+#include "hns_roce_dca.h"
static void flush_work_handle(struct work_struct *work)
{
@@ -604,8 +605,21 @@ static int set_user_sq_size(struct hns_roce_dev *hr_dev,
return 0;
}
+static bool check_dca_is_enable(struct hns_roce_dev *hr_dev, bool is_user,
+ unsigned long addr)
+{
+ if (!(hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_DCA_MODE))
+ return false;
+
+ /* If the user QP's buffer addr is 0, the DCA mode should be enabled */
+ if (is_user)
+ return !addr;
+
+ return false;
+}
+
static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
- struct hns_roce_qp *hr_qp,
+ struct hns_roce_qp *hr_qp, bool dca_en,
struct hns_roce_buf_attr *buf_attr)
{
int buf_size;
@@ -649,9 +663,21 @@ static int set_wqe_buf_attr(struct hns_roce_dev *hr_dev,
if (hr_qp->buff_size < 1)
return -EINVAL;
- buf_attr->page_shift = HNS_HW_PAGE_SHIFT + hr_dev->caps.mtt_buf_pg_sz;
buf_attr->region_count = idx;
+ if (dca_en) {
+ /*
+ * When enable DCA, there's no need to alloc buffer now, and
+ * the page shift should be fixed to 4K.
+ */
+ buf_attr->mtt_only = true;
+ buf_attr->page_shift = HNS_HW_PAGE_SHIFT;
+ } else {
+ buf_attr->mtt_only = false;
+ buf_attr->page_shift = HNS_HW_PAGE_SHIFT +
+ hr_dev->caps.mtt_buf_pg_sz;
+ }
+
return 0;
}
@@ -748,12 +774,48 @@ static void free_rq_inline_buf(struct hns_roce_qp *hr_qp)
kfree(hr_qp->rq_inl_buf.wqe_list);
}
-static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
+static int alloc_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
+ bool dca_en, struct hns_roce_buf_attr *buf_attr,
+ struct ib_udata *udata, unsigned long addr)
+{
+ struct ib_device *ibdev = &hr_dev->ib_dev;
+ int ret;
+
+ if (dca_en) {
+ /* DCA must be enabled after the buffer attr is configured. */
+ hns_roce_enable_dca(hr_dev, hr_qp);
+
+ hr_qp->en_flags |= HNS_ROCE_QP_CAP_DYNAMIC_CTX_ATTACH;
+ }
+
+ ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, buf_attr,
+ HNS_HW_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);
+ if (dca_en)
+ hns_roce_disable_dca(hr_dev, hr_qp);
+ }
+
+ return ret;
+}
+
+static void free_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
+ struct ib_udata *udata)
+{
+ hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
+
+ if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DYNAMIC_CTX_ATTACH)
+ hns_roce_disable_dca(hr_dev, hr_qp);
+}
+
+static int alloc_qp_wqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
struct ib_qp_init_attr *init_attr,
struct ib_udata *udata, unsigned long addr)
{
struct ib_device *ibdev = &hr_dev->ib_dev;
struct hns_roce_buf_attr buf_attr = {};
+ bool dca_en;
int ret;
if (!udata && hr_qp->rq_inl_buf.wqe_cnt) {
@@ -768,16 +830,16 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
hr_qp->rq_inl_buf.wqe_list = NULL;
}
- ret = set_wqe_buf_attr(hr_dev, hr_qp, &buf_attr);
+ dca_en = check_dca_is_enable(hr_dev, !!udata, addr);
+ ret = set_wqe_buf_attr(hr_dev, hr_qp, dca_en, &buf_attr);
if (ret) {
- ibdev_err(ibdev, "failed to split WQE buf, ret = %d.\n", ret);
+ ibdev_err(ibdev, "failed to set WQE attr, ret = %d.\n", ret);
goto err_inline;
}
- ret = hns_roce_mtr_create(hr_dev, &hr_qp->mtr, &buf_attr,
- PAGE_SHIFT + hr_dev->caps.mtt_ba_pg_sz,
- udata, addr);
+
+ ret = alloc_wqe_buf(hr_dev, hr_qp, dca_en, &buf_attr, udata, addr);
if (ret) {
- ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret);
+ ibdev_err(ibdev, "failed to alloc WQE buf, ret = %d.\n", ret);
goto err_inline;
}
@@ -788,9 +850,10 @@ static int alloc_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
return ret;
}
-static void free_qp_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp)
+static void free_qp_wqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
+ struct ib_udata *udata)
{
- hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr);
+ free_wqe_buf(hr_dev, hr_qp, udata);
free_rq_inline_buf(hr_qp);
}
@@ -848,7 +911,6 @@ static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
goto err_out;
}
hr_qp->en_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
- resp->cap_flags |= HNS_ROCE_QP_CAP_SQ_RECORD_DB;
}
if (user_qp_has_rdb(hr_dev, init_attr, udata, resp)) {
@@ -861,7 +923,6 @@ static int alloc_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
goto err_sdb;
}
