* [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs
@ 2025-12-24 4:25 Sriharsha Basavapatna
2025-12-24 4:25 ` [PATCH rdma-next v6 1/4] RDMA/bnxt_re: Move the UAPI methods to a dedicated file Sriharsha Basavapatna
` (3 more replies)
0 siblings, 4 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2025-12-24 4:25 UTC (permalink / raw)
To: leon, jgg
Cc: linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
Hi,
This patchset supports Direct Verbs in the bnxt_re driver.
This is required by vendor specific applications that need to manage
the HW resources directly and to implement the datapath in the
application.
To support this, the library and the driver are being enhanced to
provide Direct Verbs using which the application can allocate and
manage the HW resources (Queues, Doorbell etc) . The Direct Verbs
enable the application to implement the control path.
Patch#1 Move uapi methods to a separate file
Patch#2 Refactor existing bnxt_qplib_create_qp() function
Patch#3 Support dbr direct verbs
Patch#4 Support cq and qp direct verbs
Thanks,
-Harsha
******
Changes:
v6:
- Minor updates in Patch#3:
- Removed unused variables.
- Renamed & updated a uverbs method to a global.
- Minor updates in Patch#4:
- Removed unused variables, stray hunks.
v5:
- Design changes to address previous round of comments:
- Reverted changes in rdma-core (removed V4-Patch#1).
- Removed driver support for umem-reg/dereg DVs (Patch#3).
- Enhanced driver specific udata to avoid new CQ/QP ioctls (Patch#4).
- Removed additional driver functions in modify/query QP (Patch#4).
- Utilized queue-va in udata for deferred pinning (Patch#4).
v4:
- Added a new (rdma core) patch.
- Addressed code review comments in patch 5.
v3:
- Addressed code review comments in patches 1, 2 and 4.
v2:
- Fixed build warnings reported by test robot in patches 3 and 4.
v5: https://lore.kernel.org/linux-rdma/20251129165441.75274-1-sriharsha.basavapatna@broadcom.com/
v4: https://lore.kernel.org/linux-rdma/20251117061741.15752-1-sriharsha.basavapatna@broadcom.com/
v3: https://lore.kernel.org/linux-rdma/20251110145628.290296-1-sriharsha.basavapatna@broadcom.com/
v2: https://lore.kernel.org/linux-rdma/20251104072320.210596-1-sriharsha.basavapatna@broadcom.com/
v1: https://lore.kernel.org/linux-rdma/20251103105033.205586-1-sriharsha.basavapatna@broadcom.com/
******
Kalesh AP (3):
RDMA/bnxt_re: Move the UAPI methods to a dedicated file
RDMA/bnxt_re: Refactor bnxt_qplib_create_qp() function
RDMA/bnxt_re: Direct Verbs: Support DBR verbs
Sriharsha Basavapatna (1):
RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
drivers/infiniband/hw/bnxt_re/Makefile | 2 +-
drivers/infiniband/hw/bnxt_re/bnxt_re.h | 7 +
drivers/infiniband/hw/bnxt_re/dv.c | 939 ++++++++++++++++++++++
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 552 +++++--------
drivers/infiniband/hw/bnxt_re/ib_verbs.h | 21 +
drivers/infiniband/hw/bnxt_re/main.c | 1 +
drivers/infiniband/hw/bnxt_re/qplib_fp.c | 310 +++----
drivers/infiniband/hw/bnxt_re/qplib_fp.h | 10 +-
drivers/infiniband/hw/bnxt_re/qplib_res.c | 43 +
drivers/infiniband/hw/bnxt_re/qplib_res.h | 10 +
include/uapi/rdma/bnxt_re-abi.h | 45 ++
11 files changed, 1395 insertions(+), 545 deletions(-)
create mode 100644 drivers/infiniband/hw/bnxt_re/dv.c
--
2.51.2.636.ga99f379adf
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH rdma-next v6 1/4] RDMA/bnxt_re: Move the UAPI methods to a dedicated file
2025-12-24 4:25 [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs Sriharsha Basavapatna
@ 2025-12-24 4:25 ` Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 2/4] RDMA/bnxt_re: Refactor bnxt_qplib_create_qp() function Sriharsha Basavapatna
` (2 subsequent siblings)
3 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2025-12-24 4:25 UTC (permalink / raw)
To: leon, jgg
Cc: linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
From: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
This is in preparation for upcoming patches in the series.
Driver has to support additional UAPIs for Direct verbs.
Moving current UAPI implementation to a new file, dv.c.
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Reviewed-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
drivers/infiniband/hw/bnxt_re/Makefile | 2 +-
drivers/infiniband/hw/bnxt_re/dv.c | 339 +++++++++++++++++++++++
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 305 +-------------------
drivers/infiniband/hw/bnxt_re/ib_verbs.h | 3 +
4 files changed, 344 insertions(+), 305 deletions(-)
create mode 100644 drivers/infiniband/hw/bnxt_re/dv.c
diff --git a/drivers/infiniband/hw/bnxt_re/Makefile b/drivers/infiniband/hw/bnxt_re/Makefile
index f63417d2ccc6..b82d12df6269 100644
--- a/drivers/infiniband/hw/bnxt_re/Makefile
+++ b/drivers/infiniband/hw/bnxt_re/Makefile
@@ -5,4 +5,4 @@ obj-$(CONFIG_INFINIBAND_BNXT_RE) += bnxt_re.o
bnxt_re-y := main.o ib_verbs.o \
qplib_res.o qplib_rcfw.o \
qplib_sp.o qplib_fp.o hw_counters.o \
- debugfs.o
+ debugfs.o dv.o
diff --git a/drivers/infiniband/hw/bnxt_re/dv.c b/drivers/infiniband/hw/bnxt_re/dv.c
new file mode 100644
index 000000000000..5655c6176af4
--- /dev/null
+++ b/drivers/infiniband/hw/bnxt_re/dv.c
@@ -0,0 +1,339 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
+/*
+ * Copyright (c) 2025, Broadcom. All rights reserved. The term
+ * Broadcom refers to Broadcom Limited and/or its subsidiaries.
+ *
+ * Description: Direct Verbs interpreter
+ */
+
+#include <rdma/ib_addr.h>
+#include <rdma/uverbs_types.h>
+#include <rdma/uverbs_std_types.h>
+#include <rdma/ib_user_ioctl_cmds.h>
+#define UVERBS_MODULE_NAME bnxt_re
+#include <rdma/uverbs_named_ioctl.h>
+#include <rdma/bnxt_re-abi.h>
+
+#include "roce_hsi.h"
+#include "qplib_res.h"
+#include "qplib_sp.h"
+#include "qplib_fp.h"
+#include "qplib_rcfw.h"
+#include "bnxt_re.h"
+#include "ib_verbs.h"
+
+static struct bnxt_re_cq *bnxt_re_search_for_cq(struct bnxt_re_dev *rdev, u32 cq_id)
+{
+ struct bnxt_re_cq *cq = NULL, *tmp_cq;
+
+ hash_for_each_possible(rdev->cq_hash, tmp_cq, hash_entry, cq_id) {
+ if (tmp_cq->qplib_cq.id == cq_id) {
+ cq = tmp_cq;
+ break;
+ }
+ }
+ return cq;
+}
+
+static struct bnxt_re_srq *bnxt_re_search_for_srq(struct bnxt_re_dev *rdev, u32 srq_id)
+{
+ struct bnxt_re_srq *srq = NULL, *tmp_srq;
+
+ hash_for_each_possible(rdev->srq_hash, tmp_srq, hash_entry, srq_id) {
+ if (tmp_srq->qplib_srq.id == srq_id) {
+ srq = tmp_srq;
+ break;
+ }
+ }
+ return srq;
+}
+
+static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_ucontext *uctx;
+ struct ib_ucontext *ib_uctx;
+
+ ib_uctx = ib_uverbs_get_ucontext(attrs);
+ if (IS_ERR(ib_uctx))
+ return PTR_ERR(ib_uctx);
+
+ uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
+ if (IS_ERR(uctx))
+ return PTR_ERR(uctx);
+
+ bnxt_re_pacing_alert(uctx->rdev);
+ return 0;
+}
+
+static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs)
+{
+ struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
+ enum bnxt_re_alloc_page_type alloc_type;
+ struct bnxt_re_user_mmap_entry *entry;
+ enum bnxt_re_mmap_flag mmap_flag;
+ struct bnxt_qplib_chip_ctx *cctx;
+ struct bnxt_re_ucontext *uctx;
+ struct ib_ucontext *ib_uctx;
+ struct bnxt_re_dev *rdev;
+ u64 mmap_offset;
+ u32 length;
+ u32 dpi;
+ u64 addr;
+ int err;
+
+ ib_uctx = ib_uverbs_get_ucontext(attrs);
+ if (IS_ERR(ib_uctx))
+ return PTR_ERR(ib_uctx);
+
+ uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
+ if (IS_ERR(uctx))
+ return PTR_ERR(uctx);
+
+ err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE);
+ if (err)
+ return err;
+
+ rdev = uctx->rdev;
+ cctx = rdev->chip_ctx;
+
+ switch (alloc_type) {
+ case BNXT_RE_ALLOC_WC_PAGE:
+ if (cctx->modes.db_push) {
+ if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi,
+ uctx, BNXT_QPLIB_DPI_TYPE_WC))
+ return -ENOMEM;
+ length = PAGE_SIZE;
+ dpi = uctx->wcdpi.dpi;
+ addr = (u64)uctx->wcdpi.umdbr;
+ mmap_flag = BNXT_RE_MMAP_WC_DB;
+ } else {
+ return -EINVAL;
+ }
+
+ break;
+ case BNXT_RE_ALLOC_DBR_BAR_PAGE:
+ length = PAGE_SIZE;
+ addr = (u64)rdev->pacing.dbr_bar_addr;
+ mmap_flag = BNXT_RE_MMAP_DBR_BAR;
+ break;
+
+ case BNXT_RE_ALLOC_DBR_PAGE:
+ length = PAGE_SIZE;
+ addr = (u64)rdev->pacing.dbr_page;
+ mmap_flag = BNXT_RE_MMAP_DBR_PAGE;
+ break;
+
+ default:
+ return -EOPNOTSUPP;
+ }
+
+ entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mmap_offset);
+ if (!entry)
+ return -ENOMEM;
+
+ uobj->object = entry;
+ uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
+ err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
+ &mmap_offset, sizeof(mmap_offset));
+ if (err)
+ return err;
+
+ err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
+ &length, sizeof(length));
+ if (err)
+ return err;
+
+ err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI,
+ &dpi, sizeof(dpi));
+ if (err)
+ return err;
+
+ return 0;
+}
+
+static int alloc_page_obj_cleanup(struct ib_uobject *uobject,
+ enum rdma_remove_reason why,
+ struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_user_mmap_entry *entry = uobject->object;
+ struct bnxt_re_ucontext *uctx = entry->uctx;
+
+ switch (entry->mmap_flag) {
+ case BNXT_RE_MMAP_WC_DB:
+ if (uctx && uctx->wcdpi.dbr) {
+ struct bnxt_re_dev *rdev = uctx->rdev;
+
+ bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
+ uctx->wcdpi.dbr = NULL;
+ }
+ break;
+ case BNXT_RE_MMAP_DBR_BAR:
+ case BNXT_RE_MMAP_DBR_PAGE:
+ break;
+ default:
+ goto exit;
+ }
+ rdma_user_mmap_entry_remove(&entry->rdma_entry);
+exit:
+ return 0;
+}
+
+DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE,
+ UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE,
+ BNXT_RE_OBJECT_ALLOC_PAGE,
+ UVERBS_ACCESS_NEW,
+ UA_MANDATORY),
+ UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE,
+ enum bnxt_re_alloc_page_type,
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
+ UVERBS_ATTR_TYPE(u64),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
+ UVERBS_ATTR_TYPE(u32),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI,
+ UVERBS_ATTR_TYPE(u32),
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE,
+ UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE,
+ BNXT_RE_OBJECT_ALLOC_PAGE,
+ UVERBS_ACCESS_DESTROY,
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE,
+ UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup),
+ &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE),
+ &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE));
+
+DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_NOTIFY_DRV);
+
+DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_NOTIFY_DRV,
+ &UVERBS_METHOD(BNXT_RE_METHOD_NOTIFY_DRV));
+
+/* Toggle MEM */
+static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bundle *attrs)
+{
+ struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_TOGGLE_MEM_HANDLE);
+ enum bnxt_re_mmap_flag mmap_flag = BNXT_RE_MMAP_TOGGLE_PAGE;
+ enum bnxt_re_get_toggle_mem_type res_type;
+ struct bnxt_re_user_mmap_entry *entry;
+ struct bnxt_re_ucontext *uctx;
+ struct ib_ucontext *ib_uctx;
+ struct bnxt_re_dev *rdev;
+ struct bnxt_re_srq *srq;
+ u32 length = PAGE_SIZE;
+ struct bnxt_re_cq *cq;
+ u64 mem_offset;
+ u32 offset = 0;
+ u64 addr = 0;
+ u32 res_id;
+ int err;
+
+ ib_uctx = ib_uverbs_get_ucontext(attrs);
+ if (IS_ERR(ib_uctx))
+ return PTR_ERR(ib_uctx);
+
+ err = uverbs_get_const(&res_type, attrs, BNXT_RE_TOGGLE_MEM_TYPE);
+ if (err)
+ return err;
+
+ uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
+ rdev = uctx->rdev;
+ err = uverbs_copy_from(&res_id, attrs, BNXT_RE_TOGGLE_MEM_RES_ID);
+ if (err)
+ return err;
+
+ switch (res_type) {
+ case BNXT_RE_CQ_TOGGLE_MEM:
+ cq = bnxt_re_search_for_cq(rdev, res_id);
+ if (!cq)
+ return -EINVAL;
+
+ addr = (u64)cq->uctx_cq_page;
+ break;
+ case BNXT_RE_SRQ_TOGGLE_MEM:
+ srq = bnxt_re_search_for_srq(rdev, res_id);
+ if (!srq)
+ return -EINVAL;
+
+ addr = (u64)srq->uctx_srq_page;
+ break;
+
+ default:
+ return -EOPNOTSUPP;
+ }
+
+ entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mem_offset);
+ if (!entry)
+ return -ENOMEM;
+
+ uobj->object = entry;
+ uverbs_finalize_uobj_create(attrs, BNXT_RE_TOGGLE_MEM_HANDLE);
+ err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_PAGE,
+ &mem_offset, sizeof(mem_offset));
+ if (err)
+ return err;
+
+ err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_LENGTH,
+ &length, sizeof(length));
+ if (err)
+ return err;
+
+ err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_OFFSET,
+ &offset, sizeof(offset));
+ if (err)
+ return err;
+
+ return 0;
+}
+
+static int get_toggle_mem_obj_cleanup(struct ib_uobject *uobject,
+ enum rdma_remove_reason why,
+ struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_user_mmap_entry *entry = uobject->object;
+
+ rdma_user_mmap_entry_remove(&entry->rdma_entry);
+ return 0;
+}
+
+DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM,
+ UVERBS_ATTR_IDR(BNXT_RE_TOGGLE_MEM_HANDLE,
+ BNXT_RE_OBJECT_GET_TOGGLE_MEM,
+ UVERBS_ACCESS_NEW,
+ UA_MANDATORY),
+ UVERBS_ATTR_CONST_IN(BNXT_RE_TOGGLE_MEM_TYPE,
+ enum bnxt_re_get_toggle_mem_type,
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_IN(BNXT_RE_TOGGLE_MEM_RES_ID,
+ UVERBS_ATTR_TYPE(u32),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_PAGE,
+ UVERBS_ATTR_TYPE(u64),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_OFFSET,
+ UVERBS_ATTR_TYPE(u32),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_LENGTH,
+ UVERBS_ATTR_TYPE(u32),
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM,
+ UVERBS_ATTR_IDR(BNXT_RE_RELEASE_TOGGLE_MEM_HANDLE,
+ BNXT_RE_OBJECT_GET_TOGGLE_MEM,
+ UVERBS_ACCESS_DESTROY,
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_GET_TOGGLE_MEM,
+ UVERBS_TYPE_ALLOC_IDR(get_toggle_mem_obj_cleanup),
+ &UVERBS_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM),
+ &UVERBS_METHOD(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM));
+
+const struct uapi_definition bnxt_re_uapi_defs[] = {
+ UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
+ UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
+ UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_GET_TOGGLE_MEM),
+ {}
+};
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
index f19b55c13d58..f758f92ba72b 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
@@ -627,7 +627,7 @@ static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
return rc;
}
-static struct bnxt_re_user_mmap_entry*
+struct bnxt_re_user_mmap_entry*
bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
enum bnxt_re_mmap_flag mmap_flag, u64 *offset)
{
@@ -4531,32 +4531,6 @@ int bnxt_re_destroy_flow(struct ib_flow *flow_id)
return rc;
}
-static struct bnxt_re_cq *bnxt_re_search_for_cq(struct bnxt_re_dev *rdev, u32 cq_id)
-{
- struct bnxt_re_cq *cq = NULL, *tmp_cq;
-
- hash_for_each_possible(rdev->cq_hash, tmp_cq, hash_entry, cq_id) {
- if (tmp_cq->qplib_cq.id == cq_id) {
- cq = tmp_cq;
- break;
- }
- }
- return cq;
-}
-
-static struct bnxt_re_srq *bnxt_re_search_for_srq(struct bnxt_re_dev *rdev, u32 srq_id)
-{
- struct bnxt_re_srq *srq = NULL, *tmp_srq;
-
- hash_for_each_possible(rdev->srq_hash, tmp_srq, hash_entry, srq_id) {
- if (tmp_srq->qplib_srq.id == srq_id) {
- srq = tmp_srq;
- break;
- }
- }
- return srq;
-}
-
/* Helper function to mmap the virtual memory from user app */
int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
{
@@ -4659,280 +4633,3 @@ int bnxt_re_process_mad(struct ib_device *ibdev, int mad_flags,
ret |= IB_MAD_RESULT_REPLY;
return ret;
}
-
-static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs)
-{
- struct bnxt_re_ucontext *uctx;
-
- uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
- bnxt_re_pacing_alert(uctx->rdev);
- return 0;
-}
-
-static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs)
-{
- struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
- enum bnxt_re_alloc_page_type alloc_type;
- struct bnxt_re_user_mmap_entry *entry;
- enum bnxt_re_mmap_flag mmap_flag;
- struct bnxt_qplib_chip_ctx *cctx;
- struct bnxt_re_ucontext *uctx;
- struct bnxt_re_dev *rdev;
- u64 mmap_offset;
- u32 length;
- u32 dpi;
- u64 addr;
- int err;
-
- uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
- if (IS_ERR(uctx))
- return PTR_ERR(uctx);
-
- err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE);
- if (err)
- return err;
-
- rdev = uctx->rdev;
- cctx = rdev->chip_ctx;
-
- switch (alloc_type) {
- case BNXT_RE_ALLOC_WC_PAGE:
- if (cctx->modes.db_push) {
- if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi,
- uctx, BNXT_QPLIB_DPI_TYPE_WC))
- return -ENOMEM;
- length = PAGE_SIZE;
- dpi = uctx->wcdpi.dpi;
- addr = (u64)uctx->wcdpi.umdbr;
- mmap_flag = BNXT_RE_MMAP_WC_DB;
- } else {
- return -EINVAL;
- }
-
- break;
- case BNXT_RE_ALLOC_DBR_BAR_PAGE:
- length = PAGE_SIZE;
- addr = (u64)rdev->pacing.dbr_bar_addr;
- mmap_flag = BNXT_RE_MMAP_DBR_BAR;
- break;
-
- case BNXT_RE_ALLOC_DBR_PAGE:
- length = PAGE_SIZE;
- addr = (u64)rdev->pacing.dbr_page;
- mmap_flag = BNXT_RE_MMAP_DBR_PAGE;
- break;
-
- default:
- return -EOPNOTSUPP;
- }
-
- entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mmap_offset);
- if (!entry)
- return -ENOMEM;
-
- uobj->object = entry;
- uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
- err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
- &mmap_offset, sizeof(mmap_offset));
- if (err)
- return err;
-
- err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
- &length, sizeof(length));
- if (err)
- return err;
-
- err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI,
- &dpi, sizeof(dpi));
- if (err)
- return err;
-
- return 0;
-}
-
-static int alloc_page_obj_cleanup(struct ib_uobject *uobject,
- enum rdma_remove_reason why,
- struct uverbs_attr_bundle *attrs)
-{
- struct bnxt_re_user_mmap_entry *entry = uobject->object;
- struct bnxt_re_ucontext *uctx = entry->uctx;
-
- switch (entry->mmap_flag) {
- case BNXT_RE_MMAP_WC_DB:
- if (uctx && uctx->wcdpi.dbr) {
- struct bnxt_re_dev *rdev = uctx->rdev;
-
- bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
- uctx->wcdpi.dbr = NULL;
- }
- break;
- case BNXT_RE_MMAP_DBR_BAR:
- case BNXT_RE_MMAP_DBR_PAGE:
- break;
- default:
- goto exit;
- }
- rdma_user_mmap_entry_remove(&entry->rdma_entry);
-exit:
- return 0;
-}
-
-DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE,
- UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE,
- BNXT_RE_OBJECT_ALLOC_PAGE,
- UVERBS_ACCESS_NEW,
- UA_MANDATORY),
- UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE,
- enum bnxt_re_alloc_page_type,
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
- UVERBS_ATTR_TYPE(u64),
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
- UVERBS_ATTR_TYPE(u32),
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI,
- UVERBS_ATTR_TYPE(u32),
- UA_MANDATORY));
-
-DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE,
- UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE,
- BNXT_RE_OBJECT_ALLOC_PAGE,
- UVERBS_ACCESS_DESTROY,
- UA_MANDATORY));
-
-DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE,
- UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup),
- &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE),
- &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE));
-
-DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_NOTIFY_DRV);
-
-DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_NOTIFY_DRV,
- &UVERBS_METHOD(BNXT_RE_METHOD_NOTIFY_DRV));
-
-/* Toggle MEM */
-static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bundle *attrs)
-{
- struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_TOGGLE_MEM_HANDLE);
- enum bnxt_re_mmap_flag mmap_flag = BNXT_RE_MMAP_TOGGLE_PAGE;
- enum bnxt_re_get_toggle_mem_type res_type;
- struct bnxt_re_user_mmap_entry *entry;
- struct bnxt_re_ucontext *uctx;
- struct ib_ucontext *ib_uctx;
- struct bnxt_re_dev *rdev;
- struct bnxt_re_srq *srq;
- u32 length = PAGE_SIZE;
- struct bnxt_re_cq *cq;
- u64 mem_offset;
- u32 offset = 0;
- u64 addr = 0;
- u32 res_id;
- int err;
-
- ib_uctx = ib_uverbs_get_ucontext(attrs);
- if (IS_ERR(ib_uctx))
- return PTR_ERR(ib_uctx);
-
- err = uverbs_get_const(&res_type, attrs, BNXT_RE_TOGGLE_MEM_TYPE);
- if (err)
- return err;
-
- uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
- rdev = uctx->rdev;
- err = uverbs_copy_from(&res_id, attrs, BNXT_RE_TOGGLE_MEM_RES_ID);
- if (err)
- return err;
-
- switch (res_type) {
- case BNXT_RE_CQ_TOGGLE_MEM:
- cq = bnxt_re_search_for_cq(rdev, res_id);
- if (!cq)
- return -EINVAL;
-
- addr = (u64)cq->uctx_cq_page;
- break;
- case BNXT_RE_SRQ_TOGGLE_MEM:
- srq = bnxt_re_search_for_srq(rdev, res_id);
- if (!srq)
- return -EINVAL;
-
- addr = (u64)srq->uctx_srq_page;
- break;
-
- default:
- return -EOPNOTSUPP;
- }
-
- entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mem_offset);
- if (!entry)
- return -ENOMEM;
-
- uobj->object = entry;
- uverbs_finalize_uobj_create(attrs, BNXT_RE_TOGGLE_MEM_HANDLE);
- err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_PAGE,
- &mem_offset, sizeof(mem_offset));
- if (err)
- return err;
-
- err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_LENGTH,
- &length, sizeof(length));
- if (err)
- return err;
-
- err = uverbs_copy_to(attrs, BNXT_RE_TOGGLE_MEM_MMAP_OFFSET,
- &offset, sizeof(offset));
- if (err)
- return err;
-
- return 0;
-}
-
-static int get_toggle_mem_obj_cleanup(struct ib_uobject *uobject,
- enum rdma_remove_reason why,
- struct uverbs_attr_bundle *attrs)
-{
- struct bnxt_re_user_mmap_entry *entry = uobject->object;
-
- rdma_user_mmap_entry_remove(&entry->rdma_entry);
- return 0;
-}
-
-DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM,
- UVERBS_ATTR_IDR(BNXT_RE_TOGGLE_MEM_HANDLE,
- BNXT_RE_OBJECT_GET_TOGGLE_MEM,
- UVERBS_ACCESS_NEW,
- UA_MANDATORY),
- UVERBS_ATTR_CONST_IN(BNXT_RE_TOGGLE_MEM_TYPE,
- enum bnxt_re_get_toggle_mem_type,
- UA_MANDATORY),
- UVERBS_ATTR_PTR_IN(BNXT_RE_TOGGLE_MEM_RES_ID,
- UVERBS_ATTR_TYPE(u32),
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_PAGE,
- UVERBS_ATTR_TYPE(u64),
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_OFFSET,
- UVERBS_ATTR_TYPE(u32),
- UA_MANDATORY),
- UVERBS_ATTR_PTR_OUT(BNXT_RE_TOGGLE_MEM_MMAP_LENGTH,
- UVERBS_ATTR_TYPE(u32),
- UA_MANDATORY));
-
-DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM,
- UVERBS_ATTR_IDR(BNXT_RE_RELEASE_TOGGLE_MEM_HANDLE,
- BNXT_RE_OBJECT_GET_TOGGLE_MEM,
- UVERBS_ACCESS_DESTROY,
- UA_MANDATORY));
-
-DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_GET_TOGGLE_MEM,
- UVERBS_TYPE_ALLOC_IDR(get_toggle_mem_obj_cleanup),
- &UVERBS_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM),
- &UVERBS_METHOD(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM));
-
-const struct uapi_definition bnxt_re_uapi_defs[] = {
- UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
- UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
- UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_GET_TOGGLE_MEM),
- {}
-};
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
index 76ba9ab04d5c..a11f56730a31 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
@@ -293,4 +293,7 @@ static inline u32 __to_ib_port_num(u16 port_id)
unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp);
void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, unsigned long flags);
+struct bnxt_re_user_mmap_entry*
+bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
+ enum bnxt_re_mmap_flag mmap_flag, u64 *offset);
#endif /* __BNXT_RE_IB_VERBS_H__ */
--
2.51.2.636.ga99f379adf
^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH rdma-next v6 2/4] RDMA/bnxt_re: Refactor bnxt_qplib_create_qp() function
2025-12-24 4:25 [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs Sriharsha Basavapatna
2025-12-24 4:25 ` [PATCH rdma-next v6 1/4] RDMA/bnxt_re: Move the UAPI methods to a dedicated file Sriharsha Basavapatna
@ 2025-12-24 4:26 ` Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
3 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2025-12-24 4:26 UTC (permalink / raw)
To: leon, jgg
Cc: linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
From: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Inside bnxt_qplib_create_qp(), driver currently is doing
a lot of things like allocating HWQ memory for SQ/RQ/ORRQ/IRRQ,
initializing few of qplib_qp fields etc.
Refactored the code such that all memory allocation for HWQs
have been moved to bnxt_re_init_qp_attr() function and inside
bnxt_qplib_create_qp() function just initialize the request
structure and issue the HWRM command to firmware.
Introduced couple of new functions bnxt_re_setup_qp_hwqs() and
bnxt_re_setup_qp_swqs() moved the hwq and swq memory allocation
logic there.
This patch also introduces a change to store the PD id in
bnxt_qplib_qp. Instead of keeping a pointer to "struct
bnxt_qplib_pd", store PD id directly in "struct bnxt_qplib_qp".
This change is needed for a subsequent change in this patch
series. This PD ID value will be used in new DV implementation
for create_qp(). There is no functional change.
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Reviewed-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 205 ++++++++++++--
drivers/infiniband/hw/bnxt_re/qplib_fp.c | 310 +++++++---------------
drivers/infiniband/hw/bnxt_re/qplib_fp.h | 10 +-
drivers/infiniband/hw/bnxt_re/qplib_res.h | 6 +
4 files changed, 301 insertions(+), 230 deletions(-)
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
index f758f92ba72b..0d95eaee3885 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
@@ -967,6 +967,12 @@ static void bnxt_re_del_unique_gid(struct bnxt_re_dev *rdev)
dev_err(rdev_to_dev(rdev), "Failed to delete unique GID, rc: %d\n", rc);
}
+static void bnxt_re_qp_free_umem(struct bnxt_re_qp *qp)
+{
+ ib_umem_release(qp->rumem);
+ ib_umem_release(qp->sumem);
+}
+
/* Queue Pairs */
int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
{
@@ -1009,8 +1015,7 @@ int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_RAW_ETHERTYPE)
bnxt_re_del_unique_gid(rdev);
- ib_umem_release(qp->rumem);
- ib_umem_release(qp->sumem);
+ bnxt_re_qp_free_umem(qp);
/* Flush all the entries of notification queue associated with
* given qp.
@@ -1154,6 +1159,7 @@ static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
}
qplib_qp->dpi = &cntx->dpi;
+ qplib_qp->is_user = true;
return 0;
rqfail:
ib_umem_release(qp->sumem);
@@ -1211,6 +1217,114 @@ static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
return NULL;
}
+static int bnxt_re_qp_alloc_init_xrrq(struct bnxt_re_qp *qp)
+{
+ struct bnxt_qplib_res *res = &qp->rdev->qplib_res;
+ struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
+ struct bnxt_qplib_hwq_attr hwq_attr = {};
+ struct bnxt_qplib_sg_info sginfo = {};
+ struct bnxt_qplib_hwq *irrq, *orrq;
+ int rc, req_size;
+
+ orrq = &qplib_qp->orrq;
+ orrq->max_elements =
+ ORD_LIMIT_TO_ORRQ_SLOTS(qplib_qp->max_rd_atomic);
+ req_size = orrq->max_elements *
+ BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE + PAGE_SIZE - 1;
+ req_size &= ~(PAGE_SIZE - 1);
+ sginfo.pgsize = req_size;
+ sginfo.pgshft = PAGE_SHIFT;
+
+ hwq_attr.res = res;
+ hwq_attr.sginfo = &sginfo;
+ hwq_attr.depth = orrq->max_elements;
+ hwq_attr.stride = BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE;
+ hwq_attr.aux_stride = 0;
+ hwq_attr.aux_depth = 0;
+ hwq_attr.type = HWQ_TYPE_CTX;
+ rc = bnxt_qplib_alloc_init_hwq(orrq, &hwq_attr);
+ if (rc)
+ return rc;
+
+ irrq = &qplib_qp->irrq;
+ irrq->max_elements =
+ IRD_LIMIT_TO_IRRQ_SLOTS(qplib_qp->max_dest_rd_atomic);
+ req_size = irrq->max_elements *
+ BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE + PAGE_SIZE - 1;
+ req_size &= ~(PAGE_SIZE - 1);
+ sginfo.pgsize = req_size;
+ hwq_attr.sginfo = &sginfo;
+ hwq_attr.depth = irrq->max_elements;
+ hwq_attr.stride = BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE;
+ rc = bnxt_qplib_alloc_init_hwq(irrq, &hwq_attr);
+ if (rc)
+ goto free_orrq_hwq;
+ return 0;
+free_orrq_hwq:
+ bnxt_qplib_free_hwq(res, orrq);
+ return rc;
+}
+
+static int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp)
+{
+ struct bnxt_qplib_res *res = &qp->rdev->qplib_res;
+ struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
+ struct bnxt_qplib_hwq_attr hwq_attr = {};
+ struct bnxt_qplib_q *sq = &qplib_qp->sq;
+ struct bnxt_qplib_q *rq = &qplib_qp->rq;
+ u8 wqe_mode = qplib_qp->wqe_mode;
+ u8 pg_sz_lvl;
+ int rc;
+
+ hwq_attr.res = res;
+ hwq_attr.sginfo = &sq->sg_info;
+ hwq_attr.stride = bnxt_qplib_get_stride();
+ hwq_attr.depth = bnxt_qplib_get_depth(sq, wqe_mode, true);
+ hwq_attr.aux_stride = qplib_qp->psn_sz;
+ hwq_attr.aux_depth = (qplib_qp->psn_sz) ?
+ bnxt_qplib_set_sq_size(sq, wqe_mode) : 0;
+ if (qplib_qp->is_host_msn_tbl && qplib_qp->psn_sz)
+ hwq_attr.aux_depth = qplib_qp->msn_tbl_sz;
+ hwq_attr.type = HWQ_TYPE_QUEUE;
+ rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr);
+ if (rc)
+ return rc;
+
+ pg_sz_lvl = bnxt_qplib_base_pg_size(&sq->hwq) << CMDQ_CREATE_QP_SQ_PG_SIZE_SFT;
+ pg_sz_lvl |= ((sq->hwq.level & CMDQ_CREATE_QP_SQ_LVL_MASK) <<
+ CMDQ_CREATE_QP_SQ_LVL_SFT);
+ sq->hwq.pg_sz_lvl = pg_sz_lvl;
+
+ hwq_attr.res = res;
+ hwq_attr.sginfo = &rq->sg_info;
+ hwq_attr.stride = bnxt_qplib_get_stride();
+ hwq_attr.depth = bnxt_qplib_get_depth(rq, qplib_qp->wqe_mode, false);
+ hwq_attr.aux_stride = 0;
+ hwq_attr.aux_depth = 0;
+ hwq_attr.type = HWQ_TYPE_QUEUE;
+ rc = bnxt_qplib_alloc_init_hwq(&rq->hwq, &hwq_attr);
+ if (rc)
+ goto free_sq_hwq;
+ pg_sz_lvl = bnxt_qplib_base_pg_size(&rq->hwq) <<
+ CMDQ_CREATE_QP_RQ_PG_SIZE_SFT;
+ pg_sz_lvl |= ((rq->hwq.level & CMDQ_CREATE_QP_RQ_LVL_MASK) <<
+ CMDQ_CREATE_QP_RQ_LVL_SFT);
+ rq->hwq.pg_sz_lvl = pg_sz_lvl;
+
+ if (qplib_qp->psn_sz) {
+ rc = bnxt_re_qp_alloc_init_xrrq(qp);
+ if (rc)
+ goto free_rq_hwq;
+ }
+
+ return 0;
+free_rq_hwq:
+ bnxt_qplib_free_hwq(res, &rq->hwq);
+free_sq_hwq:
+ bnxt_qplib_free_hwq(res, &sq->hwq);
+ return rc;
+}
+
static struct bnxt_re_qp *bnxt_re_create_shadow_qp
(struct bnxt_re_pd *pd,
struct bnxt_qplib_res *qp1_res,
@@ -1229,9 +1343,10 @@ static struct bnxt_re_qp *bnxt_re_create_shadow_qp
/* Initialize the shadow QP structure from the QP1 values */
ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
- qp->qplib_qp.pd = &pd->qplib_pd;
+ qp->qplib_qp.pd_id = pd->qplib_pd.id;
qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
qp->qplib_qp.type = IB_QPT_UD;
+ qp->qplib_qp.cctx = rdev->chip_ctx;
qp->qplib_qp.max_inline_data = 0;
qp->qplib_qp.sig_type = true;
@@ -1264,10 +1379,14 @@ static struct bnxt_re_qp *bnxt_re_create_shadow_qp
qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
qp->qplib_qp.dpi = &rdev->dpi_privileged;
- rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
+ rc = bnxt_re_setup_qp_hwqs(qp);
if (rc)
goto fail;
+ rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
+ if (rc)
+ goto free_hwq;
+
spin_lock_init(&qp->sq_lock);
INIT_LIST_HEAD(&qp->list);
mutex_lock(&rdev->qp_lock);
@@ -1275,6 +1394,9 @@ static struct bnxt_re_qp *bnxt_re_create_shadow_qp
atomic_inc(&rdev->stats.res.qp_count);
mutex_unlock(&rdev->qp_lock);
return qp;
+
+free_hwq:
+ bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
fail:
kfree(qp);
return NULL;
@@ -1445,6 +1567,39 @@ static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
return qptype;
}
+static void bnxt_re_qp_calculate_msn_psn_size(struct bnxt_re_qp *qp)
+{
+ struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
+ struct bnxt_qplib_q *sq = &qplib_qp->sq;
+ struct bnxt_re_dev *rdev = qp->rdev;
+ u8 wqe_mode = qplib_qp->wqe_mode;
+
+ if (rdev->dev_attr)
+ qplib_qp->is_host_msn_tbl =
+ _is_host_msn_table(rdev->dev_attr->dev_cap_flags2);
+
+ if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
+ qplib_qp->psn_sz = bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx) ?