hr_qp->en_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
- resp->cap_flags |= HNS_ROCE_QP_CAP_RQ_RECORD_DB;
}
} else {
if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
@@ -1040,18 +1101,18 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
}
}
- ret = alloc_qp_buf(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
- if (ret) {
- ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
- goto err_buf;
- }
-
ret = alloc_qpn(hr_dev, hr_qp);
if (ret) {
ibdev_err(ibdev, "failed to alloc QPN, ret = %d.\n", ret);
goto err_qpn;
}
+ ret = alloc_qp_wqe(hr_dev, hr_qp, init_attr, udata, ucmd.buf_addr);
+ if (ret) {
+ ibdev_err(ibdev, "failed to alloc QP buffer, ret = %d.\n", ret);
+ goto err_buf;
+ }
+
ret = alloc_qp_db(hr_dev, hr_qp, init_attr, udata, &ucmd, &resp);
if (ret) {
ibdev_err(ibdev, "failed to alloc QP doorbell, ret = %d.\n",
@@ -1073,6 +1134,7 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
}
if (udata) {
+ resp.cap_flags = hr_qp->en_flags;
ret = ib_copy_to_udata(udata, &resp,
min(udata->outlen, sizeof(resp)));
if (ret) {
@@ -1101,10 +1163,10 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
err_qpc:
free_qp_db(hr_dev, hr_qp, udata);
err_db:
+ free_qp_wqe(hr_dev, hr_qp, udata);
+err_buf:
free_qpn(hr_dev, hr_qp);
err_qpn:
- free_qp_buf(hr_dev, hr_qp);
-err_buf:
free_kernel_wrid(hr_qp);
return ret;
}
@@ -1118,7 +1180,7 @@ void hns_roce_qp_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,
free_qpc(hr_dev, hr_qp);
free_qpn(hr_dev, hr_qp);
- free_qp_buf(hr_dev, hr_qp);
+ free_qp_wqe(hr_dev, hr_qp, udata);
free_kernel_wrid(hr_qp);
free_qp_db(hr_dev, hr_qp, udata);
diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h
index bcca5be..4452b17 100644
--- a/include/uapi/rdma/hns-abi.h
+++ b/include/uapi/rdma/hns-abi.h
@@ -77,6 +77,7 @@ enum hns_roce_qp_cap_flags {
HNS_ROCE_QP_CAP_RQ_RECORD_DB = 1 << 0,
HNS_ROCE_QP_CAP_SQ_RECORD_DB = 1 << 1,
HNS_ROCE_QP_CAP_OWNER_DB = 1 << 2,
+ HNS_ROCE_QP_CAP_DYNAMIC_CTX_ATTACH = 1 << 4,
};
struct hns_roce_ib_create_qp_resp {
--
2.8.1
next prev parent reply other threads:[~2021-07-29 2:23 UTC|newest]
Thread overview: 22+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-07-29 2:19 [PATCH v4 for-next 00/12] RDMA/hns: Add support for Dynamic Context Attachment Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 01/12] RDMA/hns: Introduce DCA for RC QP Wenpeng Liang
2021-08-19 23:54 ` Jason Gunthorpe
2021-08-21 9:37 ` Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 02/12] RDMA/hns: Add method for shrinking DCA memory pool Wenpeng Liang
2021-07-29 2:19 ` Wenpeng Liang [this message]
2021-07-29 2:19 ` [PATCH v4 for-next 04/12] RDMA/hns: Refactor QP modify flow Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 05/12] RDMA/hns: Add method for attaching WQE buffer Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 06/12] RDMA/hns: Setup the configuration of WQE addressing to QPC Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 07/12] RDMA/hns: Add method to detach WQE buffer Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 08/12] RDMA/hns: Add method to query WQE buffer's address Wenpeng Liang
2021-08-19 23:58 ` Jason Gunthorpe
2021-08-21 9:39 ` Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 09/12] RDMA/hns: Add a shared memory to sync DCA status Wenpeng Liang
2021-08-19 23:43 ` Jason Gunthorpe
2021-08-21 9:42 ` Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 10/12] RDMA/hns: Sync DCA status by the shared memory Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 11/12] RDMA/nldev: Add detailed CTX information support Wenpeng Liang
2021-07-29 2:19 ` [PATCH v4 for-next 12/12] RDMA/hns: Dump detailed driver-specific UCTX Wenpeng Liang
2021-08-12 12:42 ` [PATCH v4 for-next 00/12] RDMA/hns: Add support for Dynamic Context Attachment Wenpeng Liang
2021-08-21 9:34 ` Wenpeng Liang
2021-08-22 22:25 ` Jason Gunthorpe
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1627525163-1683-4-git-send-email-liangwenpeng@huawei.com \
--to=liangwenpeng@huawei.com \
--cc=dledford@redhat.com \
--cc=jgg@nvidia.com \
--cc=leon@kernel.org \
--cc=linux-rdma@vger.kernel.org \
--cc=linuxarm@huawei.com \
--cc=wangxi11@huawei.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).