+ sizeof(struct sq_psn_search_ext) :
+ sizeof(struct sq_psn_search);
+ if (qplib_qp->is_host_msn_tbl) {
+ qplib_qp->psn_sz = sizeof(struct sq_msn_search);
+ qplib_qp->msn = 0;
+ }
+ }
+
+ /* Update msn tbl size */
+ if (qplib_qp->is_host_msn_tbl && qplib_qp->psn_sz) {
+ if (wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
+ qplib_qp->msn_tbl_sz =
+ roundup_pow_of_two(bnxt_qplib_set_sq_size(sq, wqe_mode));
+ else
+ qplib_qp->msn_tbl_sz =
+ roundup_pow_of_two(bnxt_qplib_set_sq_size(sq, wqe_mode)) / 2;
+ qplib_qp->msn = 0;
+ }
+}
+
static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
struct ib_qp_init_attr *init_attr,
struct bnxt_re_ucontext *uctx,
@@ -1462,17 +1617,17 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
/* Setup misc params */
ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
- qplqp->pd = &pd->qplib_pd;
+ qplqp->pd_id = pd->qplib_pd.id;
qplqp->qp_handle = (u64)qplqp;
qplqp->max_inline_data = init_attr->cap.max_inline_data;
qplqp->sig_type = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
qptype = bnxt_re_init_qp_type(rdev, init_attr);
- if (qptype < 0) {
- rc = qptype;
- goto out;
- }
+ if (qptype < 0)
+ return qptype;
qplqp->type = (u8)qptype;
qplqp->wqe_mode = bnxt_re_is_var_size_supported(rdev, uctx);
+ qplqp->dev_cap_flags = dev_attr->dev_cap_flags;
+ qplqp->cctx = rdev->chip_ctx;
if (init_attr->qp_type == IB_QPT_RC) {
qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
@@ -1502,20 +1657,32 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
/* Setup RQ/SRQ */
rc = bnxt_re_init_rq_attr(qp, init_attr, uctx);
if (rc)
- goto out;
+ return rc;
if (init_attr->qp_type == IB_QPT_GSI)
bnxt_re_adjust_gsi_rq_attr(qp);
/* Setup SQ */
rc = bnxt_re_init_sq_attr(qp, init_attr, uctx, ureq);
if (rc)
- goto out;
+ return rc;
if (init_attr->qp_type == IB_QPT_GSI)
bnxt_re_adjust_gsi_sq_attr(qp, init_attr, uctx);
- if (uctx) /* This will update DPI and qp_handle */
+ if (uctx) { /* This will update DPI and qp_handle */
rc = bnxt_re_init_user_qp(rdev, pd, qp, uctx, ureq);
-out:
+ if (rc)
+ return rc;
+ }
+
+ bnxt_re_qp_calculate_msn_psn_size(qp);
+
+ rc = bnxt_re_setup_qp_hwqs(qp);
+ if (rc)
+ goto free_umem;
+
+ return 0;
+free_umem:
+ bnxt_re_qp_free_umem(qp);
return rc;
}
@@ -1573,6 +1740,7 @@ static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
rdev = qp->rdev;
qplqp = &qp->qplib_qp;
+ qplqp->cctx = rdev->chip_ctx;
qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
@@ -1676,13 +1844,14 @@ int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
if (rc == -ENODEV)
goto qp_destroy;
if (rc)
- goto fail;
+ goto free_hwq;
} else {
rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
if (rc) {
ibdev_err(&rdev->ibdev, "Failed to create HW QP");
- goto free_umem;
+ goto free_hwq;
}
+
if (udata) {
struct bnxt_re_qp_resp resp;
@@ -1733,9 +1902,9 @@ int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
return 0;
qp_destroy:
bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
-free_umem:
- ib_umem_release(qp->rumem);
- ib_umem_release(qp->sumem);
+free_hwq:
+ bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
+ bnxt_re_qp_free_umem(qp);
fail:
return rc;
}
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_fp.c b/drivers/infiniband/hw/bnxt_re/qplib_fp.c
index c88f049136fc..43535c6ec70b 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_fp.c
+++ b/drivers/infiniband/hw/bnxt_re/qplib_fp.c
@@ -793,8 +793,6 @@ int bnxt_qplib_post_srq_recv(struct bnxt_qplib_srq *srq,
return 0;
}
-/* QP */
-
static int bnxt_qplib_alloc_init_swq(struct bnxt_qplib_q *que)
{
int indx;
@@ -813,9 +811,71 @@ static int bnxt_qplib_alloc_init_swq(struct bnxt_qplib_q *que)
return 0;
}
+static int bnxt_re_setup_qp_swqs(struct bnxt_qplib_qp *qplqp)
+{
+ struct bnxt_qplib_q *sq = &qplqp->sq;
+ struct bnxt_qplib_q *rq = &qplqp->rq;
+ int rc;
+
+ if (qplqp->is_user)
+ return 0;
+
+ rc = bnxt_qplib_alloc_init_swq(sq);
+ if (rc)
+ return rc;
+
+ if (!qplqp->srq) {
+ rc = bnxt_qplib_alloc_init_swq(rq);
+ if (rc)
+ goto free_sq_swq;
+ }
+
+ return 0;
+free_sq_swq:
+ kfree(sq->swq);
+ return rc;
+}
+
+static void bnxt_qp_init_dbinfo(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
+{
+ struct bnxt_qplib_q *sq = &qp->sq;
+ struct bnxt_qplib_q *rq = &qp->rq;
+
+ sq->dbinfo.hwq = &sq->hwq;
+ sq->dbinfo.xid = qp->id;
+ sq->dbinfo.db = qp->dpi->dbr;
+ sq->dbinfo.max_slot = bnxt_qplib_set_sq_max_slot(qp->wqe_mode);
+ sq->dbinfo.flags = 0;
+ if (rq->max_wqe) {
+ rq->dbinfo.hwq = &rq->hwq;
+ rq->dbinfo.xid = qp->id;
+ rq->dbinfo.db = qp->dpi->dbr;
+ rq->dbinfo.max_slot = bnxt_qplib_set_rq_max_slot(rq->wqe_size);
+ rq->dbinfo.flags = 0;
+ }
+}
+
+static void bnxt_qplib_init_psn_ptr(struct bnxt_qplib_qp *qp, int size)
+{
+ struct bnxt_qplib_hwq *sq_hwq;
+ struct bnxt_qplib_q *sq;
+ u64 fpsne, psn_pg;
+ u16 indx_pad = 0;
+
+ sq = &qp->sq;
+ sq_hwq = &sq->hwq;
+ /* First psn entry */
+ fpsne = (u64)bnxt_qplib_get_qe(sq_hwq, sq_hwq->depth, &psn_pg);
+ if (!IS_ALIGNED(fpsne, PAGE_SIZE))
+ indx_pad = (fpsne & ~PAGE_MASK) / size;
+ sq_hwq->pad_pgofft = indx_pad;
+ sq_hwq->pad_pg = (u64 *)psn_pg;
+ sq_hwq->pad_stride = size;
+}
+
+/* QP */
int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
{
- struct bnxt_qplib_hwq_attr hwq_attr = {};
struct bnxt_qplib_rcfw *rcfw = res->rcfw;
struct creq_create_qp1_resp resp = {};
struct bnxt_qplib_cmdqmsg msg = {};
@@ -824,7 +884,6 @@ int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
struct cmdq_create_qp1 req = {};
struct bnxt_qplib_pbl *pbl;
u32 qp_flags = 0;
- u8 pg_sz_lvl;
u32 tbl_indx;
int rc;
@@ -838,26 +897,12 @@ int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
req.qp_handle = cpu_to_le64(qp->qp_handle);
/* SQ */
- hwq_attr.res = res;
- hwq_attr.sginfo = &sq->sg_info;
- hwq_attr.stride = sizeof(struct sq_sge);
- hwq_attr.depth = bnxt_qplib_get_depth(sq, qp->wqe_mode, false);
- hwq_attr.type = HWQ_TYPE_QUEUE;
- rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr);
- if (rc)
- return rc;
-
- rc = bnxt_qplib_alloc_init_swq(sq);
- if (rc)
- goto fail_sq;
+ sq->max_sw_wqe = bnxt_qplib_get_depth(sq, qp->wqe_mode, true);
+ req.sq_size = cpu_to_le32(sq->max_sw_wqe);
+ req.sq_pg_size_sq_lvl = sq->hwq.pg_sz_lvl;
- req.sq_size = cpu_to_le32(bnxt_qplib_set_sq_size(sq, qp->wqe_mode));
pbl = &sq->hwq.pbl[PBL_LVL_0];
req.sq_pbl = cpu_to_le64(pbl->pg_map_arr[0]);
- pg_sz_lvl = (bnxt_qplib_base_pg_size(&sq->hwq) <<
- CMDQ_CREATE_QP1_SQ_PG_SIZE_SFT);
- pg_sz_lvl |= (sq->hwq.level & CMDQ_CREATE_QP1_SQ_LVL_MASK);
- req.sq_pg_size_sq_lvl = pg_sz_lvl;
req.sq_fwo_sq_sge =
cpu_to_le16((sq->max_sge & CMDQ_CREATE_QP1_SQ_SGE_MASK) <<
CMDQ_CREATE_QP1_SQ_SGE_SFT);
@@ -866,24 +911,10 @@ int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
/* RQ */
if (rq->max_wqe) {
rq->dbinfo.flags = 0;
- hwq_attr.res = res;
- hwq_attr.sginfo = &rq->sg_info;
- hwq_attr.stride = sizeof(struct sq_sge);
- hwq_attr.depth = bnxt_qplib_get_depth(rq, qp->wqe_mode, false);
- hwq_attr.type = HWQ_TYPE_QUEUE;
- rc = bnxt_qplib_alloc_init_hwq(&rq->hwq, &hwq_attr);
- if (rc)
- goto sq_swq;
- rc = bnxt_qplib_alloc_init_swq(rq);
- if (rc)
- goto fail_rq;
req.rq_size = cpu_to_le32(rq->max_wqe);
pbl = &rq->hwq.pbl[PBL_LVL_0];
req.rq_pbl = cpu_to_le64(pbl->pg_map_arr[0]);
- pg_sz_lvl = (bnxt_qplib_base_pg_size(&rq->hwq) <<
- CMDQ_CREATE_QP1_RQ_PG_SIZE_SFT);
- pg_sz_lvl |= (rq->hwq.level & CMDQ_CREATE_QP1_RQ_LVL_MASK);
- req.rq_pg_size_rq_lvl = pg_sz_lvl;
+ req.rq_pg_size_rq_lvl = rq->hwq.pg_sz_lvl;
req.rq_fwo_rq_sge =
cpu_to_le16((rq->max_sge &
CMDQ_CREATE_QP1_RQ_SGE_MASK) <<
@@ -894,11 +925,11 @@ int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
rc = bnxt_qplib_alloc_qp_hdr_buf(res, qp);
if (rc) {
rc = -ENOMEM;
- goto rq_rwq;
+ return rc;
}
qp_flags |= CMDQ_CREATE_QP1_QP_FLAGS_RESERVED_LKEY_ENABLE;
req.qp_flags = cpu_to_le32(qp_flags);
- req.pd_id = cpu_to_le32(qp->pd->id);
+ req.pd_id = cpu_to_le32(qp->pd_id);
bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0);
rc = bnxt_qplib_rcfw_send_message(rcfw, &msg);
@@ -907,73 +938,39 @@ int bnxt_qplib_create_qp1(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
qp->id = le32_to_cpu(resp.xid);
qp->cur_qp_state = CMDQ_MODIFY_QP_NEW_STATE_RESET;
- qp->cctx = res->cctx;
- sq->dbinfo.hwq = &sq->hwq;
- sq->dbinfo.xid = qp->id;
- sq->dbinfo.db = qp->dpi->dbr;
- sq->dbinfo.max_slot = bnxt_qplib_set_sq_max_slot(qp->wqe_mode);
- if (rq->max_wqe) {
- rq->dbinfo.hwq = &rq->hwq;
- rq->dbinfo.xid = qp->id;
- rq->dbinfo.db = qp->dpi->dbr;
- rq->dbinfo.max_slot = bnxt_qplib_set_rq_max_slot(rq->wqe_size);
- }
+
+ rc = bnxt_re_setup_qp_swqs(qp);
+ if (rc)
+ goto destroy_qp;
+ bnxt_qp_init_dbinfo(res, qp);
+
tbl_indx = map_qp_id_to_tbl_indx(qp->id, rcfw);
rcfw->qp_tbl[tbl_indx].qp_id = qp->id;
rcfw->qp_tbl[tbl_indx].qp_handle = (void *)qp;
return 0;
+destroy_qp:
+ bnxt_qplib_destroy_qp(res, qp);
fail:
bnxt_qplib_free_qp_hdr_buf(res, qp);
-rq_rwq:
- kfree(rq->swq);
-fail_rq:
- bnxt_qplib_free_hwq(res, &rq->hwq);
-sq_swq:
- kfree(sq->swq);
-fail_sq:
- bnxt_qplib_free_hwq(res, &sq->hwq);
return rc;
}
-static void bnxt_qplib_init_psn_ptr(struct bnxt_qplib_qp *qp, int size)
-{
- struct bnxt_qplib_hwq *hwq;
- struct bnxt_qplib_q *sq;
- u64 fpsne, psn_pg;
- u16 indx_pad = 0;
-
- sq = &qp->sq;
- hwq = &sq->hwq;
- /* First psn entry */
- fpsne = (u64)bnxt_qplib_get_qe(hwq, hwq->depth, &psn_pg);
- if (!IS_ALIGNED(fpsne, PAGE_SIZE))
- indx_pad = (fpsne & ~PAGE_MASK) / size;
- hwq->pad_pgofft = indx_pad;
- hwq->pad_pg = (u64 *)psn_pg;
- hwq->pad_stride = size;
-}
-
int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
{
struct bnxt_qplib_rcfw *rcfw = res->rcfw;
- struct bnxt_qplib_hwq_attr hwq_attr = {};
- struct bnxt_qplib_sg_info sginfo = {};
struct creq_create_qp_resp resp = {};
struct bnxt_qplib_cmdqmsg msg = {};
struct bnxt_qplib_q *sq = &qp->sq;
struct bnxt_qplib_q *rq = &qp->rq;
struct cmdq_create_qp req = {};
- int rc, req_size, psn_sz = 0;
- struct bnxt_qplib_hwq *xrrq;
struct bnxt_qplib_pbl *pbl;
u32 qp_flags = 0;
- u8 pg_sz_lvl;
u32 tbl_indx;
u16 nsge;
+ int rc;
- qp->is_host_msn_tbl = _is_host_msn_table(res->dattr->dev_cap_flags2);
sq->dbinfo.flags = 0;
bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req,
CMDQ_BASE_OPCODE_CREATE_QP,
@@ -985,56 +982,10 @@ int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
req.qp_handle = cpu_to_le64(qp->qp_handle);
/* SQ */
- if (qp->type == CMDQ_CREATE_QP_TYPE_RC) {
- psn_sz = bnxt_qplib_is_chip_gen_p5_p7(res->cctx) ?
- sizeof(struct sq_psn_search_ext) :
- sizeof(struct sq_psn_search);
-
- if (qp->is_host_msn_tbl) {
- psn_sz = sizeof(struct sq_msn_search);
- qp->msn = 0;
- }
- }
-
- hwq_attr.res = res;
- hwq_attr.sginfo = &sq->sg_info;
- hwq_attr.stride = sizeof(struct sq_sge);
- hwq_attr.depth = bnxt_qplib_get_depth(sq, qp->wqe_mode, true);
- hwq_attr.aux_stride = psn_sz;
- hwq_attr.aux_depth = psn_sz ? bnxt_qplib_set_sq_size(sq, qp->wqe_mode)
- : 0;
- /* Update msn tbl size */
- if (qp->is_host_msn_tbl && psn_sz) {
- if (qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
- hwq_attr.aux_depth =
- roundup_pow_of_two(bnxt_qplib_set_sq_size(sq, qp->wqe_mode));
- else
- hwq_attr.aux_depth =
- roundup_pow_of_two(bnxt_qplib_set_sq_size(sq, qp->wqe_mode)) / 2;
- qp->msn_tbl_sz = hwq_attr.aux_depth;
- qp->msn = 0;
- }
-
- hwq_attr.type = HWQ_TYPE_QUEUE;
- rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr);
- if (rc)
- return rc;
-
- if (!sq->hwq.is_user) {
- rc = bnxt_qplib_alloc_init_swq(sq);
- if (rc)
- goto fail_sq;
-
- if (psn_sz)
- bnxt_qplib_init_psn_ptr(qp, psn_sz);
- }
- req.sq_size = cpu_to_le32(bnxt_qplib_set_sq_size(sq, qp->wqe_mode));
+ req.sq_size = cpu_to_le32(sq->max_sw_wqe);
pbl = &sq->hwq.pbl[PBL_LVL_0];
req.sq_pbl = cpu_to_le64(pbl->pg_map_arr[0]);
- pg_sz_lvl = (bnxt_qplib_base_pg_size(&sq->hwq) <<
- CMDQ_CREATE_QP_SQ_PG_SIZE_SFT);
- pg_sz_lvl |= (sq->hwq.level & CMDQ_CREATE_QP_SQ_LVL_MASK);
- req.sq_pg_size_sq_lvl = pg_sz_lvl;
+ req.sq_pg_size_sq_lvl = sq->hwq.pg_sz_lvl;
req.sq_fwo_sq_sge =
cpu_to_le16(((sq->max_sge & CMDQ_CREATE_QP_SQ_SGE_MASK) <<
CMDQ_CREATE_QP_SQ_SGE_SFT) | 0);
@@ -1043,29 +994,10 @@ int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
/* RQ */
if (!qp->srq) {
rq->dbinfo.flags = 0;
- hwq_attr.res = res;
- hwq_attr.sginfo = &rq->sg_info;
- hwq_attr.stride = sizeof(struct sq_sge);
- hwq_attr.depth = bnxt_qplib_get_depth(rq, qp->wqe_mode, false);
- hwq_attr.aux_stride = 0;
- hwq_attr.aux_depth = 0;
- hwq_attr.type = HWQ_TYPE_QUEUE;
- rc = bnxt_qplib_alloc_init_hwq(&rq->hwq, &hwq_attr);
- if (rc)
- goto sq_swq;
- if (!rq->hwq.is_user) {
- rc = bnxt_qplib_alloc_init_swq(rq);
- if (rc)
- goto fail_rq;
- }
-
req.rq_size = cpu_to_le32(rq->max_wqe);
pbl = &rq->hwq.pbl[PBL_LVL_0];
req.rq_pbl = cpu_to_le64(pbl->pg_map_arr[0]);
- pg_sz_lvl = (bnxt_qplib_base_pg_size(&rq->hwq) <<
- CMDQ_CREATE_QP_RQ_PG_SIZE_SFT);
- pg_sz_lvl |= (rq->hwq.level & CMDQ_CREATE_QP_RQ_LVL_MASK);
- req.rq_pg_size_rq_lvl = pg_sz_lvl;
+ req.rq_pg_size_rq_lvl = rq->hwq.pg_sz_lvl;
nsge = (qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ?
6 : rq->max_sge;
req.rq_fwo_rq_sge =
@@ -1091,68 +1023,34 @@ int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
req.qp_flags = cpu_to_le32(qp_flags);
/* ORRQ and IRRQ */
- if (psn_sz) {
- xrrq = &qp->orrq;
- xrrq->max_elements =
- ORD_LIMIT_TO_ORRQ_SLOTS(qp->max_rd_atomic);
- req_size = xrrq->max_elements *
- BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE + PAGE_SIZE - 1;
- req_size &= ~(PAGE_SIZE - 1);
- sginfo.pgsize = req_size;
- sginfo.pgshft = PAGE_SHIFT;
-
- hwq_attr.res = res;
- hwq_attr.sginfo = &sginfo;
- hwq_attr.depth = xrrq->max_elements;
- hwq_attr.stride = BNXT_QPLIB_MAX_ORRQE_ENTRY_SIZE;
- hwq_attr.aux_stride = 0;
- hwq_attr.aux_depth = 0;
- hwq_attr.type = HWQ_TYPE_CTX;
- rc = bnxt_qplib_alloc_init_hwq(xrrq, &hwq_attr);
- if (rc)
- goto rq_swq;
- pbl = &xrrq->pbl[PBL_LVL_0];
- req.orrq_addr = cpu_to_le64(pbl->pg_map_arr[0]);
-
- xrrq = &qp->irrq;
- xrrq->max_elements = IRD_LIMIT_TO_IRRQ_SLOTS(
- qp->max_dest_rd_atomic);
- req_size = xrrq->max_elements *
- BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE + PAGE_SIZE - 1;
- req_size &= ~(PAGE_SIZE - 1);
- sginfo.pgsize = req_size;
- hwq_attr.depth = xrrq->max_elements;
- hwq_attr.stride = BNXT_QPLIB_MAX_IRRQE_ENTRY_SIZE;
- rc = bnxt_qplib_alloc_init_hwq(xrrq, &hwq_attr);
- if (rc)
- goto fail_orrq;
-
- pbl = &xrrq->pbl[PBL_LVL_0];
- req.irrq_addr = cpu_to_le64(pbl->pg_map_arr[0]);
+ if (qp->psn_sz) {
+ req.orrq_addr = cpu_to_le64(bnxt_qplib_get_base_addr(&qp->orrq));
+ req.irrq_addr = cpu_to_le64(bnxt_qplib_get_base_addr(&qp->irrq));
}
- req.pd_id = cpu_to_le32(qp->pd->id);
+
+ req.pd_id = cpu_to_le32(qp->pd_id);
bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req),
sizeof(resp), 0);
rc = bnxt_qplib_rcfw_send_message(rcfw, &msg);
if (rc)
- goto fail;
+ return rc;
qp->id = le32_to_cpu(resp.xid);
+
+ if (!qp->is_user) {
+ rc = bnxt_re_setup_qp_swqs(qp);
+ if (rc)
+ goto destroy_qp;
+ }
+ bnxt_qp_init_dbinfo(res, qp);
+ if (qp->psn_sz)
+ bnxt_qplib_init_psn_ptr(qp, qp->psn_sz);
+
qp->cur_qp_state = CMDQ_MODIFY_QP_NEW_STATE_RESET;
INIT_LIST_HEAD(&qp->sq_flush);
INIT_LIST_HEAD(&qp->rq_flush);
qp->cctx = res->cctx;
- sq->dbinfo.hwq = &sq->hwq;
- sq->dbinfo.xid = qp->id;
- sq->dbinfo.db = qp->dpi->dbr;
- sq->dbinfo.max_slot = bnxt_qplib_set_sq_max_slot(qp->wqe_mode);
- if (rq->max_wqe) {
- rq->dbinfo.hwq = &rq->hwq;
- rq->dbinfo.xid = qp->id;
- rq->dbinfo.db = qp->dpi->dbr;
- rq->dbinfo.max_slot = bnxt_qplib_set_rq_max_slot(rq->wqe_size);
- }
spin_lock_bh(&rcfw->tbl_lock);
tbl_indx = map_qp_id_to_tbl_indx(qp->id, rcfw);
rcfw->qp_tbl[tbl_indx].qp_id = qp->id;
@@ -1160,18 +1058,8 @@ int bnxt_qplib_create_qp(struct bnxt_qplib_res *res, struct bnxt_qplib_qp *qp)
spin_unlock_bh(&rcfw->tbl_lock);
return 0;
-fail:
- bnxt_qplib_free_hwq(res, &qp->irrq);
-fail_orrq:
- bnxt_qplib_free_hwq(res, &qp->orrq);
-rq_swq:
- kfree(rq->swq);
-fail_rq:
- bnxt_qplib_free_hwq(res, &rq->hwq);
-sq_swq:
- kfree(sq->swq);
-fail_sq:
- bnxt_qplib_free_hwq(res, &sq->hwq);
+destroy_qp:
+ bnxt_qplib_destroy_qp(res, qp);
return rc;
}
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_fp.h b/drivers/infiniband/hw/bnxt_re/qplib_fp.h
index 1b414a73b46d..c862fa7ba499 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_fp.h
+++ b/drivers/infiniband/hw/bnxt_re/qplib_fp.h
@@ -268,7 +268,7 @@ struct bnxt_qplib_q {
};
struct bnxt_qplib_qp {
- struct bnxt_qplib_pd *pd;
+ u32 pd_id;
struct bnxt_qplib_dpi *dpi;
struct bnxt_qplib_chip_ctx *cctx;
u64 qp_handle;
@@ -279,6 +279,7 @@ struct bnxt_qplib_qp {
u8 wqe_mode;
u8 state;
u8 cur_qp_state;
+ u8 is_user;
u64 modify_flags;
u32 max_inline_data;
u32 mtu;
@@ -343,9 +344,11 @@ struct bnxt_qplib_qp {
struct list_head rq_flush;
u32 msn;
u32 msn_tbl_sz;
+ u32 psn_sz;
bool is_host_msn_tbl;
u8 tos_dscp;
u32 ugid_index;
+ u16 dev_cap_flags;
};
#define BNXT_RE_MAX_MSG_SIZE 0x80000000
@@ -614,6 +617,11 @@ static inline void bnxt_qplib_swq_mod_start(struct bnxt_qplib_q *que, u32 idx)
que->swq_start = que->swq[idx].next_idx;
}
+static inline u32 bnxt_qplib_get_stride(void)
+{
+ return sizeof(struct sq_sge);
+}
+
static inline u32 bnxt_qplib_get_depth(struct bnxt_qplib_q *que, u8 wqe_mode, bool is_sq)
{
u32 slots;
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_res.h b/drivers/infiniband/hw/bnxt_re/qplib_res.h
index 2ea3b7f232a3..ccdab938d707 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_res.h
+++ b/drivers/infiniband/hw/bnxt_re/qplib_res.h
@@ -198,6 +198,7 @@ struct bnxt_qplib_hwq {
u32 cons; /* raw */
u8 cp_bit;
u8 is_user;
+ u8 pg_sz_lvl;
u64 *pad_pg;
u32 pad_stride;
u32 pad_pgofft;
@@ -358,6 +359,11 @@ static inline u8 bnxt_qplib_get_ring_type(struct bnxt_qplib_chip_ctx *cctx)
RING_ALLOC_REQ_RING_TYPE_ROCE_CMPL;
}
+static inline u64 bnxt_qplib_get_base_addr(struct bnxt_qplib_hwq *hwq)
+{
+ return hwq->pbl[PBL_LVL_0].pg_map_arr[0];
+}
+
static inline u8 bnxt_qplib_base_pg_size(struct bnxt_qplib_hwq *hwq)
{
u8 pg_size = BNXT_QPLIB_HWRM_PG_SIZE_4K;
--
2.51.2.636.ga99f379adf
^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2025-12-24 4:25 [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs Sriharsha Basavapatna
2025-12-24 4:25 ` [PATCH rdma-next v6 1/4] RDMA/bnxt_re: Move the UAPI methods to a dedicated file Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 2/4] RDMA/bnxt_re: Refactor bnxt_qplib_create_qp() function Sriharsha Basavapatna
@ 2025-12-24 4:26 ` Sriharsha Basavapatna
2026-01-09 18:35 ` Jason Gunthorpe
` (3 more replies)
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
3 siblings, 4 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2025-12-24 4:26 UTC (permalink / raw)
To: leon, jgg
Cc: linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
From: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
The following Direct Verbs (DV) methods have been implemented in
this patch.
Doorbell Region Direct Verbs:
-----------------------------
- BNXT_RE_METHOD_DBR_ALLOC:
This will allow the appliation to create extra doorbell regions
and use the associated doorbell page index in DV_CREATE_QP and
use the associated DB address while ringing the doorbell.
- BNXT_RE_METHOD_DBR_FREE:
Free the allocated doorbell region.
- BNXT_RE_METHOD_GET_DEFAULT_DBR:
Return the default doorbell page index and doorbell page address
associated with the ucontext.
Co-developed-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Reviewed-by: Selvin Xavier <selvin.xavier@broadcom.com>
---
drivers/infiniband/hw/bnxt_re/bnxt_re.h | 2 +
drivers/infiniband/hw/bnxt_re/dv.c | 135 ++++++++++++++++++++++
drivers/infiniband/hw/bnxt_re/ib_verbs.h | 7 ++
drivers/infiniband/hw/bnxt_re/main.c | 1 +
drivers/infiniband/hw/bnxt_re/qplib_res.c | 43 +++++++
drivers/infiniband/hw/bnxt_re/qplib_res.h | 4 +
include/uapi/rdma/bnxt_re-abi.h | 28 +++++
7 files changed, 220 insertions(+)
diff --git a/drivers/infiniband/hw/bnxt_re/bnxt_re.h b/drivers/infiniband/hw/bnxt_re/bnxt_re.h
index 3a7ce4729fcf..ce717398b980 100644
--- a/drivers/infiniband/hw/bnxt_re/bnxt_re.h
+++ b/drivers/infiniband/hw/bnxt_re/bnxt_re.h
@@ -160,6 +160,7 @@ struct bnxt_re_nq_record {
#define MAX_CQ_HASH_BITS (16)
#define MAX_SRQ_HASH_BITS (16)
+#define MAX_DPI_HASH_BITS (16)
static inline bool bnxt_re_chip_gen_p7(u16 chip_num)
{
@@ -217,6 +218,7 @@ struct bnxt_re_dev {
struct delayed_work dbq_pacing_work;
DECLARE_HASHTABLE(cq_hash, MAX_CQ_HASH_BITS);
DECLARE_HASHTABLE(srq_hash, MAX_SRQ_HASH_BITS);
+ DECLARE_HASHTABLE(dpi_hash, MAX_DPI_HASH_BITS);
struct dentry *dbg_root;
struct dentry *qp_debugfs;
unsigned long event_bitmap;
diff --git a/drivers/infiniband/hw/bnxt_re/dv.c b/drivers/infiniband/hw/bnxt_re/dv.c
index 5655c6176af4..6bc275c7413c 100644
--- a/drivers/infiniband/hw/bnxt_re/dv.c
+++ b/drivers/infiniband/hw/bnxt_re/dv.c
@@ -331,9 +331,144 @@ DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_GET_TOGGLE_MEM,
&UVERBS_METHOD(BNXT_RE_METHOD_GET_TOGGLE_MEM),
&UVERBS_METHOD(BNXT_RE_METHOD_RELEASE_TOGGLE_MEM));
+static int UVERBS_HANDLER(BNXT_RE_METHOD_DBR_ALLOC)(struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_dv_db_region dbr = {};
+ struct bnxt_re_alloc_dbr_obj *obj;
+ struct bnxt_re_ucontext *uctx;
+ struct ib_ucontext *ib_uctx;
+ struct bnxt_qplib_dpi *dpi;
+ struct bnxt_re_dev *rdev;
+ struct ib_uobject *uobj;
+ u64 mmap_offset;
+ int ret;
+
+ ib_uctx = ib_uverbs_get_ucontext(attrs);
+ if (IS_ERR(ib_uctx))
+ return PTR_ERR(ib_uctx);
+
+ uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
+ rdev = uctx->rdev;
+ uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
+
+ obj = kzalloc(sizeof(*obj), GFP_KERNEL);
+ if (!obj)
+ return -ENOMEM;
+
+ dpi = &obj->dpi;
+ ret = bnxt_qplib_alloc_uc_dpi(&rdev->qplib_res, dpi);
+ if (ret)
+ goto free_mem;
+
+ obj->entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_UC_DB,
+ &mmap_offset);
+ if (!obj->entry) {
+ ret = -ENOMEM;
+ goto free_dpi;
+ }
+
+ obj->rdev = rdev;
+ dbr.umdbr = dpi->umdbr;
+ dbr.dpi = dpi->dpi;
+
+ ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
+ &dbr, sizeof(dbr));
+ if (ret)
+ goto free_entry;
+
+ ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
+ &mmap_offset, sizeof(mmap_offset));
+ if (ret)
+ goto free_entry;
+
+ uobj->object = obj;
+ uverbs_finalize_uobj_create(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
+ hash_add(rdev->dpi_hash, &obj->hash_entry, dpi->dpi);
+ return 0;
+free_entry:
+ rdma_user_mmap_entry_remove(&obj->entry->rdma_entry);
+free_dpi:
+ bnxt_qplib_free_uc_dpi(&rdev->qplib_res, dpi);
+free_mem:
+ kfree(obj);
+ return ret;
+}
+
+static int bnxt_re_dv_dbr_cleanup(struct ib_uobject *uobject,
+ enum rdma_remove_reason why,
+ struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_alloc_dbr_obj *obj = uobject->object;
+ struct bnxt_re_dev *rdev = obj->rdev;
+
+ hash_del(&obj->hash_entry);
+ rdma_user_mmap_entry_remove(&obj->entry->rdma_entry);
+ bnxt_qplib_free_uc_dpi(&rdev->qplib_res, &obj->dpi);
+ kfree(obj);
+ return 0;
+}
+
+static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_DEFAULT_DBR)(struct uverbs_attr_bundle *attrs)
+{
+ struct bnxt_re_dv_db_region dpi = {};
+ struct bnxt_re_ucontext *uctx;
+ struct ib_ucontext *ib_uctx;
+ int ret;
+
+ ib_uctx = ib_uverbs_get_ucontext(attrs);
+ if (IS_ERR(ib_uctx))
+ return PTR_ERR(ib_uctx);
+
+ uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
+ dpi.umdbr = uctx->dpi.umdbr;
+ dpi.dpi = uctx->dpi.dpi;
+
+ ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_DEFAULT_DBR_ATTR,
+ &dpi, sizeof(dpi));
+ if (ret)
+ return ret;
+
+ return 0;
+}
+
+DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_DBR_ALLOC,
+ UVERBS_ATTR_IDR(BNXT_RE_DV_ALLOC_DBR_HANDLE,
+ BNXT_RE_OBJECT_DBR,
+ UVERBS_ACCESS_NEW,
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_DV_ALLOC_DBR_ATTR,
+ UVERBS_ATTR_STRUCT(struct bnxt_re_dv_db_region,
+ umdbr),
+ UA_MANDATORY),
+ UVERBS_ATTR_PTR_IN(BNXT_RE_DV_ALLOC_DBR_OFFSET,
+ UVERBS_ATTR_TYPE(u64),
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DBR_FREE,
+ UVERBS_ATTR_IDR(BNXT_RE_DV_FREE_DBR_HANDLE,
+ BNXT_RE_OBJECT_DBR,
+ UVERBS_ACCESS_DESTROY,
+ UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_DBR,
+ UVERBS_TYPE_ALLOC_IDR(bnxt_re_dv_dbr_cleanup),
+ &UVERBS_METHOD(BNXT_RE_METHOD_DBR_ALLOC),
+ &UVERBS_METHOD(BNXT_RE_METHOD_DBR_FREE));
+
+DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_GET_DEFAULT_DBR,
+ UVERBS_ATTR_PTR_OUT(BNXT_RE_DV_DEFAULT_DBR_ATTR,
+ UVERBS_ATTR_STRUCT(struct bnxt_re_dv_db_region,
+ umdbr),
+ UA_MANDATORY));
+
+DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_DEFAULT_DBR,
+ &UVERBS_METHOD(BNXT_RE_METHOD_GET_DEFAULT_DBR));
+
const struct uapi_definition bnxt_re_uapi_defs[] = {
UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_GET_TOGGLE_MEM),
+ UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_DBR),
+ UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_DEFAULT_DBR),
{}
};
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
index a11f56730a31..8fbc57484bab 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
@@ -164,6 +164,13 @@ struct bnxt_re_user_mmap_entry {
u8 mmap_flag;
};
+struct bnxt_re_alloc_dbr_obj {
+ struct bnxt_re_dev *rdev;
+ struct bnxt_qplib_dpi dpi;
+ struct bnxt_re_user_mmap_entry *entry;
+ struct hlist_node hash_entry;
+};
+
struct bnxt_re_flow {
struct ib_flow ib_flow;
struct bnxt_re_dev *rdev;
diff --git a/drivers/infiniband/hw/bnxt_re/main.c b/drivers/infiniband/hw/bnxt_re/main.c
index 73003ad25ee8..2f5f174ed5e7 100644
--- a/drivers/infiniband/hw/bnxt_re/main.c
+++ b/drivers/infiniband/hw/bnxt_re/main.c
@@ -2338,6 +2338,7 @@ static int bnxt_re_dev_init(struct bnxt_re_dev *rdev, u8 op_type)
bnxt_re_vf_res_config(rdev);
}
hash_init(rdev->cq_hash);
+ hash_init(rdev->dpi_hash);
if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT)
hash_init(rdev->srq_hash);
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_res.c b/drivers/infiniband/hw/bnxt_re/qplib_res.c
index 875d7b52c06a..30cc2d64a9ae 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_res.c
+++ b/drivers/infiniband/hw/bnxt_re/qplib_res.c
@@ -685,6 +685,49 @@ static int bnxt_qplib_alloc_pd_tbl(struct bnxt_qplib_res *res,
}
/* DPIs */
+int bnxt_qplib_alloc_uc_dpi(struct bnxt_qplib_res *res, struct bnxt_qplib_dpi *dpi)
+{
+ struct bnxt_qplib_dpi_tbl *dpit = &res->dpi_tbl;
+ struct bnxt_qplib_reg_desc *reg;
+ u32 bit_num;
+ int rc = 0;
+
+ reg = &dpit->wcreg;
+ mutex_lock(&res->dpi_tbl_lock);
+ bit_num = find_first_bit(dpit->tbl, dpit->max);
+ if (bit_num >= dpit->max) {
+ rc = -ENOMEM;
+ goto unlock;
+ }
+ /* Found unused DPI */
+ clear_bit(bit_num, dpit->tbl);
+ dpi->bit = bit_num;
+ dpi->dpi = bit_num + (reg->offset - dpit->ucreg.offset) / PAGE_SIZE;
+ dpi->umdbr = reg->bar_base + reg->offset + bit_num * PAGE_SIZE;
+unlock:
+ mutex_unlock(&res->dpi_tbl_lock);
+ return rc;
+}
+
+int bnxt_qplib_free_uc_dpi(struct bnxt_qplib_res *res, struct bnxt_qplib_dpi *dpi)
+{
+ struct bnxt_qplib_dpi_tbl *dpit = &res->dpi_tbl;
+ int rc = 0;
+
+ mutex_lock(&res->dpi_tbl_lock);
+ if (dpi->bit >= dpit->max) {
+ rc = -EINVAL;
+ goto unlock;
+ }
+
+ if (test_and_set_bit(dpi->bit, dpit->tbl))
+ rc = -EINVAL;
+ memset(dpi, 0, sizeof(*dpi));
+unlock:
+ mutex_unlock(&res->dpi_tbl_lock);
+ return rc;
+}
+
int bnxt_qplib_alloc_dpi(struct bnxt_qplib_res *res,
struct bnxt_qplib_dpi *dpi,
void *app, u8 type)
diff --git a/drivers/infiniband/hw/bnxt_re/qplib_res.h b/drivers/infiniband/hw/bnxt_re/qplib_res.h
index ccdab938d707..3a8162ef4c33 100644
--- a/drivers/infiniband/hw/bnxt_re/qplib_res.h
+++ b/drivers/infiniband/hw/bnxt_re/qplib_res.h
@@ -436,6 +436,10 @@ int bnxt_qplib_alloc_dpi(struct bnxt_qplib_res *res,
void *app, u8 type);
int bnxt_qplib_dealloc_dpi(struct bnxt_qplib_res *res,
struct bnxt_qplib_dpi *dpi);
+int bnxt_qplib_alloc_uc_dpi(struct bnxt_qplib_res *res,
+ struct bnxt_qplib_dpi *dpi);
+int bnxt_qplib_free_uc_dpi(struct bnxt_qplib_res *res,
+ struct bnxt_qplib_dpi *dpi);
void bnxt_qplib_cleanup_res(struct bnxt_qplib_res *res);
int bnxt_qplib_init_res(struct bnxt_qplib_res *res);
void bnxt_qplib_free_res(struct bnxt_qplib_res *res);
diff --git a/include/uapi/rdma/bnxt_re-abi.h b/include/uapi/rdma/bnxt_re-abi.h
index faa9d62b3b30..8df073714bbc 100644
--- a/include/uapi/rdma/bnxt_re-abi.h
+++ b/include/uapi/rdma/bnxt_re-abi.h
@@ -162,6 +162,8 @@ enum bnxt_re_objects {
BNXT_RE_OBJECT_ALLOC_PAGE = (1U << UVERBS_ID_NS_SHIFT),
BNXT_RE_OBJECT_NOTIFY_DRV,
BNXT_RE_OBJECT_GET_TOGGLE_MEM,
+ BNXT_RE_OBJECT_DBR,
+ BNXT_RE_OBJECT_DEFAULT_DBR,
};
enum bnxt_re_alloc_page_type {
@@ -215,4 +217,30 @@ enum bnxt_re_toggle_mem_methods {
BNXT_RE_METHOD_GET_TOGGLE_MEM = (1U << UVERBS_ID_NS_SHIFT),
BNXT_RE_METHOD_RELEASE_TOGGLE_MEM,
};
+
+struct bnxt_re_dv_db_region {
+ __u32 dpi;
+ __aligned_u64 umdbr;
+};
+
+enum bnxt_re_obj_dbr_alloc_attrs {
+ BNXT_RE_DV_ALLOC_DBR_HANDLE = (1U << UVERBS_ID_NS_SHIFT),
+ BNXT_RE_DV_ALLOC_DBR_ATTR,
+ BNXT_RE_DV_ALLOC_DBR_OFFSET,
+};
+
+enum bnxt_re_obj_dbr_free_attrs {
+ BNXT_RE_DV_FREE_DBR_HANDLE = (1U << UVERBS_ID_NS_SHIFT),
+};
+
+enum bnxt_re_obj_default_dbr_attrs {
+ BNXT_RE_DV_DEFAULT_DBR_ATTR = (1U << UVERBS_ID_NS_SHIFT),
+};
+
+enum bnxt_re_obj_dpi_methods {
+ BNXT_RE_METHOD_DBR_ALLOC = (1U << UVERBS_ID_NS_SHIFT),
+ BNXT_RE_METHOD_DBR_FREE,
+ BNXT_RE_METHOD_GET_DEFAULT_DBR,
+};
+
#endif /* __BNXT_RE_UVERBS_ABI_H__*/
--
2.51.2.636.ga99f379adf
^ permalink raw reply related [flat|nested] 23+ messages in thread
* [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2025-12-24 4:25 [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs Sriharsha Basavapatna
` (2 preceding siblings ...)
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
@ 2025-12-24 4:26 ` Sriharsha Basavapatna
2026-01-09 18:37 ` Jason Gunthorpe
` (2 more replies)
3 siblings, 3 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2025-12-24 4:26 UTC (permalink / raw)
To: leon, jgg
Cc: linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
The following Direct Verbs have been implemented, by enhancing the
driver specific udata in existing verbs.
CQ Direct Verbs:
----------------
- CREATE_CQ:
Create a CQ using the specified udata (struct bnxt_re_cq_req).
The driver maps/pins the CQ user memory and registers it with the
hardware.
- DESTROY_CQ:
Unmap the user memory and destroy the CQ.
QP Direct Verbs:
----------------
- CREATE_QP:
Create a QP using the specified udata (struct bnxt_re_qp_req).
The driver maps/pins the SQ/RQ user memory and registers it
with the hardware.
- DESTROY_QP:
Unmap SQ/RQ user memory and destroy the QP.
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Co-developed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Co-developed-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
---
drivers/infiniband/hw/bnxt_re/bnxt_re.h | 5 +
drivers/infiniband/hw/bnxt_re/dv.c | 465 +++++++++++++++++++++++
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 60 ++-
drivers/infiniband/hw/bnxt_re/ib_verbs.h | 11 +
include/uapi/rdma/bnxt_re-abi.h | 17 +
5 files changed, 539 insertions(+), 19 deletions(-)
diff --git a/drivers/infiniband/hw/bnxt_re/bnxt_re.h b/drivers/infiniband/hw/bnxt_re/bnxt_re.h
index ce717398b980..e77f19f551ee 100644
--- a/drivers/infiniband/hw/bnxt_re/bnxt_re.h
+++ b/drivers/infiniband/hw/bnxt_re/bnxt_re.h
@@ -235,6 +235,8 @@ struct bnxt_re_dev {
union ib_gid ugid;
u32 ugid_index;
u8 sniffer_flow_created : 1;
+ atomic_t dv_cq_count;
+ atomic_t dv_qp_count;
};
#define to_bnxt_re_dev(ptr, member) \
@@ -278,6 +280,9 @@ static inline int bnxt_re_read_context_allowed(struct bnxt_re_dev *rdev)
return 0;
}
+struct bnxt_qplib_nq *bnxt_re_get_nq(struct bnxt_re_dev *rdev);
+void bnxt_re_put_nq(struct bnxt_re_dev *rdev, struct bnxt_qplib_nq *nq);
+
#define BNXT_RE_CONTEXT_TYPE_QPC_SIZE_P5 1088
#define BNXT_RE_CONTEXT_TYPE_CQ_SIZE_P5 128
#define BNXT_RE_CONTEXT_TYPE_MRW_SIZE_P5 128
diff --git a/drivers/infiniband/hw/bnxt_re/dv.c b/drivers/infiniband/hw/bnxt_re/dv.c
index 6bc275c7413c..d7baadec0fdd 100644
--- a/drivers/infiniband/hw/bnxt_re/dv.c
+++ b/drivers/infiniband/hw/bnxt_re/dv.c
@@ -12,6 +12,7 @@
#include <rdma/ib_user_ioctl_cmds.h>
#define UVERBS_MODULE_NAME bnxt_re
#include <rdma/uverbs_named_ioctl.h>
+#include <rdma/ib_umem.h>
#include <rdma/bnxt_re-abi.h>
#include "roce_hsi.h"
@@ -48,6 +49,19 @@ static struct bnxt_re_srq *bnxt_re_search_for_srq(struct bnxt_re_dev *rdev, u32
return srq;
}
+static struct bnxt_re_alloc_dbr_obj *bnxt_re_search_for_dpi(struct bnxt_re_dev *rdev, u32 dpi)
+{
+ struct bnxt_re_alloc_dbr_obj *obj = NULL, *tmp_obj;
+
+ hash_for_each_possible(rdev->dpi_hash, tmp_obj, hash_entry, dpi) {
+ if (tmp_obj->dpi.dpi == dpi) {
+ obj = tmp_obj;
+ break;
+ }
+ }
+ return obj;
+}
+
static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs)
{
struct bnxt_re_ucontext *uctx;
@@ -464,6 +478,457 @@ DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_GET_DEFAULT_DBR,
DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_DEFAULT_DBR,
&UVERBS_METHOD(BNXT_RE_METHOD_GET_DEFAULT_DBR));
+static int bnxt_re_dv_create_cq_resp(struct bnxt_re_dev *rdev,
+ struct bnxt_re_cq *cq,
+ struct bnxt_re_cq_resp *resp)
+{
+ struct bnxt_qplib_cq *qplcq = &cq->qplib_cq;
+
+ resp->cqid = qplcq->id;
+ resp->tail = qplcq->hwq.cons;
+ resp->phase = qplcq->period;
+ resp->comp_mask = BNXT_RE_CQ_DV_SUPPORT;
+ return 0;
+}
+
+static struct ib_umem *bnxt_re_dv_umem_get(struct bnxt_re_dev *rdev,
+ struct ib_ucontext *ib_uctx,
+ int dmabuf_fd,
+ u64 addr, u64 size,
+ struct bnxt_qplib_sg_info *sg)
+{
+ int access = IB_ACCESS_LOCAL_WRITE;
+ struct ib_umem *umem;
+ int umem_pgs, rc;
+
+ if (dmabuf_fd) {
+ struct ib_umem_dmabuf *umem_dmabuf;
+
+ umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, addr, size,
+ dmabuf_fd, access);
+ if (IS_ERR(umem_dmabuf)) {
+ rc = PTR_ERR(umem_dmabuf);
+ goto umem_err;
+ }
+ umem = &umem_dmabuf->umem;
+ } else {
+ umem = ib_umem_get(&rdev->ibdev, addr, size, access);
+ if (IS_ERR(umem)) {
+ rc = PTR_ERR(umem);
+ goto umem_err;
+ }
+ }
+
+ umem_pgs = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
+ if (!umem_pgs) {
+ rc = -EINVAL;
+ goto umem_rel;
+ }
+ sg->npages = umem_pgs;
+ sg->pgshft = PAGE_SHIFT;
+ sg->pgsize = PAGE_SIZE;
+ sg->umem = umem;
+ return umem;
+
+umem_rel:
+ ib_umem_release(umem);
+umem_err:
+ return ERR_PTR(rc);
+}
+
+static int bnxt_re_dv_create_qplib_cq(struct bnxt_re_dev *rdev,
+ struct bnxt_re_ucontext *re_uctx,
+ struct bnxt_re_cq *cq,
+ struct bnxt_re_cq_req *req)
+{
+ struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr;
+ struct bnxt_qplib_cq *qplcq;
+ struct ib_umem *umem;
+ u32 cqe = req->ncqe;
+ u32 max_active_cqs;
+ int rc = -EINVAL;
+
+ if (atomic_read(&rdev->stats.res.cq_count) >= dev_attr->max_cq) {
+ ibdev_dbg(&rdev->ibdev, "Create CQ failed - max exceeded(CQs)");
+ return rc;
+ }
+
+ /* Validate CQ fields */
+ if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
+ ibdev_dbg(&rdev->ibdev, "Create CQ failed - max exceeded(CQ_WQs)");
+ return rc;
+ }
+
+ qplcq = &cq->qplib_cq;
+ qplcq->cq_handle = (u64)qplcq;
+ umem = bnxt_re_dv_umem_get(rdev, &re_uctx->ib_uctx, req->dmabuf_fd,
+ req->cq_va, cqe * sizeof(struct cq_base),
+ &qplcq->sg_info);
+ if (IS_ERR(umem)) {
+ rc = PTR_ERR(umem);
+ ibdev_dbg(&rdev->ibdev,
+ "bnxt_re_dv_umem_get() failed, rc: %d\n", rc);
+ return rc;
+ }
+ cq->umem = umem;
+ qplcq->dpi = &re_uctx->dpi;
+ qplcq->max_wqe = cqe;
+ qplcq->nq = bnxt_re_get_nq(rdev);
+ qplcq->cnq_hw_ring_id = qplcq->nq->ring_id;
+ qplcq->coalescing = &rdev->cq_coalescing;
+ rc = bnxt_qplib_create_cq(&rdev->qplib_res, qplcq);
+ if (rc) {
+ ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
+ goto fail_qpl;
+ }
+
+ cq->ib_cq.cqe = cqe;
+ cq->cq_period = qplcq->period;
+
+ max_active_cqs = atomic_inc_return(&rdev->stats.res.cq_count);
+ if (max_active_cqs > rdev->stats.res.cq_watermark)
+ rdev->stats.res.cq_watermark = max_active_cqs;
+ spin_lock_init(&cq->cq_lock);
+
+ return 0;
+
+fail_qpl:
+ ib_umem_release(cq->umem);
+ return rc;
+}
+
+static void bnxt_re_dv_init_ib_cq(struct bnxt_re_dev *rdev,
+ struct bnxt_re_cq *re_cq)
+{
+ struct ib_cq *ib_cq;
+
+ ib_cq = &re_cq->ib_cq;
+ ib_cq->device = &rdev->ibdev;
+ ib_cq->uobject = NULL;
+ ib_cq->comp_handler = NULL;
+ ib_cq->event_handler = NULL;
+ atomic_set(&ib_cq->usecnt, 0);
+}
+
+int bnxt_re_dv_create_cq(struct bnxt_re_dev *rdev, struct ib_udata *udata,
+ struct bnxt_re_cq *re_cq, struct bnxt_re_cq_req *req)
+{
+ struct bnxt_re_ucontext *re_uctx =
+ rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
+ struct bnxt_re_cq_resp resp = {};
+ int ret;
+
+ ret = bnxt_re_dv_create_qplib_cq(rdev, re_uctx, re_cq, req);
+ if (ret)
+ return ret;
+
+ ret = bnxt_re_dv_create_cq_resp(rdev, re_cq, &resp);
+ if (ret)
+ goto fail_resp;
+
+ ret = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
+ if (ret)
+ goto fail_resp;
+
+ bnxt_re_dv_init_ib_cq(rdev, re_cq);
+ re_cq->is_dv_cq = true;
+ atomic_inc(&rdev->dv_cq_count);
+ return 0;
+
+fail_resp:
+ bnxt_qplib_destroy_cq(&rdev->qplib_res, &re_cq->qplib_cq);
+ bnxt_re_put_nq(rdev, re_cq->qplib_cq.nq);
+ ib_umem_release(re_cq->umem);
+ return ret;
+};
+
+static int bnxt_re_dv_init_qp_attr(struct bnxt_re_qp *qp,
+ struct bnxt_re_ucontext *cntx,
+ struct ib_qp_init_attr *init_attr,
+ struct bnxt_re_qp_req *req,
+ struct bnxt_re_alloc_dbr_obj *dbr_obj)
+{
+ struct bnxt_qplib_dev_attr *dev_attr;
+ struct bnxt_qplib_qp *qplqp;
+ struct bnxt_re_cq *send_cq;
+ struct bnxt_re_cq *recv_cq;
+ struct bnxt_re_dev *rdev;
+ struct bnxt_qplib_q *rq;
+ struct bnxt_qplib_q *sq;
+ u32 slot_size;
+ int qptype;
+
+ rdev = qp->rdev;
+ qplqp = &qp->qplib_qp;
+ dev_attr = rdev->dev_attr;
+
+ /* Setup misc params */
+ qplqp->is_user = true;
+ qplqp->pd_id = req->pd_id;
+ qplqp->qp_handle = (u64)qplqp;
+ qplqp->sig_type = false;
+ qptype = __from_ib_qp_type(init_attr->qp_type);
+ if (qptype < 0)
+ return qptype;
+ qplqp->type = (u8)qptype;
+ qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode;
+ ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
+ qplqp->dev_cap_flags = dev_attr->dev_cap_flags;
+ qplqp->cctx = rdev->chip_ctx;
+
+ if (init_attr->qp_type == IB_QPT_RC) {
+ qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
+ qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
+ }
+ qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
+ if (dbr_obj)
+ qplqp->dpi = &dbr_obj->dpi;
+ else
+ qplqp->dpi = &cntx->dpi;
+
+ /* Setup CQs */
+ if (!init_attr->send_cq)
+ return -EINVAL;
+ send_cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
+ qplqp->scq = &send_cq->qplib_cq;
+ qp->scq = send_cq;
+
+ if (!init_attr->recv_cq)
+ return -EINVAL;
+ recv_cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
+ qplqp->rcq = &recv_cq->qplib_cq;
+ qp->rcq = recv_cq;
+
+ if (!init_attr->srq) {
+ /* Setup RQ */
+ slot_size = bnxt_qplib_get_stride();
+ rq = &qplqp->rq;
+ rq->max_sge = init_attr->cap.max_recv_sge;
+ rq->wqe_size = req->rq_wqe_sz;
+ rq->max_wqe = (req->rq_slots * slot_size) /
+ req->rq_wqe_sz;
+ rq->max_sw_wqe = rq->max_wqe;
+ rq->q_full_delta = 0;
+ rq->sg_info.pgsize = PAGE_SIZE;
+ rq->sg_info.pgshft = PAGE_SHIFT;
+ }
+
+ /* Setup SQ */
+ sq = &qplqp->sq;
+ sq->max_sge = init_attr->cap.max_send_sge;
+ sq->wqe_size = req->sq_wqe_sz;
+ sq->max_wqe = req->sq_slots; /* SQ in var-wqe mode */
+ sq->max_sw_wqe = sq->max_wqe;
+ sq->q_full_delta = 0;
+ sq->sg_info.pgsize = PAGE_SIZE;
+ sq->sg_info.pgshft = PAGE_SHIFT;
+
+ return 0;
+}
+
+static int bnxt_re_dv_init_user_qp(struct bnxt_re_dev *rdev,
+ struct bnxt_re_ucontext *cntx,
+ struct bnxt_re_qp *qp,
+ struct ib_qp_init_attr *init_attr,
+ struct bnxt_re_qp_req *req)
+{
+ struct bnxt_qplib_sg_info *sginfo;
+ struct bnxt_qplib_qp *qplib_qp;
+ struct ib_umem *umem;
+ int rc = -EINVAL;
+
+ qplib_qp = &qp->qplib_qp;
+ qplib_qp->qp_handle = req->qp_handle;
+ sginfo = &qplib_qp->sq.sg_info;
+
+ /* SQ */
+ umem = bnxt_re_dv_umem_get(rdev, &cntx->ib_uctx, req->sq_dmabuf_fd,
+ req->qpsva, req->sq_len, sginfo);
+ if (IS_ERR(umem)) {
+ rc = PTR_ERR(umem);
+ ibdev_dbg(&rdev->ibdev,
+ "bnxt_re_dv_umem_get() failed, rc: %d\n", rc);
+ return rc;
+ }
+ qp->sumem = umem;
+
+ /* SRQ */
+ if (init_attr->srq) {
+ struct bnxt_re_srq *srq;
+
+ srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
+ qplib_qp->srq = &srq->qplib_srq;
+ goto done;
+ }
+
+ /* RQ */
+ sginfo = &qplib_qp->rq.sg_info;
+ umem = bnxt_re_dv_umem_get(rdev, &cntx->ib_uctx, req->rq_dmabuf_fd,
+ req->qprva, req->rq_len, sginfo);
+ if (IS_ERR(umem)) {
+ rc = PTR_ERR(umem);
+ ibdev_dbg(&rdev->ibdev,
+ "bnxt_re_dv_umem_get() failed, rc: %d\n", rc);
+ goto rqfail;
+ }
+ qp->rumem = umem;
+done:
+ qplib_qp->is_user = true;
+ return 0;
+rqfail:
+ ib_umem_release(qp->sumem);
+ qplib_qp->sq.sg_info.umem = NULL;
+ return rc;
+}
+
+static int
+bnxt_re_dv_qp_init_msn(struct bnxt_re_dev *rdev, struct bnxt_re_qp *qp,
+ struct bnxt_re_qp_req *req)
+{
+ struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr;
+ struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
+
+ if (req->sq_npsn > dev_attr->max_qp_wqes ||
+ req->sq_psn_sz > sizeof(struct sq_psn_search_ext))
+ return -EINVAL;
+
+ qplib_qp->is_host_msn_tbl = true;
+ qplib_qp->msn = 0;
+ qplib_qp->psn_sz = req->sq_psn_sz;
+ qplib_qp->msn_tbl_sz = req->sq_psn_sz * req->sq_npsn;
+ return 0;
+}
+
+static void bnxt_re_dv_init_qp(struct bnxt_re_dev *rdev,
+ struct bnxt_re_qp *qp)
+{
+ u32 active_qps, tmp_qps;
+
+ spin_lock_init(&qp->sq_lock);
+ spin_lock_init(&qp->rq_lock);
+ INIT_LIST_HEAD(&qp->list);
+ mutex_lock(&rdev->qp_lock);
+ list_add_tail(&qp->list, &rdev->qp_list);
+ mutex_unlock(&rdev->qp_lock);
+ atomic_inc(&rdev->stats.res.qp_count);
+ active_qps = atomic_read(&rdev->stats.res.qp_count);
+ if (active_qps > rdev->stats.res.qp_watermark)
+ rdev->stats.res.qp_watermark = active_qps;
+
+ /* Get the counters for RC QPs */
+ tmp_qps = atomic_inc_return(&rdev->stats.res.rc_qp_count);
+ if (tmp_qps > rdev->stats.res.rc_qp_watermark)
+ rdev->stats.res.rc_qp_watermark = tmp_qps;
+}
+
+int bnxt_re_dv_create_qp(struct bnxt_re_dev *rdev, struct ib_udata *udata,
+ struct ib_qp_init_attr *init_attr,
+ struct bnxt_re_qp *re_qp, struct bnxt_re_qp_req *req)
+{
+ struct bnxt_re_alloc_dbr_obj *dbr_obj = NULL;
+ struct bnxt_re_cq *send_cq = NULL;
+ struct bnxt_re_cq *recv_cq = NULL;
+ struct bnxt_re_qp_resp resp = {};
+ struct bnxt_re_ucontext *uctx;
+ int ret;
+
+ uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
+ if (init_attr->send_cq) {
+ send_cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
+ re_qp->scq = send_cq;
+ }
+
+ if (init_attr->recv_cq) {
+ recv_cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
+ re_qp->rcq = recv_cq;
+ }
+
+ re_qp->rdev = rdev;
+ if (req->dpi != uctx->dpi.dpi) {
+ dbr_obj = bnxt_re_search_for_dpi(rdev, req->dpi);
+ if (!dbr_obj)
+ return -EINVAL;
+ }
+ ret = bnxt_re_dv_init_qp_attr(re_qp, uctx, init_attr, req, dbr_obj);
+ if (ret)
+ return ret;
+
+ ret = bnxt_re_dv_init_user_qp(rdev, uctx, re_qp, init_attr, req);
+ if (ret)
+ return ret;
+
+ ret = bnxt_re_dv_qp_init_msn(rdev, re_qp, req);
+ if (ret)
+ goto free_umem;
+
+ ret = bnxt_re_setup_qp_hwqs(re_qp, true);
+ if (ret)
+ goto free_umem;
+
+ ret = bnxt_qplib_create_qp(&rdev->qplib_res, &re_qp->qplib_qp);
+ if (ret) {
+ ibdev_err(&rdev->ibdev, "Failed to create HW QP");
+ goto free_hwq;
+ }
+
+ resp.qpid = re_qp->qplib_qp.id;
+ resp.comp_mask = BNXT_RE_QP_DV_SUPPORT;
+ resp.rsvd = 0;
+ ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
+ if (ret)
+ goto free_qplib;
+
+ bnxt_re_dv_init_qp(rdev, re_qp);
+ re_qp->is_dv_qp = true;
+ atomic_inc(&rdev->dv_qp_count);
+ return 0;
+
+free_qplib:
+ bnxt_qplib_destroy_qp(&rdev->qplib_res, &re_qp->qplib_qp);
+free_hwq:
+ bnxt_qplib_free_qp_res(&rdev->qplib_res, &re_qp->qplib_qp);
+free_umem:
+ bnxt_re_qp_free_umem(re_qp);
+ return ret;
+}
+
+int bnxt_re_dv_destroy_qp(struct bnxt_re_qp *qp)
+{
+ struct bnxt_re_dev *rdev = qp->rdev;
+ struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
+ struct bnxt_qplib_nq *scq_nq = NULL;
+ struct bnxt_qplib_nq *rcq_nq = NULL;
+ int rc;
+
+ mutex_lock(&rdev->qp_lock);
+ list_del(&qp->list);
+ atomic_dec(&rdev->stats.res.qp_count);
+ if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_RC)
+ atomic_dec(&rdev->stats.res.rc_qp_count);
+ mutex_unlock(&rdev->qp_lock);
+
+ rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, qplib_qp);
+ if (rc)
+ ibdev_err_ratelimited(&rdev->ibdev,
+ "id = %d failed rc = %d",
+ qplib_qp->id, rc);
+
+ bnxt_qplib_free_qp_res(&rdev->qplib_res, qplib_qp);
+ bnxt_re_qp_free_umem(qp);
+
+ /* Flush all the entries of notification queue associated with
+ * given qp.
+ */
+ scq_nq = qplib_qp->scq->nq;
+ rcq_nq = qplib_qp->rcq->nq;
+ bnxt_re_synchronize_nq(scq_nq);
+ if (scq_nq != rcq_nq)
+ bnxt_re_synchronize_nq(rcq_nq);
+
+ atomic_dec(&rdev->dv_qp_count);
+ return 0;
+}
+
const struct uapi_definition bnxt_re_uapi_defs[] = {
UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
index 0d95eaee3885..e9ac01000ef5 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
@@ -967,7 +967,7 @@ static void bnxt_re_del_unique_gid(struct bnxt_re_dev *rdev)
dev_err(rdev_to_dev(rdev), "Failed to delete unique GID, rc: %d\n", rc);
}
-static void bnxt_re_qp_free_umem(struct bnxt_re_qp *qp)
+void bnxt_re_qp_free_umem(struct bnxt_re_qp *qp)
{
ib_umem_release(qp->rumem);
ib_umem_release(qp->sumem);
@@ -984,6 +984,9 @@ int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
unsigned int flags;
int rc;
+ if (qp->is_dv_qp)
+ return bnxt_re_dv_destroy_qp(qp);
+
bnxt_re_debug_rem_qpinfo(rdev, qp);
bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
@@ -1029,7 +1032,7 @@ int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
return 0;
}
-static u8 __from_ib_qp_type(enum ib_qp_type type)
+u8 __from_ib_qp_type(enum ib_qp_type type)
{
switch (type) {
case IB_QPT_GSI:
@@ -1265,7 +1268,7 @@ static int bnxt_re_qp_alloc_init_xrrq(struct bnxt_re_qp *qp)
return rc;
}
-static int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp)
+int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp, bool is_dv_qp)
{
struct bnxt_qplib_res *res = &qp->rdev->qplib_res;
struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
@@ -1279,12 +1282,17 @@ static int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp)
hwq_attr.res = res;
hwq_attr.sginfo = &sq->sg_info;
hwq_attr.stride = bnxt_qplib_get_stride();
- hwq_attr.depth = bnxt_qplib_get_depth(sq, wqe_mode, true);
hwq_attr.aux_stride = qplib_qp->psn_sz;
- hwq_attr.aux_depth = (qplib_qp->psn_sz) ?
- bnxt_qplib_set_sq_size(sq, wqe_mode) : 0;
- if (qplib_qp->is_host_msn_tbl && qplib_qp->psn_sz)
+ if (!is_dv_qp) {
+ hwq_attr.depth = bnxt_qplib_get_depth(sq, wqe_mode, true);
+ hwq_attr.aux_depth = (qplib_qp->psn_sz) ?
+ bnxt_qplib_set_sq_size(sq, wqe_mode) : 0;
+ if (qplib_qp->is_host_msn_tbl && qplib_qp->psn_sz)
+ hwq_attr.aux_depth = qplib_qp->msn_tbl_sz;
+ } else {
+ hwq_attr.depth = sq->max_wqe;
hwq_attr.aux_depth = qplib_qp->msn_tbl_sz;
+ }
hwq_attr.type = HWQ_TYPE_QUEUE;
rc = bnxt_qplib_alloc_init_hwq(&sq->hwq, &hwq_attr);
if (rc)
@@ -1295,10 +1303,16 @@ static int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp)
CMDQ_CREATE_QP_SQ_LVL_SFT);
sq->hwq.pg_sz_lvl = pg_sz_lvl;
+ if (qplib_qp->srq)
+ goto done;
+
hwq_attr.res = res;
hwq_attr.sginfo = &rq->sg_info;
hwq_attr.stride = bnxt_qplib_get_stride();
- hwq_attr.depth = bnxt_qplib_get_depth(rq, qplib_qp->wqe_mode, false);
+ if (!is_dv_qp)
+ hwq_attr.depth = bnxt_qplib_get_depth(rq, qplib_qp->wqe_mode, false);
+ else
+ hwq_attr.depth = rq->max_wqe * 3;
hwq_attr.aux_stride = 0;
hwq_attr.aux_depth = 0;
hwq_attr.type = HWQ_TYPE_QUEUE;
@@ -1311,6 +1325,7 @@ static int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp)
CMDQ_CREATE_QP_RQ_LVL_SFT);
rq->hwq.pg_sz_lvl = pg_sz_lvl;
+done:
if (qplib_qp->psn_sz) {
rc = bnxt_re_qp_alloc_init_xrrq(qp);
if (rc)
@@ -1379,7 +1394,7 @@ static struct bnxt_re_qp *bnxt_re_create_shadow_qp
qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
qp->qplib_qp.dpi = &rdev->dpi_privileged;
- rc = bnxt_re_setup_qp_hwqs(qp);
+ rc = bnxt_re_setup_qp_hwqs(qp, false);
if (rc)
goto fail;
@@ -1676,7 +1691,7 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
bnxt_re_qp_calculate_msn_psn_size(qp);
- rc = bnxt_re_setup_qp_hwqs(qp);
+ rc = bnxt_re_setup_qp_hwqs(qp, false);
if (rc)
goto free_umem;
@@ -1823,9 +1838,12 @@ int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
- if (udata)
+ if (udata) {
if (ib_copy_from_udata(&ureq, udata, min(udata->inlen, sizeof(ureq))))
return -EFAULT;
+ if (ureq.comp_mask & BNXT_RE_QP_DV_SUPPORT)
+ return bnxt_re_dv_create_qp(rdev, udata, qp_init_attr, qp, &ureq);
+ }
rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
if (!rc) {
@@ -3241,7 +3259,7 @@ int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
return rc;
}
-static struct bnxt_qplib_nq *bnxt_re_get_nq(struct bnxt_re_dev *rdev)
+struct bnxt_qplib_nq *bnxt_re_get_nq(struct bnxt_re_dev *rdev)
{
int min, indx;
@@ -3256,7 +3274,7 @@ static struct bnxt_qplib_nq *bnxt_re_get_nq(struct bnxt_re_dev *rdev)
return &rdev->nqr->nq[min];
}
-static void bnxt_re_put_nq(struct bnxt_re_dev *rdev, struct bnxt_qplib_nq *nq)
+void bnxt_re_put_nq(struct bnxt_re_dev *rdev, struct bnxt_qplib_nq *nq)
{
mutex_lock(&rdev->nqr->load_lock);
nq->load--;
@@ -3284,6 +3302,8 @@ int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
bnxt_re_put_nq(rdev, nq);
ib_umem_release(cq->umem);
+ if (cq->is_dv_cq)
+ atomic_dec(&rdev->dv_cq_count);
atomic_dec(&rdev->stats.res.cq_count);
kfree(cq->cql);
@@ -3300,6 +3320,7 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr;
struct bnxt_qplib_chip_ctx *cctx;
+ struct bnxt_re_cq_req req;
int cqe = attr->cqe;
int rc, entries;
u32 active_cqs;
@@ -3317,6 +3338,13 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
cctx = rdev->chip_ctx;
cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
+ if (udata) {
+ if (ib_copy_from_udata(&req, udata, sizeof(req)))
+ return -EFAULT;
+ if (req.comp_mask & BNXT_RE_CQ_DV_SUPPORT)
+ return bnxt_re_dv_create_cq(rdev, udata, cq, &req);
+ }
+
entries = bnxt_re_init_depth(cqe + 1, uctx);
if (entries > dev_attr->max_cq_wqes + 1)
entries = dev_attr->max_cq_wqes + 1;
@@ -3324,12 +3352,6 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
if (udata) {
- struct bnxt_re_cq_req req;
- if (ib_copy_from_udata(&req, udata, sizeof(req))) {
- rc = -EFAULT;
- goto fail;
- }
-
cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
entries * sizeof(struct cq_base),
IB_ACCESS_LOCAL_WRITE);
diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
index 8fbc57484bab..50fb70d1e78d 100644
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h
@@ -96,6 +96,7 @@ struct bnxt_re_qp {
struct bnxt_re_cq *scq;
struct bnxt_re_cq *rcq;
struct dentry *dentry;
+ bool is_dv_qp;
};
struct bnxt_re_cq {
@@ -113,6 +114,7 @@ struct bnxt_re_cq {
int resize_cqe;
void *uctx_cq_page;
struct hlist_node hash_entry;
+ bool is_dv_cq;
};
struct bnxt_re_mr {
@@ -303,4 +305,13 @@ void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, unsigned long flags);
struct bnxt_re_user_mmap_entry*
bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
enum bnxt_re_mmap_flag mmap_flag, u64 *offset);
+u8 __from_ib_qp_type(enum ib_qp_type type);
+int bnxt_re_setup_qp_hwqs(struct bnxt_re_qp *qp, bool is_dv_qp);
+void bnxt_re_qp_free_umem(struct bnxt_re_qp *qp);
+int bnxt_re_dv_create_cq(struct bnxt_re_dev *rdev, struct ib_udata *udata,
+ struct bnxt_re_cq *re_cq, struct bnxt_re_cq_req *req);
+int bnxt_re_dv_create_qp(struct bnxt_re_dev *rdev, struct ib_udata *udata,
+ struct ib_qp_init_attr *init_attr,
+ struct bnxt_re_qp *re_qp, struct bnxt_re_qp_req *req);
+int bnxt_re_dv_destroy_qp(struct bnxt_re_qp *qp);
#endif /* __BNXT_RE_IB_VERBS_H__ */
diff --git a/include/uapi/rdma/bnxt_re-abi.h b/include/uapi/rdma/bnxt_re-abi.h
index 8df073714bbc..b38bf86b4d1e 100644
--- a/include/uapi/rdma/bnxt_re-abi.h
+++ b/include/uapi/rdma/bnxt_re-abi.h
@@ -101,10 +101,14 @@ struct bnxt_re_pd_resp {
struct bnxt_re_cq_req {
__aligned_u64 cq_va;
__aligned_u64 cq_handle;
+ __aligned_u64 comp_mask;
+ __u32 ncqe;
+ __u32 dmabuf_fd;
};
enum bnxt_re_cq_mask {
BNXT_RE_CQ_TOGGLE_PAGE_SUPPORT = 0x1,
+ BNXT_RE_CQ_DV_SUPPORT = 0x2
};
struct bnxt_re_cq_resp {
@@ -121,6 +125,7 @@ struct bnxt_re_resize_cq_req {
enum bnxt_re_qp_mask {
BNXT_RE_QP_REQ_MASK_VAR_WQE_SQ_SLOTS = 0x1,
+ BNXT_RE_QP_DV_SUPPORT = 0x2,
};
struct bnxt_re_qp_req {
@@ -129,11 +134,23 @@ struct bnxt_re_qp_req {
__aligned_u64 qp_handle;
__aligned_u64 comp_mask;
__u32 sq_slots;
+ __u32 pd_id;
+ __u32 dpi;
+ __u32 sq_dmabuf_fd;
+ __u32 sq_len; /* total len including MSN area */
+ __u32 sq_wqe_sz;
+ __u32 sq_psn_sz;
+ __u32 sq_npsn;
+ __u32 rq_dmabuf_fd;
+ __u32 rq_len;
+ __u32 rq_slots; /* == max_recv_wr */
+ __u32 rq_wqe_sz;
};
struct bnxt_re_qp_resp {
__u32 qpid;
__u32 rsvd;
+ __aligned_u64 comp_mask;
};
struct bnxt_re_srq_req {
--
2.51.2.636.ga99f379adf
^ permalink raw reply related [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
@ 2026-01-09 18:35 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 18:50 ` Jason Gunthorpe
` (2 subsequent siblings)
3 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 18:35 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> +struct bnxt_re_dv_db_region {
> + __u32 dpi;
> + __aligned_u64 umdbr;
> +};
This needs a '__u32 reserved' to mark the implicit padding..
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
@ 2026-01-09 18:37 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 19:08 ` Jason Gunthorpe
2026-01-09 19:37 ` Jason Gunthorpe
2 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 18:37 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> @@ -101,10 +101,14 @@ struct bnxt_re_pd_resp {
> struct bnxt_re_cq_req {
> __aligned_u64 cq_va;
> __aligned_u64 cq_handle;
> + __aligned_u64 comp_mask;
> + __u32 ncqe;
> + __u32 dmabuf_fd;
> };
> + __u32 pd_id;
> + __u32 dpi;
> + __u32 sq_dmabuf_fd;
> + __u32 sq_len; /* total len including MSN area */
> + __u32 sq_wqe_sz;
> + __u32 sq_psn_sz;
> + __u32 sq_npsn;
> + __u32 rq_dmabuf_fd;
All these fds are supposed to be __s32
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
2026-01-09 18:35 ` Jason Gunthorpe
@ 2026-01-09 18:50 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 18:55 ` Jason Gunthorpe
2026-01-09 19:01 ` Jason Gunthorpe
3 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 18:50 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> + ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
> + &dbr, sizeof(dbr));
> + if (ret)
> + goto free_entry;
> +
> + ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
> + &mmap_offset, sizeof(mmap_offset));
^^^^^^^^^^^^
> +DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_DBR_ALLOC,
> + UVERBS_ATTR_IDR(BNXT_RE_DV_ALLOC_DBR_HANDLE,
> + BNXT_RE_OBJECT_DBR,
> + UVERBS_ACCESS_NEW,
> + UA_MANDATORY),
> + UVERBS_ATTR_PTR_OUT(BNXT_RE_DV_ALLOC_DBR_ATTR,
> + UVERBS_ATTR_STRUCT(struct bnxt_re_dv_db_region,
> + umdbr),
> + UA_MANDATORY),
> + UVERBS_ATTR_PTR_IN(BNXT_RE_DV_ALLOC_DBR_OFFSET,
^^^^^^^
ptr_in shuld not be used with copy_to
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
2026-01-09 18:35 ` Jason Gunthorpe
2026-01-09 18:50 ` Jason Gunthorpe
@ 2026-01-09 18:55 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 19:01 ` Jason Gunthorpe
3 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 18:55 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> +static int UVERBS_HANDLER(BNXT_RE_METHOD_DBR_ALLOC)(struct uverbs_attr_bundle *attrs)
> +{
> + struct bnxt_re_dv_db_region dbr = {};
> + struct bnxt_re_alloc_dbr_obj *obj;
> + struct bnxt_re_ucontext *uctx;
> + struct ib_ucontext *ib_uctx;
> + struct bnxt_qplib_dpi *dpi;
> + struct bnxt_re_dev *rdev;
> + struct ib_uobject *uobj;
> + u64 mmap_offset;
> + int ret;
> +
> + ib_uctx = ib_uverbs_get_ucontext(attrs);
> + if (IS_ERR(ib_uctx))
> + return PTR_ERR(ib_uctx);
> +
> + uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
> + rdev = uctx->rdev;
> + uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
> +
> + obj = kzalloc(sizeof(*obj), GFP_KERNEL);
> + if (!obj)
> + return -ENOMEM;
> +
> + dpi = &obj->dpi;
> + ret = bnxt_qplib_alloc_uc_dpi(&rdev->qplib_res, dpi);
> + if (ret)
> + goto free_mem;
> +
> + obj->entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_UC_DB,
> + &mmap_offset);
> + if (!obj->entry) {
> + ret = -ENOMEM;
> + goto free_dpi;
> + }
> +
> + obj->rdev = rdev;
> + dbr.umdbr = dpi->umdbr;
> + dbr.dpi = dpi->dpi;
> +
> + ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
> + &dbr, sizeof(dbr));
> + if (ret)
> + goto free_entry;
> +
> + ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
> + &mmap_offset, sizeof(mmap_offset));
> + if (ret)
> + goto free_entry;
> +
> + uobj->object = obj;
> + uverbs_finalize_uobj_create(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
> + hash_add(rdev->dpi_hash, &obj->hash_entry, dpi->dpi);
This is out of order, hash_del is done inside bnxt_re_dv_dbr_cleanup()
so it should be before finalize, but it isn't a bug.
It should probably be written like this:
uobj->object = obj;
hash_add(rdev->dpi_hash, &obj->hash_entry, dpi->dpi);
uverbs_finalize_uobj_create(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
&dbr, sizeof(dbr));
if (ret)
return ret;
ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
&mmap_offset, sizeof(mmap_offset));
if (ret)
return ret;
return 0;
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
` (2 preceding siblings ...)
2026-01-09 18:55 ` Jason Gunthorpe
@ 2026-01-09 19:01 ` Jason Gunthorpe
2026-01-13 17:14 ` Sriharsha Basavapatna
3 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 19:01 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> @@ -217,6 +218,7 @@ struct bnxt_re_dev {
> struct delayed_work dbq_pacing_work;
> DECLARE_HASHTABLE(cq_hash, MAX_CQ_HASH_BITS);
> DECLARE_HASHTABLE(srq_hash, MAX_SRQ_HASH_BITS);
> + DECLARE_HASHTABLE(dpi_hash, MAX_DPI_HASH_BITS);
> struct dentry *dbg_root;
> struct dentry *qp_debugfs;
> unsigned long event_bitmap;
hash tables require locking.
hash_for_each_possible(rdev->dpi_hash, tmp_obj, hash_entry, dpi) {
if (tmp_obj->dpi.dpi == dpi) {
obj = tmp_obj;
break;
}
}
vs
hash_del(&obj->hash_entry);
kfree(obj);
Has no lock I can see and is not safe without one.
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
2026-01-09 18:37 ` Jason Gunthorpe
@ 2026-01-09 19:08 ` Jason Gunthorpe
2026-01-13 17:14 ` Sriharsha Basavapatna
2026-01-09 19:37 ` Jason Gunthorpe
2 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 19:08 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> +static struct ib_umem *bnxt_re_dv_umem_get(struct bnxt_re_dev *rdev,
> + struct ib_ucontext *ib_uctx,
> + int dmabuf_fd,
> + u64 addr, u64 size,
> + struct bnxt_qplib_sg_info *sg)
> +{
> + int access = IB_ACCESS_LOCAL_WRITE;
> + struct ib_umem *umem;
> + int umem_pgs, rc;
> +
> + if (dmabuf_fd) {
> + struct ib_umem_dmabuf *umem_dmabuf;
> +
> + umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, addr, size,
> + dmabuf_fd, access);
> + if (IS_ERR(umem_dmabuf)) {
> + rc = PTR_ERR(umem_dmabuf);
> + goto umem_err;
> + }
> + umem = &umem_dmabuf->umem;
> + } else {
> + umem = ib_umem_get(&rdev->ibdev, addr, size, access);
> + if (IS_ERR(umem)) {
> + rc = PTR_ERR(umem);
> + goto umem_err;
> + }
> + }
> +
> + umem_pgs = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
I should never see PAGE_SIZE passed to dma_blocks, and you can't call
dma_blocks without previously calling ib_umem_find_best_pgsz() to
validate that the umem is compatible.
I assume you want to use SZ_4K here, as any sizing of the umem should
be derived from absolute hardware capability, never PAGE_SIZE.
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
2026-01-09 18:37 ` Jason Gunthorpe
2026-01-09 19:08 ` Jason Gunthorpe
@ 2026-01-09 19:37 ` Jason Gunthorpe
2026-01-13 17:15 ` Sriharsha Basavapatna
2 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-09 19:37 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> +static int bnxt_re_dv_create_qplib_cq(struct bnxt_re_dev *rdev,
> + struct bnxt_re_ucontext *re_uctx,
> + struct bnxt_re_cq *cq,
> + struct bnxt_re_cq_req *req)
> +{
> + struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr;
> + struct bnxt_qplib_cq *qplcq;
> + struct ib_umem *umem;
> + u32 cqe = req->ncqe;
> + u32 max_active_cqs;
> + int rc = -EINVAL;
> +
> + if (atomic_read(&rdev->stats.res.cq_count) >= dev_attr->max_cq) {
> + ibdev_dbg(&rdev->ibdev, "Create CQ failed - max exceeded(CQs)");
> + return rc;
> + }
This is a racy way to use atomics, this should be an
atomic_inc_return, then check the return code, doing a dec on all
error paths.
> +static void bnxt_re_dv_init_ib_cq(struct bnxt_re_dev *rdev,
> + struct bnxt_re_cq *re_cq)
> +{
> + struct ib_cq *ib_cq;
> +
> + ib_cq = &re_cq->ib_cq;
> + ib_cq->device = &rdev->ibdev;
> + ib_cq->uobject = NULL;
> + ib_cq->comp_handler = NULL;
> + ib_cq->event_handler = NULL;
> + atomic_set(&ib_cq->usecnt, 0);
> +}
All these should already be 0 since this was freshly zallocated, no?
> +int bnxt_re_dv_create_cq(struct bnxt_re_dev *rdev, struct ib_udata *udata,
> + struct bnxt_re_cq *re_cq, struct bnxt_re_cq_req *req)
> +{
> + struct bnxt_re_ucontext *re_uctx =
> + rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
> + struct bnxt_re_cq_resp resp = {};
> + int ret;
> +
> + ret = bnxt_re_dv_create_qplib_cq(rdev, re_uctx, re_cq, req);
> + if (ret)
> + return ret;
> +
> + ret = bnxt_re_dv_create_cq_resp(rdev, re_cq, &resp);
> + if (ret)
> + goto fail_resp;
> +
> + ret = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
> + if (ret)
> + goto fail_resp;
> +
> + bnxt_re_dv_init_ib_cq(rdev, re_cq);
> + re_cq->is_dv_cq = true;
> + atomic_inc(&rdev->dv_cq_count);
> + return 0;
> +
> +fail_resp:
> + bnxt_qplib_destroy_cq(&rdev->qplib_res, &re_cq->qplib_cq);
> + bnxt_re_put_nq(rdev, re_cq->qplib_cq.nq);
> + ib_umem_release(re_cq->umem);
This seems really weird error unwinding, I expect to see functions
with error unwinds that match the calls within the function.
So why doesn't bnxt_qplib_destroy_cq() do the umem release and
req_put_nq?
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2026-01-09 18:35 ` Jason Gunthorpe
@ 2026-01-13 17:13 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:13 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 353 bytes --]
On Sat, Jan 10, 2026 at 12:05 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> > +struct bnxt_re_dv_db_region {
> > + __u32 dpi;
> > + __aligned_u64 umdbr;
> > +};
>
> This needs a '__u32 reserved' to mark the implicit padding..
Ack.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-09 18:37 ` Jason Gunthorpe
@ 2026-01-13 17:13 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:13 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 733 bytes --]
On Sat, Jan 10, 2026 at 12:07 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> > @@ -101,10 +101,14 @@ struct bnxt_re_pd_resp {
> > struct bnxt_re_cq_req {
> > __aligned_u64 cq_va;
> > __aligned_u64 cq_handle;
> > + __aligned_u64 comp_mask;
> > + __u32 ncqe;
> > + __u32 dmabuf_fd;
> > };
>
> > + __u32 pd_id;
> > + __u32 dpi;
> > + __u32 sq_dmabuf_fd;
> > + __u32 sq_len; /* total len including MSN area */
> > + __u32 sq_wqe_sz;
> > + __u32 sq_psn_sz;
> > + __u32 sq_npsn;
> > + __u32 rq_dmabuf_fd;
>
> All these fds are supposed to be __s32
Ack.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2026-01-09 18:50 ` Jason Gunthorpe
@ 2026-01-13 17:13 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:13 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 1416 bytes --]
On Sat, Jan 10, 2026 at 12:20 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> > + ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
> > + &dbr, sizeof(dbr));
> > + if (ret)
> > + goto free_entry;
> > +
> > + ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
> > + &mmap_offset, sizeof(mmap_offset));
> ^^^^^^^^^^^^
>
>
> > +DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_DBR_ALLOC,
> > + UVERBS_ATTR_IDR(BNXT_RE_DV_ALLOC_DBR_HANDLE,
> > + BNXT_RE_OBJECT_DBR,
> > + UVERBS_ACCESS_NEW,
> > + UA_MANDATORY),
> > + UVERBS_ATTR_PTR_OUT(BNXT_RE_DV_ALLOC_DBR_ATTR,
> > + UVERBS_ATTR_STRUCT(struct bnxt_re_dv_db_region,
> > + umdbr),
> > + UA_MANDATORY),
> > + UVERBS_ATTR_PTR_IN(BNXT_RE_DV_ALLOC_DBR_OFFSET,
> ^^^^^^^
>
> ptr_in shuld not be used with copy_to
Ack.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2026-01-09 18:55 ` Jason Gunthorpe
@ 2026-01-13 17:13 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:13 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 2963 bytes --]
On Sat, Jan 10, 2026 at 12:25 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> > +static int UVERBS_HANDLER(BNXT_RE_METHOD_DBR_ALLOC)(struct uverbs_attr_bundle *attrs)
> > +{
> > + struct bnxt_re_dv_db_region dbr = {};
> > + struct bnxt_re_alloc_dbr_obj *obj;
> > + struct bnxt_re_ucontext *uctx;
> > + struct ib_ucontext *ib_uctx;
> > + struct bnxt_qplib_dpi *dpi;
> > + struct bnxt_re_dev *rdev;
> > + struct ib_uobject *uobj;
> > + u64 mmap_offset;
> > + int ret;
> > +
> > + ib_uctx = ib_uverbs_get_ucontext(attrs);
> > + if (IS_ERR(ib_uctx))
> > + return PTR_ERR(ib_uctx);
> > +
> > + uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
> > + rdev = uctx->rdev;
> > + uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
> > +
> > + obj = kzalloc(sizeof(*obj), GFP_KERNEL);
> > + if (!obj)
> > + return -ENOMEM;
> > +
> > + dpi = &obj->dpi;
> > + ret = bnxt_qplib_alloc_uc_dpi(&rdev->qplib_res, dpi);
> > + if (ret)
> > + goto free_mem;
> > +
> > + obj->entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_UC_DB,
> > + &mmap_offset);
> > + if (!obj->entry) {
> > + ret = -ENOMEM;
> > + goto free_dpi;
> > + }
> > +
> > + obj->rdev = rdev;
> > + dbr.umdbr = dpi->umdbr;
> > + dbr.dpi = dpi->dpi;
> > +
> > + ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
> > + &dbr, sizeof(dbr));
> > + if (ret)
> > + goto free_entry;
> > +
> > + ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
> > + &mmap_offset, sizeof(mmap_offset));
> > + if (ret)
> > + goto free_entry;
> > +
> > + uobj->object = obj;
> > + uverbs_finalize_uobj_create(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
> > + hash_add(rdev->dpi_hash, &obj->hash_entry, dpi->dpi);
>
> This is out of order, hash_del is done inside bnxt_re_dv_dbr_cleanup()
> so it should be before finalize, but it isn't a bug.
>
> It should probably be written like this:
>
> uobj->object = obj;
> hash_add(rdev->dpi_hash, &obj->hash_entry, dpi->dpi);
> uverbs_finalize_uobj_create(attrs, BNXT_RE_DV_ALLOC_DBR_HANDLE);
>
> ret = uverbs_copy_to_struct_or_zero(attrs, BNXT_RE_DV_ALLOC_DBR_ATTR,
> &dbr, sizeof(dbr));
> if (ret)
> return ret;
>
> ret = uverbs_copy_to(attrs, BNXT_RE_DV_ALLOC_DBR_OFFSET,
> &mmap_offset, sizeof(mmap_offset));
> if (ret)
> return ret;
> return 0;
Ack, reordered as above.
Thanks,
-Harsha
>
> Jason
>
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs
2026-01-09 19:01 ` Jason Gunthorpe
@ 2026-01-13 17:14 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:14 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 1053 bytes --]
On Sat, Jan 10, 2026 at 12:31 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:01AM +0530, Sriharsha Basavapatna wrote:
> > @@ -217,6 +218,7 @@ struct bnxt_re_dev {
> > struct delayed_work dbq_pacing_work;
> > DECLARE_HASHTABLE(cq_hash, MAX_CQ_HASH_BITS);
> > DECLARE_HASHTABLE(srq_hash, MAX_SRQ_HASH_BITS);
> > + DECLARE_HASHTABLE(dpi_hash, MAX_DPI_HASH_BITS);
> > struct dentry *dbg_root;
> > struct dentry *qp_debugfs;
> > unsigned long event_bitmap;
>
> hash tables require locking.
>
> hash_for_each_possible(rdev->dpi_hash, tmp_obj, hash_entry, dpi) {
> if (tmp_obj->dpi.dpi == dpi) {
> obj = tmp_obj;
> break;
> }
> }
>
> vs
>
> hash_del(&obj->hash_entry);
> kfree(obj);
>
> Has no lock I can see and is not safe without one.
Updated to use rcu + mutex.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-09 19:08 ` Jason Gunthorpe
@ 2026-01-13 17:14 ` Sriharsha Basavapatna
2026-01-13 17:27 ` Jason Gunthorpe
0 siblings, 1 reply; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:14 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 1798 bytes --]
On Sat, Jan 10, 2026 at 12:38 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> > +static struct ib_umem *bnxt_re_dv_umem_get(struct bnxt_re_dev *rdev,
> > + struct ib_ucontext *ib_uctx,
> > + int dmabuf_fd,
> > + u64 addr, u64 size,
> > + struct bnxt_qplib_sg_info *sg)
> > +{
> > + int access = IB_ACCESS_LOCAL_WRITE;
> > + struct ib_umem *umem;
> > + int umem_pgs, rc;
> > +
> > + if (dmabuf_fd) {
> > + struct ib_umem_dmabuf *umem_dmabuf;
> > +
> > + umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, addr, size,
> > + dmabuf_fd, access);
> > + if (IS_ERR(umem_dmabuf)) {
> > + rc = PTR_ERR(umem_dmabuf);
> > + goto umem_err;
> > + }
> > + umem = &umem_dmabuf->umem;
> > + } else {
> > + umem = ib_umem_get(&rdev->ibdev, addr, size, access);
> > + if (IS_ERR(umem)) {
> > + rc = PTR_ERR(umem);
> > + goto umem_err;
> > + }
> > + }
> > +
> > + umem_pgs = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
>
> I should never see PAGE_SIZE passed to dma_blocks, and you can't call
> dma_blocks without previously calling ib_umem_find_best_pgsz() to
> validate that the umem is compatible.
>
> I assume you want to use SZ_4K here, as any sizing of the umem should
> be derived from absolute hardware capability, never PAGE_SIZE.
Changed to use SZ_4K.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-09 19:37 ` Jason Gunthorpe
@ 2026-01-13 17:15 ` Sriharsha Basavapatna
2026-01-13 17:28 ` Jason Gunthorpe
0 siblings, 1 reply; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-13 17:15 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 3152 bytes --]
On Sat, Jan 10, 2026 at 1:07 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> > +static int bnxt_re_dv_create_qplib_cq(struct bnxt_re_dev *rdev,
> > + struct bnxt_re_ucontext *re_uctx,
> > + struct bnxt_re_cq *cq,
> > + struct bnxt_re_cq_req *req)
> > +{
> > + struct bnxt_qplib_dev_attr *dev_attr = rdev->dev_attr;
> > + struct bnxt_qplib_cq *qplcq;
> > + struct ib_umem *umem;
> > + u32 cqe = req->ncqe;
> > + u32 max_active_cqs;
> > + int rc = -EINVAL;
> > +
> > + if (atomic_read(&rdev->stats.res.cq_count) >= dev_attr->max_cq) {
> > + ibdev_dbg(&rdev->ibdev, "Create CQ failed - max exceeded(CQs)");
> > + return rc;
> > + }
>
> This is a racy way to use atomics, this should be an
> atomic_inc_return, then check the return code, doing a dec on all
> error paths.
Ack.
>
> > +static void bnxt_re_dv_init_ib_cq(struct bnxt_re_dev *rdev,
> > + struct bnxt_re_cq *re_cq)
> > +{
> > + struct ib_cq *ib_cq;
> > +
> > + ib_cq = &re_cq->ib_cq;
> > + ib_cq->device = &rdev->ibdev;
>
>
> > + ib_cq->uobject = NULL;
> > + ib_cq->comp_handler = NULL;
> > + ib_cq->event_handler = NULL;
> > + atomic_set(&ib_cq->usecnt, 0);
> > +}
>
> All these should already be 0 since this was freshly zallocated, no?
This function is not needed, deleted (design changed from direct
uverbs to driver specific uverbs extension).
>
> > +int bnxt_re_dv_create_cq(struct bnxt_re_dev *rdev, struct ib_udata *udata,
> > + struct bnxt_re_cq *re_cq, struct bnxt_re_cq_req *req)
> > +{
> > + struct bnxt_re_ucontext *re_uctx =
> > + rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
> > + struct bnxt_re_cq_resp resp = {};
> > + int ret;
> > +
> > + ret = bnxt_re_dv_create_qplib_cq(rdev, re_uctx, re_cq, req);
> > + if (ret)
> > + return ret;
> > +
> > + ret = bnxt_re_dv_create_cq_resp(rdev, re_cq, &resp);
> > + if (ret)
> > + goto fail_resp;
> > +
> > + ret = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
> > + if (ret)
> > + goto fail_resp;
> > +
> > + bnxt_re_dv_init_ib_cq(rdev, re_cq);
> > + re_cq->is_dv_cq = true;
> > + atomic_inc(&rdev->dv_cq_count);
> > + return 0;
> > +
> > +fail_resp:
> > + bnxt_qplib_destroy_cq(&rdev->qplib_res, &re_cq->qplib_cq);
> > + bnxt_re_put_nq(rdev, re_cq->qplib_cq.nq);
> > + ib_umem_release(re_cq->umem);
>
> This seems really weird error unwinding, I expect to see functions
> with error unwinds that match the calls within the function.
>
> So why doesn't bnxt_qplib_destroy_cq() do the umem release and
> req_put_nq?
Because bnxt_qplib_destroy_cq() deals with only the FW side and the
wrapper function bnxt_re_dv_create_qplib_cq() calls umem_get() and
get_nq().
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-13 17:14 ` Sriharsha Basavapatna
@ 2026-01-13 17:27 ` Jason Gunthorpe
2026-01-17 8:07 ` Sriharsha Basavapatna
0 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-13 17:27 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Tue, Jan 13, 2026 at 10:44:21PM +0530, Sriharsha Basavapatna wrote:
> On Sat, Jan 10, 2026 at 12:38 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
> >
> > On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> > > +static struct ib_umem *bnxt_re_dv_umem_get(struct bnxt_re_dev *rdev,
> > > + struct ib_ucontext *ib_uctx,
> > > + int dmabuf_fd,
> > > + u64 addr, u64 size,
> > > + struct bnxt_qplib_sg_info *sg)
> > > +{
> > > + int access = IB_ACCESS_LOCAL_WRITE;
> > > + struct ib_umem *umem;
> > > + int umem_pgs, rc;
> > > +
> > > + if (dmabuf_fd) {
> > > + struct ib_umem_dmabuf *umem_dmabuf;
> > > +
> > > + umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, addr, size,
> > > + dmabuf_fd, access);
> > > + if (IS_ERR(umem_dmabuf)) {
> > > + rc = PTR_ERR(umem_dmabuf);
> > > + goto umem_err;
> > > + }
> > > + umem = &umem_dmabuf->umem;
> > > + } else {
> > > + umem = ib_umem_get(&rdev->ibdev, addr, size, access);
> > > + if (IS_ERR(umem)) {
> > > + rc = PTR_ERR(umem);
> > > + goto umem_err;
> > > + }
> > > + }
> > > +
> > > + umem_pgs = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
> >
> > I should never see PAGE_SIZE passed to dma_blocks, and you can't call
> > dma_blocks without previously calling ib_umem_find_best_pgsz() to
> > validate that the umem is compatible.
> >
> > I assume you want to use SZ_4K here, as any sizing of the umem should
> > be derived from absolute hardware capability, never PAGE_SIZE.
> Changed to use SZ_4K.
You also MUST call ib_umem_find_best_pgsz()
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-13 17:15 ` Sriharsha Basavapatna
@ 2026-01-13 17:28 ` Jason Gunthorpe
2026-01-17 8:07 ` Sriharsha Basavapatna
0 siblings, 1 reply; 23+ messages in thread
From: Jason Gunthorpe @ 2026-01-13 17:28 UTC (permalink / raw)
To: Sriharsha Basavapatna
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil
On Tue, Jan 13, 2026 at 10:45:22PM +0530, Sriharsha Basavapatna wrote:
> > So why doesn't bnxt_qplib_destroy_cq() do the umem release and
> > req_put_nq?
> Because bnxt_qplib_destroy_cq() deals with only the FW side and the
> wrapper function bnxt_re_dv_create_qplib_cq() calls umem_get() and
> get_nq().
So make a wrapper function for destroy too..
Jason
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-13 17:27 ` Jason Gunthorpe
@ 2026-01-17 8:07 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-17 8:07 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 2165 bytes --]
On Tue, Jan 13, 2026 at 10:57 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Tue, Jan 13, 2026 at 10:44:21PM +0530, Sriharsha Basavapatna wrote:
> > On Sat, Jan 10, 2026 at 12:38 AM Jason Gunthorpe <jgg@ziepe.ca> wrote:
> > >
> > > On Wed, Dec 24, 2025 at 09:56:02AM +0530, Sriharsha Basavapatna wrote:
> > > > +static struct ib_umem *bnxt_re_dv_umem_get(struct bnxt_re_dev *rdev,
> > > > + struct ib_ucontext *ib_uctx,
> > > > + int dmabuf_fd,
> > > > + u64 addr, u64 size,
> > > > + struct bnxt_qplib_sg_info *sg)
> > > > +{
> > > > + int access = IB_ACCESS_LOCAL_WRITE;
> > > > + struct ib_umem *umem;
> > > > + int umem_pgs, rc;
> > > > +
> > > > + if (dmabuf_fd) {
> > > > + struct ib_umem_dmabuf *umem_dmabuf;
> > > > +
> > > > + umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, addr, size,
> > > > + dmabuf_fd, access);
> > > > + if (IS_ERR(umem_dmabuf)) {
> > > > + rc = PTR_ERR(umem_dmabuf);
> > > > + goto umem_err;
> > > > + }
> > > > + umem = &umem_dmabuf->umem;
> > > > + } else {
> > > > + umem = ib_umem_get(&rdev->ibdev, addr, size, access);
> > > > + if (IS_ERR(umem)) {
> > > > + rc = PTR_ERR(umem);
> > > > + goto umem_err;
> > > > + }
> > > > + }
> > > > +
> > > > + umem_pgs = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
> > >
> > > I should never see PAGE_SIZE passed to dma_blocks, and you can't call
> > > dma_blocks without previously calling ib_umem_find_best_pgsz() to
> > > validate that the umem is compatible.
> > >
> > > I assume you want to use SZ_4K here, as any sizing of the umem should
> > > be derived from absolute hardware capability, never PAGE_SIZE.
> > Changed to use SZ_4K.
>
> You also MUST call ib_umem_find_best_pgsz()
>
> Jason
Ack.
Thanks,
-Harsha
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
* Re: [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs
2026-01-13 17:28 ` Jason Gunthorpe
@ 2026-01-17 8:07 ` Sriharsha Basavapatna
0 siblings, 0 replies; 23+ messages in thread
From: Sriharsha Basavapatna @ 2026-01-17 8:07 UTC (permalink / raw)
To: Jason Gunthorpe
Cc: leon, linux-rdma, andrew.gospodarek, selvin.xavier,
kalesh-anakkur.purayil, Sriharsha Basavapatna
[-- Attachment #1: Type: text/plain, Size: 487 bytes --]
On Tue, Jan 13, 2026 at 10:58 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Tue, Jan 13, 2026 at 10:45:22PM +0530, Sriharsha Basavapatna wrote:
> > > So why doesn't bnxt_qplib_destroy_cq() do the umem release and
> > > req_put_nq?
> > Because bnxt_qplib_destroy_cq() deals with only the FW side and the
> > wrapper function bnxt_re_dv_create_qplib_cq() calls umem_get() and
> > get_nq().
>
> So make a wrapper function for destroy too..
Ack.
Thanks,
-Harsha
>
> Jason
[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 5505 bytes --]
^ permalink raw reply [flat|nested] 23+ messages in thread
end of thread, other threads:[~2026-01-17 8:07 UTC | newest]
Thread overview: 23+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-12-24 4:25 [PATCH rdma-next v6 0/4] RDMA/bnxt_re: Support direct verbs Sriharsha Basavapatna
2025-12-24 4:25 ` [PATCH rdma-next v6 1/4] RDMA/bnxt_re: Move the UAPI methods to a dedicated file Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 2/4] RDMA/bnxt_re: Refactor bnxt_qplib_create_qp() function Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 3/4] RDMA/bnxt_re: Direct Verbs: Support DBR verbs Sriharsha Basavapatna
2026-01-09 18:35 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 18:50 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 18:55 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 19:01 ` Jason Gunthorpe
2026-01-13 17:14 ` Sriharsha Basavapatna
2025-12-24 4:26 ` [PATCH rdma-next v6 4/4] RDMA/bnxt_re: Direct Verbs: Support CQ and QP verbs Sriharsha Basavapatna
2026-01-09 18:37 ` Jason Gunthorpe
2026-01-13 17:13 ` Sriharsha Basavapatna
2026-01-09 19:08 ` Jason Gunthorpe
2026-01-13 17:14 ` Sriharsha Basavapatna
2026-01-13 17:27 ` Jason Gunthorpe
2026-01-17 8:07 ` Sriharsha Basavapatna
2026-01-09 19:37 ` Jason Gunthorpe
2026-01-13 17:15 ` Sriharsha Basavapatna
2026-01-13 17:28 ` Jason Gunthorpe
2026-01-17 8:07 ` Sriharsha Basavapatna
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox