Skip to contentSkip to conversation

RDMA: Extend uverbs umem support for QP buffers and doorbell records

PATCH In review Jiri Pirko 4 participants 34 messages Feb 3 – Feb 4

Subsystems (3)
lore
Share & tools
lore

Jiri Pirko

author
#0 lore raw

cover RDMA: Extend uverbs umem support for QP buffers and doorbell records

From: Jiri Pirko <[email protected]>

This patchset extends the existing CQ umem buffer infrastructure to QP creation and adds doorbell record (DBR) umem support for both CQ and QP.

The kernel already supports CQ creation with externally provided umem buffers. This patchset:

1. Adds umem reference counting to simplify memory lifecycle management.
2. Adds mlx5 driver implementation for the existing CQ buffer umem support.
3. Extends QP creation to support external SQ/RQ buffer umems.
4. Adds doorbell record umem support for both CQ and QP.
5. Implements all extensions in mlx5 driver.

This enables integration with dmabuf exporters for specialized memory types, such as GPU memory for GPU-direct RDMA scenarios.

Patch #1 introduces reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.
Patch #2 adopts the new ib_umem refcounting model for create_cq_umem. Patch #3 implements mlx5 driver support for the existing CQ buffer umem interface.
Patch #4 factors out common buffer umem parsing into a shared helper. Patch #5 adds core framework for QP buffer umem support with new uverbs attributes for SQ/RQ buffers.
Patch #6 implements mlx5 driver support for QP buffer umem extension. Patch #7 extends create_cq_umem device op with DBR umem parameter and adds new uverbs attributes for CQ DBR.
Patch #8 implements mlx5 driver support for CQ DBR umem and extends the doorbell mapping infrastructure.
Patch #9 extends create_qp_umem device op with DBR umem parameters and adds new uverbs attributes for SQ/RQ DBR.
Patch #10 implements mlx5 driver support for QP DBR umem.

Jiri Pirko (10):
RDMA/umem: Add reference counting to ib_umem
RDMA/uverbs: Use umem refcounting for CQ creation with external buffer RDMA/mlx5: Add support for CQ creation with external umem buffer RDMA/uverbs: Factor out common buffer umem parsing into helper RDMA/core: Add support for QP buffer umem in QP creation RDMA/mlx5: Add support for QP creation with external umem buffers RDMA/uverbs: Add doorbell record umem support to CQ creation RDMA/mlx5: Add external doorbell record umem support for CQ RDMA/uverbs: Add doorbell record umem support to QP creation RDMA/mlx5: Add external doorbell record umem support for QP

 drivers/infiniband/core/core_priv.h           |   9 ++
 drivers/infiniband/core/device.c              |   1 +
 drivers/infiniband/core/umem.c                |   5 +
 drivers/infiniband/core/umem_dmabuf.c         |   1 +
 drivers/infiniband/core/umem_odp.c            |   3 +
 drivers/infiniband/core/uverbs_ioctl.c        |  68 +++++++++++
 drivers/infiniband/core/uverbs_std_types_cq.c | 104 +++++++---------
 drivers/infiniband/core/uverbs_std_types_qp.c | 114 +++++++++++++++++-
 drivers/infiniband/core/verbs.c               |  72 +++++++++--
 drivers/infiniband/hw/efa/efa.h               |   3 +-
 drivers/infiniband/hw/efa/efa_verbs.c         |  10 +-
 drivers/infiniband/hw/mlx5/cq.c               |  47 ++++++--
 drivers/infiniband/hw/mlx5/doorbell.c         |  27 ++++-
 drivers/infiniband/hw/mlx5/main.c             |   2 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |   9 +-
 drivers/infiniband/hw/mlx5/qp.c               |  93 +++++++++++---
 drivers/infiniband/hw/mlx5/srq.c              |   2 +-
 include/rdma/ib_umem.h                        |   9 ++
 include/rdma/ib_verbs.h                       |   7 ++
 include/rdma/uverbs_ioctl.h                   |   7 ++
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  20 +++
 21 files changed, 500 insertions(+), 113 deletions(-)
Signature · 1 line
2.51.1
author
#1 lore raw

PATCH 1/10 RDMA/umem: Add reference counting to ib_umem

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
#2 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 11 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

This sentence doesn't align with the proposed change.

Quoted 9 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful information, particularly the auto‑generated summary at the end.

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
author
#3 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 7 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 6 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

Quoted 7 lines
Quoted 5 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

Quoted 8 lines
Quoted 8 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :) I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
#4 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Quoted 11 lines
Quoted 8 lines
Quoted 6 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

The above paragraphs have clear AI pattern as they don't explain why, but only how.

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
author
#5 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle. Consider various error paths in the calle some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Quoted 25 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 2 lines

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Quoted 139 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
#6 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 30 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 13 lines
Quoted 12 lines
Quoted 9 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

Quoted 8 lines
Quoted 3 lines

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Quoted 7 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

Quoted 27 lines
Quoted 27 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 5 lines
Quoted 2 lines

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Commit message should provide an additional information, which is not available in the code itself. Description like "Core allocates umem with refcount=1" has zero value.

Thanks

Quoted 141 lines
Quoted 141 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
author
#7 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 35 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 34 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 16 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Quoted 7 lines
Quoted 4 lines

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

Quoted 42 lines
Quoted 8 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

Quoted 30 lines
Quoted 29 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Quoted 6 lines
Quoted 2 lines

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Commit message should provide an additional information, which is not available in the code itself. Description like "Core allocates umem with refcount=1" has zero value.

:), sure

Quoted 144 lines

Thanks

Quoted 141 lines
Quoted 141 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
#8 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 41 lines · Jiri Pirko
Quoted 40 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 38 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

Quoted 4 lines
Quoted 2 lines

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

It depends on your timeline. We are in -rc8 right now, so at least for next 3 weeks (one week till 6.19 + two weeks merge window) from now, your series won't be applied.

Thanks

Quoted 198 lines
Quoted 44 lines
Quoted 10 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

Quoted 30 lines
Quoted 29 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Quoted 6 lines
Quoted 2 lines

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Commit message should provide an additional information, which is not available in the code itself. Description like "Core allocates umem with refcount=1" has zero value.

:), sure

Quoted 144 lines

Thanks

Quoted 141 lines
Quoted 141 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
author
#9 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 42 lines · Tue, Feb 03, 2026 at 02:32:28PM +0100, [email protected] wrote:
Quoted 41 lines · Jiri Pirko
Quoted 40 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 38 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

It depends on your timeline. We are in -rc8 right now, so at least for next 3 weeks (one week till 6.19 + two weeks merge window) from now, your series won't be applied.

Well, I don't think I have a different option, do I :)

Quoted 201 lines

Thanks

Quoted 198 lines
Quoted 44 lines
Quoted 10 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

Quoted 30 lines
Quoted 29 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Quoted 6 lines
Quoted 2 lines

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Commit message should provide an additional information, which is not available in the code itself. Description like "Core allocates umem with refcount=1" has zero value.

:), sure

Quoted 144 lines

Thanks

Quoted 141 lines
Quoted 141 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (!refcount_dec_and_test(&umem->refcount))
+		return;
 	if (umem->is_dmabuf)
 		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
 	}
 
 	umem = &umem_dmabuf->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->length = size;
 	umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
 	if (!odp_data)
 		return ERR_PTR(-ENOMEM);
 	umem = &odp_data->umem;
+	refcount_set(&umem->refcount, 1);
 	umem->ibdev = root->umem.ibdev;
 	umem->length     = size;
 	umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
+	refcount_set(&umem_odp->umem.refcount, 1);
 	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -0
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
 	struct ib_device       *ibdev;
 	struct mm_struct       *owning_mm;
+	refcount_t		refcount;
 	u64 iova;
 	size_t			length;
 	unsigned long		address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+	refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
 	return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		      		    size_t length) {
Signature · 1 line
2.51.1
#10 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

In reply to Jiri Pirko
Quoted 42 lines · Jiri Pirko
Quoted 40 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 38 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

Quoted 17 lines
Quoted 2 lines

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

Quoted 11 lines
Quoted 8 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

I'm seeing a kernel crash when perftest is stopped, I'm still debugging it, but I'm wondering if it is related to this change. I see this inline comment in the uverbs handler:

        /* Driver took a reference, release ours */
        ib_umem_release(rq_dbr_umem);
        ib_umem_release(sq_dbr_umem);
        ib_umem_release(rq_umem);
        ib_umem_release(sq_umem);

What does it mean by "Driver took a reference"? If the driver returns success from create_qp_umem(), the umem it is still using gets released above? Is there something that the driver should call to hold a reference? It is not obvious from the create_qp_umem() dev op.

Quoted 185 lines
Quoted 31 lines
Quoted 31 lines
Quoted 30 lines
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
Change-Id Ifb1765ea3b14dab3329294633ea5df063c74420d

Please remove the Change-Ids and write the commit message yourself, without relying on AI. The current message provides no meaningful

I'm new in RDMA. Not sure what you mean by meaningful information :)

This part of commit message is clearly generated by Cursor:
With reference counting:
- Core allocates umem with refcount=1
- Driver calls ib_umem_get_ref() to take a reference
- Both layers can unconditionally call ib_umem_release()
- The umem is freed only when the last reference is dropped

The above paragraphs have clear AI pattern as they don't explain why, but only how.

Quoted 6 lines
Quoted 2 lines

Why is explained above, isn't it?
If you don't want the "how part", I can remove it, no problem.

Commit message should provide an additional information, which is not available in the code itself. Description like "Core allocates umem with refcount=1" has zero value.

:), sure

Quoted 144 lines

Thanks

Quoted 141 lines
Quoted 141 lines

I'm seeing the SAME commit messages in many internals and externals patches.

Even my AI review is agreed with me :)
...
"AI-authorship-score": "medium"
...

Quoted 130 lines

I'm always trying to provide it.

Quoted 1 line

information, particularly the auto‑generated summary at the end.

Doh, the changeIDs :) Sorry about that.

Quoted 123 lines

Thanks

Quoted 120 lines
---
 drivers/infiniband/core/umem.c        | 5 +++++
 drivers/infiniband/core/umem_dmabuf.c | 1 +
 drivers/infiniband/core/umem_odp.c    | 3 +++
 include/rdma/ib_umem.h                | 9 +++++++++
 4 files changed, 18 insertions(+)
drivers/infiniband/core/umem.c +5 -0
index 8137031c2a65..09ce694d66ea 100644
@@ -192,6 +192,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
       umem = kzalloc(sizeof(*umem), GFP_KERNEL);
       if (!umem)
               return ERR_PTR(-ENOMEM);
+      refcount_set(&umem->refcount, 1);
       umem->ibdev      = device;
       umem->length     = size;
       umem->address    = addr;
@@ -280,11 +281,15 @@ EXPORT_SYMBOL(ib_umem_get);
 /**
  * ib_umem_release - release memory pinned with ib_umem_get
  * @umem: umem struct to release
+ *
+ * Decrement the reference count and free the umem when it reaches zero.
  */
 void ib_umem_release(struct ib_umem *umem)
 {
       if (!umem)
               return;
+      if (!refcount_dec_and_test(&umem->refcount))
+              return;
       if (umem->is_dmabuf)
               return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
       if (umem->is_odp)
drivers/infiniband/core/umem_dmabuf.c +1 -0
index 939da49b0dcc..5c5286092fca 100644
@@ -143,6 +143,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,
       }
 
       umem = &umem_dmabuf->umem;
+      refcount_set(&umem->refcount, 1);
       umem->ibdev = device;
       umem->length = size;
       umem->address = offset;
drivers/infiniband/core/umem_odp.c +3 -0
index 572a91a62a7b..7be30fda57e3 100644
@@ -144,6 +144,7 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
       if (!umem_odp)
               return ERR_PTR(-ENOMEM);
       umem = &umem_odp->umem;
+      refcount_set(&umem->refcount, 1);
       umem->ibdev = device;
       umem->writable = ib_access_writable(access);
       umem->owning_mm = current->mm;
@@ -185,6 +186,7 @@ ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
       if (!odp_data)
               return ERR_PTR(-ENOMEM);
       umem = &odp_data->umem;
+      refcount_set(&umem->refcount, 1);
       umem->ibdev = root->umem.ibdev;
       umem->length     = size;
       umem->address    = addr;
@@ -245,6 +247,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
       if (!umem_odp)
               return ERR_PTR(-ENOMEM);
 
+      refcount_set(&umem_odp->umem.refcount, 1);
       umem_odp->umem.ibdev = device;
       umem_odp->umem.length = size;
       umem_odp->umem.address = addr;
include/rdma/ib_umem.h +9 -1
index 0a8e092c0ea8..44264f78eab3 100644
@@ -10,6 +10,7 @@
 #include <linux/list.h>
 #include <linux/scatterlist.h>
 #include <linux/workqueue.h>
+#include <linux/refcount.h>
 #include <rdma/ib_verbs.h>
 
 struct ib_ucontext;
@@ -19,6 +20,7 @@ struct dma_buf_attach_ops;
 struct ib_umem {
       struct ib_device       *ibdev;
       struct mm_struct       *owning_mm;
+      refcount_t              refcount;
       u64 iova;
       size_t                  length;
       unsigned long           address;
@@ -110,6 +112,12 @@ static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
 
 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
                           size_t size, int access);
+
+static inline void ib_umem_get_ref(struct ib_umem *umem)
+{
+      refcount_inc(&umem->refcount);
+}
+
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
                     size_t length);
@@ -188,6 +196,7 @@ static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 {
       return ERR_PTR(-EOPNOTSUPP);
 }
+static inline void ib_umem_get_ref(struct ib_umem *umem) { }
 static inline void ib_umem_release(struct ib_umem *umem) { }
 static inline int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
                                   size_t length) {
--

2.51.1

author
#11 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 71 lines · Tue, Feb 03, 2026 at 02:49:33PM +0100, [email protected] wrote:
Quoted 42 lines · Jiri Pirko
Quoted 40 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 38 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

Quoted 17 lines
Quoted 2 lines

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

Quoted 11 lines
Quoted 8 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

I'm seeing a kernel crash when perftest is stopped, I'm still debugging it, but I'm wondering if it is related to this change. I see this inline comment in the uverbs handler:

       /* Driver took a reference, release ours */
       ib_umem_release(rq_dbr_umem);
       ib_umem_release(sq_dbr_umem);
       ib_umem_release(rq_umem);
       ib_umem_release(sq_umem);
Quoted 4 lines

What does it mean by "Driver took a reference"? If the driver returns success from create_qp_umem(), the umem it is still using gets released above? Is there something that the driver should call to hold a reference? It is not obvious from the create_qp_umem() dev op.

Yes, see "RDMA/uverbs: Use umem refcounting for CQ creation with external buffer"

[...]

Quoted 1 line
#12 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 78 lines · Jiri Pirko
Quoted 71 lines · Tue, Feb 03, 2026 at 02:49:33PM +0100, [email protected] wrote:
Quoted 42 lines · Jiri Pirko
Quoted 40 lines · Tue, Feb 03, 2026 at 02:03:35PM +0100, [email protected] wrote:
Quoted 38 lines · Jiri Pirko
Quoted 11 lines · Tue, Feb 03, 2026 at 01:26:18PM +0100, [email protected] wrote:
Quoted 7 lines · Jiri Pirko
Quoted 6 lines · Tue, Feb 03, 2026 at 11:03:27AM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

The lifecycle should be confined either to the core or to the driver, but it should not mix responsibilities across both.

Okay, I can re-phrase. The point is, the last holding reference actually does the release.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

When the core RDMA layer allocates an ib_umem and passes it to a driver (e.g., for CQ or QP creation with external buffers), both layers need to manage the umem lifecycle. Without reference counting, this requires complex conditional release logic to avoid double-frees on error paths and leaks on success paths.

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

This sentence doesn't align with the proposed change.

Hmm, I'm not sure why you think it does not align. It exactly describes the code I had it originally without this path in place :)

There is no complex conditional release logic which this reference counting solves. That counting allows delayed, semi-random release, nothing more.

Again. Without the refcount, you have to make sure the umem is consumed by the op. Actually, check the existing create_cq_umem. On the error path, the umem is released by the caller. On success path the ownership is transferred to the calle.

Quoted 17 lines
Quoted 2 lines

We need to fix it. Exactly right now, I'm working to make sure that umem is managed by ib_core and drivers don't call to ib_umem_get() at all.

Should I wait and rebase?

Quoted 11 lines
Quoted 8 lines

Consider various error paths in the calle
some of which are shared with destroy_cq op, some umems are not actually used etc, it's a nightmare. I had the code written down like this originally, that's why I actually know.

Perhaps I'm missing your point. Is is just about the patch descriptio or about the code itself?

Description and the code. UMEM needs to be created by ib_core and ib_core should destroy them.

I'm seeing a kernel crash when perftest is stopped, I'm still debugging it, but I'm wondering if it is related to this change. I see this inline comment in the uverbs handler:

       /* Driver took a reference, release ours */
       ib_umem_release(rq_dbr_umem);
       ib_umem_release(sq_dbr_umem);
       ib_umem_release(rq_umem);
       ib_umem_release(sq_umem);
Quoted 4 lines

What does it mean by "Driver took a reference"? If the driver returns success from create_qp_umem(), the umem it is still using gets released above? Is there something that the driver should call to hold a reference? It is not obvious from the create_qp_umem() dev op.

Quoted 4 lines

Yes, see "RDMA/uverbs: Use umem refcounting for CQ creation with external buffer"

[...]

Ok, added ib_umem_get_ref() before returning (success) from create_cq/qp_umem() and it doesn't crash anymore.

Quoted 2 lines
Quoted 2 lines
#13 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

In reply to Jiri Pirko
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

  if (ops->create_cq_umem) {
     umem = umem_get
     rc = ops->create_cq_umem(umem)
     if (rc)
      umem_free(umem)
  } else {
     rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

destroy_cq()

     destry_hw
     umem_free()

This basically moves all the working code in the driver to the driver_create_cq_umem() which *always* gets a umem as a parameter.

If the user uses the drvdata path to specify the umem then the driver helper create_cq() creates the umem from the drvdata parameters, otherwise the core creates it from the common UATTRs.

We can keep things so the umem is always freed by the driver on success, which doesn't require any driver changes.

It should never be "shared", this is just a very simple unwind on error kind of pattern.

I think the challenge here is to unwind the drivers into the above three functions so they don't have a mess of convoluted error handling around the umem.

Jason

author
#14 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 31 lines · Tue, Feb 03, 2026 at 03:51:38PM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

 if (ops->create_cq_umem) {
    umem = umem_get
    rc = ops->create_cq_umem(umem)
    if (rc)
     umem_free(umem)
 } else {
    rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

Quoted 7 lines

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

destroy_cq()

    destry_hw
    umem_free()

This is how it is now. However there are couple of challenges about this flow:
1) umem usage. For example, create_qp_umem at the end of the set gets 4 umem pointers. sq,rq,sq_dbr,rq_dbr. Some driver may use only one of those, 2 of those, 3 of those. Depends. mlx5 actually uses 2 or 3. If what you suggest, the current approach stands, the user has to always take all pointers, store them and eventually release them on destroy_qp path.
2) error path. I found the error path quite odd. Then create_cq/qp_umem returns !=0, core releases all umems. However, standard cq/qp destroy path takes care of releasing umems. Since a lot of code on error path and destroy path is shared, it has to be informed to release or not release the umems. That is not nice.

That is why I tried to make this more elegant using reference, which the driver takes when it takes responsibility of the release.
I admit that if the umem_get/release will not be possible to do in drivers, just core, that would solve this issue in the most elegant way. That I why I would like to wait on Leon to do that and rebase on top of it.

Quoted 8 lines

This basically moves all the working code in the driver to the driver_create_cq_umem() which *always* gets a umem as a parameter.

If the user uses the drvdata path to specify the umem then the driver helper create_cq() creates the umem from the drvdata parameters, otherwise the core creates it from the common UATTRs.

Quoted 5 lines

We can keep things so the umem is always freed by the driver on success, which doesn't require any driver changes.

It should never be "shared", this is just a very simple unwind on error kind of pattern.

Not "shared", more like "taken over".

Quoted 6 lines

I think the challenge here is to unwind the drivers into the above three functions so they don't have a mess of convoluted error handling around the umem.

Jason

#15 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 41 lines · Jiri Pirko
Quoted 31 lines · Tue, Feb 03, 2026 at 03:51:38PM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

 if (ops->create_cq_umem) {
    umem = umem_get
    rc = ops->create_cq_umem(umem)
    if (rc)
     umem_free(umem)
 } else {
    rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

Quoted 7 lines

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

destroy_cq()

    destry_hw
    umem_free()
Quoted 8 lines

This is how it is now. However there are couple of challenges about this flow:
1) umem usage. For example, create_qp_umem at the end of the set gets 4 umem pointers. sq,rq,sq_dbr,rq_dbr. Some driver may use only one of those, 2 of those, 3 of those. Depends. mlx5 actually uses 2 or 3. If what you suggest, the current approach stands, the user has to always take all pointers, store them and eventually release them on destroy_qp path.

Userspace passing umems that are not used by the driver is an error. Fail the call.

Quoted 5 lines

2) error path. I found the error path quite odd. Then create_cq/qp_umem returns !=0, core releases all umems. However, standard cq/qp destroy path takes care of releasing umems. Since a lot of code on error path and destroy path is shared, it has to be informed to release or not release the umems. That is not nice.

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

My issue with refcounts is that this isn't a refcounted structure, it has very well defined points where it must become freed.

Like we can't get through detroy_qp without the umem being freed, that is illegal.

Jason

author
#16 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 51 lines · Tue, Feb 03, 2026 at 05:59:38PM +0100, [email protected] wrote:
Quoted 41 lines · Jiri Pirko
Quoted 31 lines · Tue, Feb 03, 2026 at 03:51:38PM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

 if (ops->create_cq_umem) {
    umem = umem_get
    rc = ops->create_cq_umem(umem)
    if (rc)
     umem_free(umem)
 } else {
    rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

Quoted 7 lines

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

destroy_cq()

    destry_hw
    umem_free()
Quoted 8 lines

This is how it is now. However there are couple of challenges about this flow:
1) umem usage. For example, create_qp_umem at the end of the set gets 4 umem pointers. sq,rq,sq_dbr,rq_dbr. Some driver may use only one of those, 2 of those, 3 of those. Depends. mlx5 actually uses 2 or 3. If what you suggest, the current approach stands, the user has to always take all pointers, store them and eventually release them on destroy_qp path.

Quoted 2 lines

Userspace passing umems that are not used by the driver is an error. Fail the call.

Okay, that makes sense. I just wanted to ignore it :)

Quoted 11 lines
Quoted 6 lines

2) error path. I found the error path quite odd. Then create_cq/qp_umem returns !=0, core releases all umems. However, standard cq/qp destroy path takes care of releasing umems. Since a lot of code on error path and destroy path is shared, it has to be informed to release or not release the umems. That is not nice.

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

Exactly.

Quoted 3 lines

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

Yeah, I wanted to avoid that. But sure.

Quoted 6 lines

My issue with refcounts is that this isn't a refcounted structure, it has very well defined points where it must become freed.

Like we can't get through detroy_qp without the umem being freed, that is illegal.

Sure, even with this patch that never happens. But I get your point.

Quoted 2 lines

Jason

author
#17 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

In reply to Jason Gunthorpe
Quoted 60 lines · Tue, Feb 03, 2026 at 05:59:38PM +0100, [email protected] wrote:
Quoted 41 lines · Jiri Pirko
Quoted 31 lines · Tue, Feb 03, 2026 at 03:51:38PM +0100, [email protected] wrote:
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

 if (ops->create_cq_umem) {
    umem = umem_get
    rc = ops->create_cq_umem(umem)
    if (rc)
     umem_free(umem)
 } else {
    rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

Quoted 7 lines

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

destroy_cq()

    destry_hw
    umem_free()
Quoted 8 lines

This is how it is now. However there are couple of challenges about this flow:
1) umem usage. For example, create_qp_umem at the end of the set gets 4 umem pointers. sq,rq,sq_dbr,rq_dbr. Some driver may use only one of those, 2 of those, 3 of those. Depends. mlx5 actually uses 2 or 3. If what you suggest, the current approach stands, the user has to always take all pointers, store them and eventually release them on destroy_qp path.

Userspace passing umems that are not used by the driver is an error. Fail the call.

Quoted 5 lines

2) error path. I found the error path quite odd. Then create_cq/qp_umem returns !=0, core releases all umems. However, standard cq/qp destroy path takes care of releasing umems. Since a lot of code on error path and destroy path is shared, it has to be informed to release or not release the umems. That is not nice.

Quoted 7 lines

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

How about we have:
int (*create_cq_umem)(struct ib_cq *cq,
const struct ib_cq_init_attr *attr, struct ib_umem **umem,
^^

struct uverbs_attr_bundle *attrs);

And instead of taking ref in the callee we just do *umem = NULL? :S

This would help to cover the error path vs destroy path differences, Tt could also be used to make sure the op consumed all umems; all should be NULLed on success.

Makes sense?

Quoted 8 lines

My issue with refcounts is that this isn't a refcounted structure, it has very well defined points where it must become freed.

Like we can't get through detroy_qp without the umem being freed, that is illegal.

Jason

#18 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 16 lines · Jiri Pirko
Quoted 7 lines

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

How about we have:
int (*create_cq_umem)(struct ib_cq *cq,
const struct ib_cq_init_attr *attr, struct ib_umem **umem,
^^

struct uverbs_attr_bundle *attrs);

Quoted 7 lines

And instead of taking ref in the callee we just do *umem = NULL? :S

This would help to cover the error path vs destroy path differences, Tt could also be used to make sure the op consumed all umems; all should be NULLed on success.

Makes sense?

I'm ok with that, though never seen the pattern before. If the core code fails on !NULL it would be pretty hard to use it wrong.

Jason

#19 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 20 lines · Jason Gunthorpe
Quoted 18 lines · Jiri Pirko
Quoted 7 lines

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

How about we have:
int (*create_cq_umem)(struct ib_cq *cq,
const struct ib_cq_init_attr *attr, struct ib_umem **umem,
^^

struct uverbs_attr_bundle *attrs);

And instead of taking ref in the callee we just do *umem = NULL? :S

Quoted 8 lines
Quoted 5 lines

This would help to cover the error path vs destroy path differences, Tt could also be used to make sure the op consumed all umems; all should be NULLed on success.

Makes sense?

I'm ok with that, though never seen the pattern before. If the core code fails on !NULL it would be pretty hard to use it wrong.

I'm less excited to see unique coding patterns. We should rather invest time in effort to layer everything correctly.

Thanks

Quoted 3 lines

Jason

author
#20 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 25 lines · Wed, Feb 04, 2026 at 06:54:11PM +0100, [email protected] wrote:
Quoted 24 lines · Jason Gunthorpe
Quoted 22 lines · Jiri Pirko
Quoted 7 lines

Generally I would not assign to the driver's umem storage until the creation is completed to avoid this. ie it stays null until committed.

But looking at mlx5 that looks like quite a maze there.. Yikes..

So maybe mlx5 adds some NULL assignments on its error paths and less convoluted drivers can use a simpler option?

How about we have:
int (*create_cq_umem)(struct ib_cq *cq,
const struct ib_cq_init_attr *attr, struct ib_umem **umem,
^^

struct uverbs_attr_bundle *attrs);

And instead of taking ref in the callee we just do *umem = NULL? :S

This would help to cover the error path vs destroy path differences, Tt could also be used to make sure the op consumed all umems; all should be NULLed on success.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Makes sense?

I'm ok with that, though never seen the pattern before. If the core code fails on !NULL it would be pretty hard to use it wrong.

I'm less excited to see unique coding patterns. We should rather invest time in effort to layer everything correctly.

Easy to leave this pattern after the effort you refer to brings fruits :)

Quoted 6 lines

Thanks

Quoted 3 lines

Jason

#21 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

In reply to Jason Gunthorpe
Quoted 49 lines · Jason Gunthorpe
Quoted 5 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce reference counting for ib_umem objects to simplify memory lifecycle management when umem buffers are shared between the core layer and device drivers.

I admit I have reservations about this too.. The flow should not be so convoluted that a refcount is necessary. The lifecycle of a umem is not uncertain at all.

I imagine'd it would be like:

core code:

  if (ops->create_cq_umem) {
     umem = umem_get
     rc = ops->create_cq_umem(umem)
     if (rc)
      umem_free(umem)
  } else {
     rc = ops->create_cq()

}

Driver:
create_cq():
copy_from_user(drvdata)
umem = umem_get()
rc = driver_create_cq_umem(umem, &drvdata))
if (rc)
umem_free(umem)

create_cq_umem()
copy_from_user(drvdata)
return driver_create_cq_umem(umem, &drvdata)

   destroy_cq()
     destry_hw
     umem_free()

This basically moves all the working code in the driver to the driver_create_cq_umem() which *always* gets a umem as a parameter.

If the user uses the drvdata path to specify the umem then the driver helper create_cq() creates the umem from the drvdata parameters, otherwise the core creates it from the common UATTRs.

We can keep things so the umem is always freed by the driver on success, which doesn't require any driver changes.

Quoted 6 lines

It should never be "shared", this is just a very simple unwind on error kind of pattern.

I think the challenge here is to unwind the drivers into the above three functions so they don't have a mess of convoluted error handling around the umem.

I opted for an even simpler approach: embed the umem directly in ib_cq, set cq->umem in ib_core’s create_cq, and clean it up in ib_core’s destroy_cq.

The beginning of the series is available here:
https://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma.git/log/?h=refactor-umem-v1

Thanks

Quoted 2 lines

Jason

#22 lore raw

Re: [PATCH rdma-next 01/10] RDMA/umem: Add reference counting to ib_umem

Quoted 2 lines · Leon Romanovsky

I opted for an even simpler approach: embed the umem directly in ib_cq, set cq->umem in ib_core’s create_cq, and clean it up in ib_core’s destroy_cq.

The issue is most drivers create the umem from parameters in their driver data so the core code cannot do it at all.

The structure I gave is a way for the drivers to parse their driver data while still keeping the umem out of the bowels of the driver.

You can lift the umems into the core structures too, but things like the DBR with their driver-specific umems would be hard to deal with.

Jason

author
#23 lore raw

PATCH 2/10 RDMA/uverbs: Use umem refcounting for CQ creation with external buffer

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Adopt the new ib_umem refcounting model for create_cq_umem and the only op implementation.

Change-Id Ifad1447a9cd688554f2eed3d41e96d62fc28ed54
---
 drivers/infiniband/core/uverbs_std_types_cq.c | 3 +++
 drivers/infiniband/hw/efa/efa_verbs.c         | 5 ++++-
 2 files changed, 7 insertions(+), 1 deletion(-)
drivers/infiniband/core/uverbs_std_types_cq.c +3 -0
index fab5d914029d..380e4e4d6fb1 100644
@@ -196,6 +196,9 @@ static int UVERBS_HANDLER(UVERBS_METHOD_CQ_CREATE)(
 	if (ret)
 		goto err_free;
 
+	/* Driver took a reference, release ours */
+	ib_umem_release(umem);
+
 	obj->uevent.uobject.object = cq;
 	obj->uevent.uobject.user_handle = user_handle;
 	rdma_restrack_add(&cq->res);
drivers/infiniband/hw/efa/efa_verbs.c +4 -1
index 22d3e25c3b9d..aca48825cd71 100644
@@ -1225,6 +1225,7 @@ int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 			goto err_out;
 		}
 
+		ib_umem_get_ref(umem);
 		cq->cpu_addr = NULL;
 		cq->dma_addr = ib_umem_start_dma_addr(umem);
 		cq->umem = umem;
@@ -1300,7 +1301,9 @@ int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 err_destroy_cq:
 	efa_destroy_cq_idx(dev, cq->cq_idx);
 err_free_mapped:
-	if (!umem)
+	if (umem)
+		ib_umem_release(umem);
+	else
 		efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
 				DMA_FROM_DEVICE);
 err_out:
Signature · 1 line
2.51.1
author
#24 lore raw

PATCH 3/10 RDMA/mlx5: Add support for CQ creation with external umem buffer

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Add mlx5_ib_create_cq_umem() function that allows creating a Completion Queue with an externally provided user memory region instead of allocating one internally.

Change-Id Ie686dc8aa76fb72299a2bff2de30e95efbcffb3e
---
 drivers/infiniband/hw/mlx5/cq.c      | 42 ++++++++++++++++++++++------
 drivers/infiniband/hw/mlx5/main.c    |  1 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  2 ++
 3 files changed, 36 insertions(+), 9 deletions(-)
drivers/infiniband/hw/mlx5/cq.c +33 -9
index 651d76bca114..2558a52c31c0 100644
@@ -718,6 +718,7 @@ static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
 			  int *cqe_size, int *index, int *inlen,
+			  struct ib_umem *ext_umem,
 			  struct uverbs_attr_bundle *attrs)
 {
 	struct mlx5_ib_create_cq ucmd = {};
@@ -749,12 +750,21 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 
 	*cqe_size = ucmd.cqe_size;
 
-	cq->buf.umem =
-		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
-			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
-	if (IS_ERR(cq->buf.umem)) {
-		err = PTR_ERR(cq->buf.umem);
-		return err;
+	if (ext_umem) {
+		if (ext_umem->length < entries * ucmd.cqe_size) {
+			mlx5_ib_dbg(dev, "External umem too small for CQ\n");
+			return -EINVAL;
+		}
+		ib_umem_get_ref(ext_umem);
+		cq->buf.umem = ext_umem;
+	} else {
+		cq->buf.umem =
+			ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
+				    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
+		if (IS_ERR(cq->buf.umem)) {
+			err = PTR_ERR(cq->buf.umem);
+			return err;
+		}
 	}
 
 	page_size = mlx5_umem_find_best_cq_quantized_pgoff(
@@ -949,8 +959,10 @@ static void notify_soft_wc_handler(struct work_struct *work)
 	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
 }
 
-int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
-		      struct uverbs_attr_bundle *attrs)
+static int __mlx5_ib_create_cq(struct ib_cq *ibcq,
+			       const struct ib_cq_init_attr *attr,
+			       struct ib_umem *ext_umem,
+			       struct uverbs_attr_bundle *attrs)
 {
 	struct ib_udata *udata = &attrs->driver_udata;
 	struct ib_device *ibdev = ibcq->device;
@@ -989,7 +1001,7 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 
 	if (udata) {
 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
-				     &index, &inlen, attrs);
+				     &index, &inlen, ext_umem, attrs);
 		if (err)
 			return err;
 	} else {
@@ -1058,6 +1070,18 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	return err;
 }
 
+int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+		      struct uverbs_attr_bundle *attrs)
+{
+	return __mlx5_ib_create_cq(ibcq, attr, NULL, attrs);
+}
+
+int mlx5_ib_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+			   struct ib_umem *umem, struct uverbs_attr_bundle *attrs)
+{
+	return __mlx5_ib_create_cq(ibcq, attr, umem, attrs);
+}
+
 int mlx5_ib_pre_destroy_cq(struct ib_cq *cq)
 {
 	struct mlx5_ib_dev *dev = to_mdev(cq->device);
drivers/infiniband/hw/mlx5/main.c +1 -0
index eba023b7af0f..a719738ebd0c 100644
@@ -4447,6 +4447,7 @@ static const struct ib_device_ops mlx5_ib_dev_ops = {
 	.check_mr_status = mlx5_ib_check_mr_status,
 	.create_ah = mlx5_ib_create_ah,
 	.create_cq = mlx5_ib_create_cq,
+	.create_cq_umem = mlx5_ib_create_cq_umem,
 	.create_qp = mlx5_ib_create_qp,
 	.create_srq = mlx5_ib_create_srq,
 	.create_user_ah = mlx5_ib_create_ah,
drivers/infiniband/hw/mlx5/mlx5_ib.h +2 -0
index 4f4114d95130..25efdae83d27 100644
@@ -1371,6 +1371,8 @@ int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
 			 size_t buflen, size_t *bc);
 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		      struct uverbs_attr_bundle *attrs);
+int mlx5_ib_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+			   struct ib_umem *umem, struct uverbs_attr_bundle *attrs);
 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata);
 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
 int mlx5_ib_pre_destroy_cq(struct ib_cq *cq);
Signature · 1 line
2.51.1
author
#25 lore raw

PATCH 4/10 RDMA/uverbs: Factor out common buffer umem parsing into helper

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Introduce uverbs_get_buffer_umem() helper function to consolidate the common pattern of parsing buffer attributes (VA or dmabuf FD) and creating an ib_umem.

Change-Id If76883934c4020fbfc0cdc9378934244dc406a49
---
 drivers/infiniband/core/uverbs_ioctl.c        | 68 ++++++++++++++++++
 drivers/infiniband/core/uverbs_std_types_cq.c | 69 ++++---------------
 include/rdma/uverbs_ioctl.h                   |  7 ++
 3 files changed, 87 insertions(+), 57 deletions(-)
drivers/infiniband/core/uverbs_ioctl.c +68 -0
index f80da6a67e24..581b0705b496 100644
@@ -32,6 +32,7 @@
 
 #include <rdma/rdma_user_ioctl.h>
 #include <rdma/uverbs_ioctl.h>
+#include <rdma/ib_umem.h>
 #include "rdma_core.h"
 #include "uverbs.h"
 
@@ -847,3 +848,70 @@ void uverbs_finalize_uobj_create(const struct uverbs_attr_bundle *bundle,
 		  pbundle->uobj_hw_obj_valid);
 }
 EXPORT_SYMBOL(uverbs_finalize_uobj_create);
+
+int uverbs_get_buffer_umem(struct ib_device *ib_dev,
+			   struct uverbs_attr_bundle *attrs,
+			   u16 va_attr, u16 len_attr,
+			   u16 fd_attr, u16 offset_attr,
+			   bool has_umem_support,
+			   int access, struct ib_umem **umem)
+{
+	struct ib_umem_dmabuf *umem_dmabuf;
+	u64 buffer_va, buffer_length, buffer_offset;
+	int buffer_fd;
+	int ret;
+
+	*umem = NULL;
+
+	if (uverbs_attr_is_valid(attrs, va_attr)) {
+		ret = uverbs_copy_from(&buffer_va, attrs, va_attr);
+		if (ret)
+			return ret;
+
+		ret = uverbs_copy_from(&buffer_length, attrs, len_attr);
+		if (ret)
+			return ret;
+
+		if (uverbs_attr_is_valid(attrs, fd_attr) ||
+		    uverbs_attr_is_valid(attrs, offset_attr) ||
+		    !has_umem_support)
+			return -EINVAL;
+
+		*umem = ib_umem_get(ib_dev, buffer_va, buffer_length, access);
+		if (IS_ERR(*umem)) {
+			ret = PTR_ERR(*umem);
+			*umem = NULL;
+			return ret;
+		}
+	} else if (uverbs_attr_is_valid(attrs, fd_attr)) {
+		ret = uverbs_get_raw_fd(&buffer_fd, attrs, fd_attr);
+		if (ret)
+			return ret;
+
+		ret = uverbs_copy_from(&buffer_offset, attrs, offset_attr);
+		if (ret)
+			return ret;
+
+		ret = uverbs_copy_from(&buffer_length, attrs, len_attr);
+		if (ret)
+			return ret;
+
+		if (uverbs_attr_is_valid(attrs, va_attr) ||
+		    !has_umem_support)
+			return -EINVAL;
+
+		umem_dmabuf = ib_umem_dmabuf_get_pinned(ib_dev, buffer_offset,
+						       buffer_length, buffer_fd,
+						       access);
+		if (IS_ERR(umem_dmabuf))
+			return PTR_ERR(umem_dmabuf);
+
+		*umem = &umem_dmabuf->umem;
+	} else if (uverbs_attr_is_valid(attrs, len_attr) ||
+		   uverbs_attr_is_valid(attrs, offset_attr)) {
+		/* Length or offset without VA/FD is invalid */
+		return -EINVAL;
+	}
+
+	return 0;
+}
drivers/infiniband/core/uverbs_std_types_cq.c +12 -57

86 lines not on this page · raw message

include/rdma/uverbs_ioctl.h +7 -0
index e6c0de227fad..2f96edf395bd 100644
@@ -861,6 +861,13 @@ int uverbs_get_flags32(u32 *to, const struct uverbs_attr_bundle *attrs_bundle,
 		       size_t idx, u64 allowed_bits);
 int uverbs_copy_to(const struct uverbs_attr_bundle *attrs_bundle, size_t idx,
 		   const void *from, size_t size);
+struct ib_umem;
+int uverbs_get_buffer_umem(struct ib_device *ib_dev,
+			   struct uverbs_attr_bundle *attrs,
+			   u16 va_attr, u16 len_attr,
+			   u16 fd_attr, u16 offset_attr,
+			   bool has_umem_support,
+			   int access, struct ib_umem **umem);
 __malloc void *_uverbs_alloc(struct uverbs_attr_bundle *bundle, size_t size,
 			     gfp_t flags);
 
Signature · 1 line
2.51.1
author
#26 lore raw

PATCH 5/10 RDMA/core: Add support for QP buffer umem in QP creation

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Extend the QP creation path to support passing SQ and RQ buffers via user memory (umem), either through virtual address or dmabuf file descriptor.

Add new uverbs attributes for QP creation:
- SQ/RQ buffer virtual address
- SQ/RQ buffer length
- SQ/RQ buffer dmabuf file descriptor
- SQ/RQ buffer offset (for dmabuf)

The VA and FD modes are mutually exclusive for each buffer.

Introduce ib_create_qp_user_umem() function that extends ib_create_qp_user() with sq_umem and rq_umem parameters. The existing ib_create_qp_user() becomes a wrapper that passes NULL for umem pointers, maintaining backward compatibility with all existing callers.

Add create_qp_umem device op for drivers to implement umem-based QP creation. When umem buffers are provided, the core calls create_qp_umem instead of create_qp, allowing drivers to use pre-pinned memory for QP buffers.

Change-Id I08b1542a15fe6b89755ab7199b461050a45a9160
---
 drivers/infiniband/core/core_priv.h           |  7 ++
 drivers/infiniband/core/device.c              |  1 +
 drivers/infiniband/core/uverbs_std_types_qp.c | 63 +++++++++++++++++-
 drivers/infiniband/core/verbs.c               | 65 +++++++++++++++----
 include/rdma/ib_verbs.h                       |  4 ++
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  8 +++
 6 files changed, 132 insertions(+), 16 deletions(-)
drivers/infiniband/core/core_priv.h +7 -0
index 05102769a918..553326c4eca9 100644
@@ -320,6 +320,13 @@ struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
 				struct ib_qp_init_attr *attr,
 				struct ib_udata *udata,
 				struct ib_uqp_object *uobj, const char *caller);
+struct ib_qp *ib_create_qp_user_umem(struct ib_device *dev, struct ib_pd *pd,
+				     struct ib_qp_init_attr *attr,
+				     struct ib_umem *sq_umem,
+				     struct ib_umem *rq_umem,
+				     struct ib_udata *udata,
+				     struct ib_uqp_object *uobj,
+				     const char *caller);
 
 void ib_qp_usecnt_inc(struct ib_qp *qp);
 void ib_qp_usecnt_dec(struct ib_qp *qp);
drivers/infiniband/core/device.c +1 -0
index 4e09f6e0995e..a9ae03f1e936 100644
@@ -2704,6 +2704,7 @@ void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
 	SET_DEVICE_OP(dev_ops, create_cq_umem);
 	SET_DEVICE_OP(dev_ops, create_flow);
 	SET_DEVICE_OP(dev_ops, create_qp);
+	SET_DEVICE_OP(dev_ops, create_qp_umem);
 	SET_DEVICE_OP(dev_ops, create_rwq_ind_table);
 	SET_DEVICE_OP(dev_ops, create_srq);
 	SET_DEVICE_OP(dev_ops, create_user_ah);
drivers/infiniband/core/uverbs_std_types_qp.c +60 -3

90 lines not on this page · raw message

drivers/infiniband/core/verbs.c +52 -13

114 lines not on this page · raw message

include/rdma/ib_verbs.h +4 -0
index 8bd020da7745..831426d27078 100644
@@ -2521,6 +2521,10 @@ struct ib_device_ops {
 	int (*destroy_srq)(struct ib_srq *srq, struct ib_udata *udata);
 	int (*create_qp)(struct ib_qp *qp, struct ib_qp_init_attr *qp_init_attr,
 			 struct ib_udata *udata);
+	int (*create_qp_umem)(struct ib_qp *qp,
+			      struct ib_qp_init_attr *qp_init_attr,
+			      struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			      struct ib_udata *udata);
 	int (*modify_qp)(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
 			 int qp_attr_mask, struct ib_udata *udata);
 	int (*query_qp)(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
include/uapi/rdma/ib_user_ioctl_cmds.h +8 -0
index 35da4026f452..0d6b4151512d 100644
@@ -157,6 +157,14 @@ enum uverbs_attrs_create_qp_cmd_attr_ids {
 	UVERBS_ATTR_CREATE_QP_EVENT_FD,
 	UVERBS_ATTR_CREATE_QP_RESP_CAP,
 	UVERBS_ATTR_CREATE_QP_RESP_QP_NUM,
+	UVERBS_ATTR_CREATE_QP_SQ_BUFFER_VA,
+	UVERBS_ATTR_CREATE_QP_SQ_BUFFER_LENGTH,
+	UVERBS_ATTR_CREATE_QP_SQ_BUFFER_FD,
+	UVERBS_ATTR_CREATE_QP_SQ_BUFFER_OFFSET,
+	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_VA,
+	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_LENGTH,
+	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_FD,
+	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_OFFSET,
 };
 
 enum uverbs_attrs_destroy_qp_cmd_attr_ids {
Signature · 1 line
2.51.1
author
#27 lore raw

PATCH 6/10 RDMA/mlx5: Add support for QP creation with external umem buffers

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Implement the create_qp_umem device operation to support QP creation with externally provided umem buffers.

Change-Id Id53b7a0bfea91b6653ef6590f019498afffb58ab
---
 drivers/infiniband/hw/mlx5/main.c    |  1 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  3 ++
 drivers/infiniband/hw/mlx5/qp.c      | 81 +++++++++++++++++++++++-----
 3 files changed, 71 insertions(+), 14 deletions(-)
drivers/infiniband/hw/mlx5/main.c +1 -0
index a719738ebd0c..e77a46e463c6 100644
@@ -4449,6 +4449,7 @@ static const struct ib_device_ops mlx5_ib_dev_ops = {
 	.create_cq = mlx5_ib_create_cq,
 	.create_cq_umem = mlx5_ib_create_cq_umem,
 	.create_qp = mlx5_ib_create_qp,
+	.create_qp_umem = mlx5_ib_create_qp_umem,
 	.create_srq = mlx5_ib_create_srq,
 	.create_user_ah = mlx5_ib_create_ah,
 	.dealloc_pd = mlx5_ib_dealloc_pd,
drivers/infiniband/hw/mlx5/mlx5_ib.h +3 -0
index 25efdae83d27..755d7c17dbfd 100644
@@ -1356,6 +1356,9 @@ int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp);
 void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp);
 int mlx5_ib_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
 		      struct ib_udata *udata);
+int mlx5_ib_create_qp_umem(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
+			   struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			   struct ib_udata *udata);
 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		      int attr_mask, struct ib_udata *udata);
 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
drivers/infiniband/hw/mlx5/qp.c +67 -14

159 lines not on this page · raw message

Signature · 1 line
2.51.1
author
#28 lore raw

PATCH 7/10 RDMA/uverbs: Add doorbell record umem support to CQ creation

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Extend the user CQ creation path to support passing doorbell record (DBR) memory via umem. This allows userspace to provide pre-pinned doorbell record buffers, which is useful for confidential computing scenarios where doorbell records need to be in decrypted memory accessible by the device.

The DBR umem can be provided via:
- Virtual address (VA) mode: UVERBS_ATTR_CREATE_CQ_DBR_VA + LENGTH
- DMA-buf file descriptor mode: UVERBS_ATTR_CREATE_CQ_DBR_FD + OFFSET + LENGTH

These modes are mutually exclusive, similar to the CQ buffer umem.

Change-Id Id0d92c0c98133dcf938664a37444535d4e0a9ea6
---
 drivers/infiniband/core/uverbs_std_types_cq.c | 38 +++++++++++++++----
 drivers/infiniband/hw/efa/efa.h               |  3 +-
 drivers/infiniband/hw/efa/efa_verbs.c         |  5 ++-
 drivers/infiniband/hw/mlx5/cq.c               |  3 +-
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |  3 +-
 include/rdma/ib_verbs.h                       |  1 +
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  4 ++
 7 files changed, 45 insertions(+), 12 deletions(-)
drivers/infiniband/core/uverbs_std_types_cq.c +31 -7
index 64c43dbd96ba..95ee0fea09b0 100644
@@ -68,7 +68,7 @@ static int UVERBS_HANDLER(UVERBS_METHOD_CQ_CREATE)(
 	struct ib_device *ib_dev = attrs->context->device;
 	struct ib_cq_init_attr attr = {};
 	struct ib_uobject *ev_file_uobj;
-	struct ib_umem *umem;
+	struct ib_umem *umem, *dbr_umem;
 	struct ib_cq *cq;
 	u64 user_handle;
 	int ret;
@@ -123,17 +123,25 @@ static int UVERBS_HANDLER(UVERBS_METHOD_CQ_CREATE)(
 	if (ret)
 		goto err_event_file;
 
+	ret = uverbs_get_buffer_umem(ib_dev, attrs,
+				     UVERBS_ATTR_CREATE_CQ_DBR_VA,
+				     UVERBS_ATTR_CREATE_CQ_DBR_LENGTH,
+				     UVERBS_ATTR_CREATE_CQ_DBR_FD,
+				     UVERBS_ATTR_CREATE_CQ_DBR_OFFSET,
+				     ib_dev->ops.create_cq_umem, 0, &dbr_umem);
+	if (ret)
+		goto err_umem;
+
 	/* If no external buffer provided, require create_cq op */
-	if (!umem && !ib_dev->ops.create_cq) {
+	if (!umem && !dbr_umem && !ib_dev->ops.create_cq) {
 		ret = -EINVAL;
-		goto err_event_file;
+		goto err_dbr_umem;
 	}
 
 	cq = rdma_zalloc_drv_obj(ib_dev, ib_cq);
 	if (!cq) {
 		ret = -ENOMEM;
-		ib_umem_release(umem);
-		goto err_event_file;
+		goto err_dbr_umem;
 	}
 
 	cq->device        = ib_dev;
@@ -146,12 +154,14 @@ static int UVERBS_HANDLER(UVERBS_METHOD_CQ_CREATE)(
 	rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
 	rdma_restrack_set_name(&cq->res, NULL);
 
-	ret = umem ? ib_dev->ops.create_cq_umem(cq, &attr, umem, attrs) :
+	ret = (umem || dbr_umem) ?
+		ib_dev->ops.create_cq_umem(cq, &attr, umem, dbr_umem, attrs) :
 		ib_dev->ops.create_cq(cq, &attr, attrs);
 	if (ret)
 		goto err_free;
 
 	/* Driver took a reference, release ours */
+	ib_umem_release(dbr_umem);
 	ib_umem_release(umem);
 
 	obj->uevent.uobject.object = cq;
@@ -164,9 +174,12 @@ static int UVERBS_HANDLER(UVERBS_METHOD_CQ_CREATE)(
 	return ret;
 
 err_free:
-	ib_umem_release(umem);
 	rdma_restrack_put(&cq->res);
 	kfree(cq);
+err_dbr_umem:
+	ib_umem_release(dbr_umem);
+err_umem:
+	ib_umem_release(umem);
 err_event_file:
 	if (obj->uevent.event_file)
 		uverbs_uobject_put(&obj->uevent.event_file->uobj);
@@ -214,6 +227,17 @@ DECLARE_UVERBS_NAMED_METHOD(
 	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_CREATE_CQ_BUFFER_OFFSET,
 			   UVERBS_ATTR_TYPE(u64),
 			   UA_OPTIONAL),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_CREATE_CQ_DBR_VA,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_OPTIONAL),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_CREATE_CQ_DBR_LENGTH,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_OPTIONAL),
+	UVERBS_ATTR_RAW_FD(UVERBS_ATTR_CREATE_CQ_DBR_FD,
+			   UA_OPTIONAL),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_CREATE_CQ_DBR_OFFSET,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_OPTIONAL),
 	UVERBS_ATTR_UHW());
 
 static int UVERBS_HANDLER(UVERBS_METHOD_CQ_DESTROY)(
drivers/infiniband/hw/efa/efa.h +2 -1
index 96f9c3bc98b2..e4be0335af51 100644
@@ -164,7 +164,8 @@ int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata);
 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		  struct uverbs_attr_bundle *attrs);
 int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
-		       struct ib_umem *umem, struct uverbs_attr_bundle *attrs);
+		       struct ib_umem *umem, struct ib_umem *dbr_umem,
+		       struct uverbs_attr_bundle *attrs);
 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
 			 u64 virt_addr, int access_flags,
 			 struct ib_dmah *dmah,
drivers/infiniband/hw/efa/efa_verbs.c +3 -2
index aca48825cd71..165a2e1e5b72 100644
@@ -1132,7 +1132,8 @@ static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
 }
 
 int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
-		       struct ib_umem *umem, struct uverbs_attr_bundle *attrs)
+		       struct ib_umem *umem, struct ib_umem *dbr_umem,
+		       struct uverbs_attr_bundle *attrs)
 {
 	struct ib_udata *udata = &attrs->driver_udata;
 	struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
@@ -1314,7 +1315,7 @@ int efa_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		  struct uverbs_attr_bundle *attrs)
 {
-	return efa_create_cq_umem(ibcq, attr, NULL, attrs);
+	return efa_create_cq_umem(ibcq, attr, NULL, NULL, attrs);
 }
 
 static int umem_to_page_list(struct efa_dev *dev,
drivers/infiniband/hw/mlx5/cq.c +2 -1
index 2558a52c31c0..35c51a91e8c4 100644
@@ -1077,7 +1077,8 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 }
 
 int mlx5_ib_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
-			   struct ib_umem *umem, struct uverbs_attr_bundle *attrs)
+			   struct ib_umem *umem, struct ib_umem *dbr_umem,
+			   struct uverbs_attr_bundle *attrs)
 {
 	return __mlx5_ib_create_cq(ibcq, attr, umem, attrs);
 }
drivers/infiniband/hw/mlx5/mlx5_ib.h +2 -1
index 755d7c17dbfd..05f764185e92 100644
@@ -1375,7 +1375,8 @@ int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		      struct uverbs_attr_bundle *attrs);
 int mlx5_ib_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
-			   struct ib_umem *umem, struct uverbs_attr_bundle *attrs);
+			   struct ib_umem *umem, struct ib_umem *dbr_umem,
+			   struct uverbs_attr_bundle *attrs);
 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata);
 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
 int mlx5_ib_pre_destroy_cq(struct ib_cq *cq);
include/rdma/ib_verbs.h +1 -0
index 831426d27078..eeafa5358b49 100644
@@ -2535,6 +2535,7 @@ struct ib_device_ops {
 	int (*create_cq_umem)(struct ib_cq *cq,
 			      const struct ib_cq_init_attr *attr,
 			      struct ib_umem *umem,
+			      struct ib_umem *dbr_umem,
 			      struct uverbs_attr_bundle *attrs);
 	int (*modify_cq)(struct ib_cq *cq, u16 cq_count, u16 cq_period);
 	int (*destroy_cq)(struct ib_cq *cq, struct ib_udata *udata);
include/uapi/rdma/ib_user_ioctl_cmds.h +4 -0
index 0d6b4151512d..ef33b96511a8 100644
@@ -116,6 +116,10 @@ enum uverbs_attrs_create_cq_cmd_attr_ids {
 	UVERBS_ATTR_CREATE_CQ_BUFFER_LENGTH,
 	UVERBS_ATTR_CREATE_CQ_BUFFER_FD,
 	UVERBS_ATTR_CREATE_CQ_BUFFER_OFFSET,
+	UVERBS_ATTR_CREATE_CQ_DBR_VA,
+	UVERBS_ATTR_CREATE_CQ_DBR_LENGTH,
+	UVERBS_ATTR_CREATE_CQ_DBR_FD,
+	UVERBS_ATTR_CREATE_CQ_DBR_OFFSET,
 };
 
 enum uverbs_attrs_destroy_cq_cmd_attr_ids {
Signature · 1 line
2.51.1
author
#29 lore raw

PATCH 8/10 RDMA/mlx5: Add external doorbell record umem support for CQ

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Extend the mlx5 doorbell mapping infrastructure to support externally provided doorbell record (DBR) umem buffers. This enables userspace to pass pre-pinned DBR memory for CQ creation.

Change-Id Ic2abbd048da9eb0d691423a7df97630a6a7f5e67
---
 drivers/infiniband/hw/mlx5/cq.c       | 12 +++++++-----
 drivers/infiniband/hw/mlx5/doorbell.c | 27 +++++++++++++++++++++++----
 drivers/infiniband/hw/mlx5/mlx5_ib.h  |  2 +-
 drivers/infiniband/hw/mlx5/qp.c       |  4 ++--
 drivers/infiniband/hw/mlx5/srq.c      |  2 +-
 5 files changed, 34 insertions(+), 13 deletions(-)
drivers/infiniband/hw/mlx5/cq.c +7 -5
index 35c51a91e8c4..3afcb81e53d4 100644
@@ -718,7 +718,7 @@ static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format)
 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 			  struct mlx5_ib_cq *cq, int entries, u32 **cqb,
 			  int *cqe_size, int *index, int *inlen,
-			  struct ib_umem *ext_umem,
+			  struct ib_umem *ext_umem, struct ib_umem *ext_dbr_umem,
 			  struct uverbs_attr_bundle *attrs)
 {
 	struct mlx5_ib_create_cq ucmd = {};
@@ -775,7 +775,7 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 		goto err_umem;
 	}
 
-	err = mlx5_ib_db_map_user(context, ucmd.db_addr, &cq->db);
+	err = mlx5_ib_db_map_user(context, ucmd.db_addr, ext_dbr_umem, &cq->db);
 	if (err)
 		goto err_umem;
 
@@ -962,6 +962,7 @@ static void notify_soft_wc_handler(struct work_struct *work)
 static int __mlx5_ib_create_cq(struct ib_cq *ibcq,
 			       const struct ib_cq_init_attr *attr,
 			       struct ib_umem *ext_umem,
+			       struct ib_umem *ext_dbr_umem,
 			       struct uverbs_attr_bundle *attrs)
 {
 	struct ib_udata *udata = &attrs->driver_udata;
@@ -1001,7 +1002,8 @@ static int __mlx5_ib_create_cq(struct ib_cq *ibcq,
 
 	if (udata) {
 		err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size,
-				     &index, &inlen, ext_umem, attrs);
+				     &index, &inlen, ext_umem, ext_dbr_umem,
+				     attrs);
 		if (err)
 			return err;
 	} else {
@@ -1073,14 +1075,14 @@ static int __mlx5_ib_create_cq(struct ib_cq *ibcq,
 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		      struct uverbs_attr_bundle *attrs)
 {
-	return __mlx5_ib_create_cq(ibcq, attr, NULL, attrs);
+	return __mlx5_ib_create_cq(ibcq, attr, NULL, NULL, attrs);
 }
 
 int mlx5_ib_create_cq_umem(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 			   struct ib_umem *umem, struct ib_umem *dbr_umem,
 			   struct uverbs_attr_bundle *attrs)
 {
-	return __mlx5_ib_create_cq(ibcq, attr, umem, attrs);
+	return __mlx5_ib_create_cq(ibcq, attr, umem, dbr_umem, attrs);
 }
 
 int mlx5_ib_pre_destroy_cq(struct ib_cq *cq)
drivers/infiniband/hw/mlx5/doorbell.c +23 -4
index e32111117a5e..fd80b3ff6145 100644
@@ -46,13 +46,31 @@ struct mlx5_ib_user_db_page {
 };
 
 int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context, unsigned long virt,
-			struct mlx5_db *db)
+			struct ib_umem *ext_umem, struct mlx5_db *db)
 {
+	unsigned long dma_offset;
 	struct mlx5_ib_user_db_page *page;
 	int err = 0;
 
 	mutex_lock(&context->db_page_mutex);
 
+	if (ext_umem) {
+		/* External umem path - no page sharing */
+		page = kzalloc(sizeof(*page), GFP_KERNEL);
+		if (!page) {
+			err = -ENOMEM;
+			goto out;
+		}
+
+		ib_umem_get_ref(ext_umem);
+		page->umem = ext_umem;
+		dma_offset = ib_umem_offset(ext_umem);
+		goto add_page;
+	}
+
+	dma_offset = virt & ~PAGE_MASK;
+
+	/* Standard path - lookup existing or create new page */
 	list_for_each_entry(page, &context->db_page_list, list)
 		if ((current->mm == page->mm) &&
 		    (page->user_virt == (virt & PAGE_MASK)))
@@ -76,11 +94,11 @@ int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context, unsigned long virt,
 	mmgrab(current->mm);
 	page->mm = current->mm;
 
+add_page:
 	list_add(&page->list, &context->db_page_list);
 
 found:
-	db->dma = sg_dma_address(page->umem->sgt_append.sgt.sgl) +
-		  (virt & ~PAGE_MASK);
+	db->dma = sg_dma_address(page->umem->sgt_append.sgt.sgl) + dma_offset;
 	db->u.user_page = page;
 	++page->refcnt;
 
@@ -96,7 +114,8 @@ void mlx5_ib_db_unmap_user(struct mlx5_ib_ucontext *context, struct mlx5_db *db)
 
 	if (!--db->u.user_page->refcnt) {
 		list_del(&db->u.user_page->list);
-		mmdrop(db->u.user_page->mm);
+		if (db->u.user_page->mm)
+			mmdrop(db->u.user_page->mm);
 		ib_umem_release(db->u.user_page->umem);
 		kfree(db->u.user_page);
 	}
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -1
index 05f764185e92..c1e60f5c1754 100644
@@ -1332,7 +1332,7 @@ to_mmmap(struct rdma_user_mmap_entry *rdma_entry)
 int mlx5_ib_dev_res_cq_init(struct mlx5_ib_dev *dev);
 int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev);
 int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context, unsigned long virt,
-			struct mlx5_db *db);
+			struct ib_umem *ext_umem, struct mlx5_db *db);
 void mlx5_ib_db_unmap_user(struct mlx5_ib_ucontext *context, struct mlx5_db *db);
 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
drivers/infiniband/hw/mlx5/qp.c +2 -2
index 010f707e209d..4b926b6f2461 100644
@@ -918,7 +918,7 @@ static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 		ib_umem_num_pages(rwq->umem), page_size, rwq->rq_num_pas,
 		offset);
 
-	err = mlx5_ib_db_map_user(ucontext, ucmd->db_addr, &rwq->db);
+	err = mlx5_ib_db_map_user(ucontext, ucmd->db_addr, NULL, &rwq->db);
 	if (err) {
 		mlx5_ib_dbg(dev, "map failed\n");
 		goto err_umem;
@@ -1064,7 +1064,7 @@ static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
 	qp->bfregn = bfregn;
 
-	err = mlx5_ib_db_map_user(context, ucmd->db_addr, &qp->db);
+	err = mlx5_ib_db_map_user(context, ucmd->db_addr, NULL, &qp->db);
 	if (err) {
 		mlx5_ib_dbg(dev, "map failed\n");
 		goto err_free;
drivers/infiniband/hw/mlx5/srq.c +1 -1
index bcb6b324af50..6169ffaf457c 100644
@@ -83,7 +83,7 @@ static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
 	}
 	in->umem = srq->umem;
 
-	err = mlx5_ib_db_map_user(ucontext, ucmd.db_addr, &srq->db);
+	err = mlx5_ib_db_map_user(ucontext, ucmd.db_addr, NULL, &srq->db);
 	if (err) {
 		mlx5_ib_dbg(dev, "map doorbell failed\n");
 		goto err_umem;
Signature · 1 line
2.51.1
author
#30 lore raw

PATCH 9/10 RDMA/uverbs: Add doorbell record umem support to QP creation

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Extend the QP creation path to support passing doorbell record (DBR) memory via umem for both SQ and RQ. This allows userspace to provide pre-pinned doorbell record buffers.

The DBR umem can be provided via:
- Virtual address (VA) mode: UVERBS_ATTR_CREATE_QP_{SQ,RQ}_DBR_VA + LENGTH
- DMA-buf file descriptor mode: UVERBS_ATTR_CREATE_QP_{SQ,RQ}_DBR_FD + OFFSET + LENGTH

These modes are mutually exclusive, similar to the QP buffer umem.

Change-Id Id2836b852d114a93c9b5b45c09bcef386386a193
---
 drivers/infiniband/core/core_priv.h           |  2 +
 drivers/infiniband/core/uverbs_std_types_qp.c | 55 ++++++++++++++++++-
 drivers/infiniband/core/verbs.c               | 25 ++++++---
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |  1 +
 drivers/infiniband/hw/mlx5/qp.c               |  1 +
 include/rdma/ib_verbs.h                       |  2 +
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  8 +++
 7 files changed, 83 insertions(+), 11 deletions(-)
drivers/infiniband/core/core_priv.h +2 -0
index 553326c4eca9..baca8cc120aa 100644
@@ -324,6 +324,8 @@ struct ib_qp *ib_create_qp_user_umem(struct ib_device *dev, struct ib_pd *pd,
 				     struct ib_qp_init_attr *attr,
 				     struct ib_umem *sq_umem,
 				     struct ib_umem *rq_umem,
+				     struct ib_umem *sq_dbr_umem,
+				     struct ib_umem *rq_dbr_umem,
 				     struct ib_udata *udata,
 				     struct ib_uqp_object *uobj,
 				     const char *caller);
drivers/infiniband/core/uverbs_std_types_qp.c +53 -2

86 lines not on this page · raw message

drivers/infiniband/core/verbs.c +16 -9
index 97073b50fc30..db0ad750a78a 100644
@@ -1265,6 +1265,8 @@ static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
 			       struct ib_qp_init_attr *attr,
 			       struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			       struct ib_umem *sq_dbr_umem,
+			       struct ib_umem *rq_dbr_umem,
 			       struct ib_udata *udata,
 			       struct ib_uqp_object *uobj, const char *caller)
 {
@@ -1272,7 +1274,7 @@ static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
 	struct ib_qp *qp;
 	int ret;
 
-	if (sq_umem || rq_umem) {
+	if (sq_umem || rq_umem || sq_dbr_umem || rq_dbr_umem) {
 		if (!dev->ops.create_qp_umem)
 			return ERR_PTR(-EOPNOTSUPP);
 	} else {
@@ -1309,9 +1311,9 @@ static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
 	WARN_ONCE(!udata && !caller, "Missing kernel QP owner");
 	rdma_restrack_set_name(&qp->res, udata ? NULL : caller);
 
-	if (sq_umem || rq_umem)
+	if (sq_umem || rq_umem || sq_dbr_umem || rq_dbr_umem)
 		ret = dev->ops.create_qp_umem(qp, attr, sq_umem, rq_umem,
-					      udata);
+					      sq_dbr_umem, rq_dbr_umem, udata);
 	else
 		ret = dev->ops.create_qp(qp, attr, udata);
 	if (ret)
@@ -1349,6 +1351,8 @@ static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
  *   the actual capabilities of the created QP.
  * @sq_umem: SQ buffer umem (optional)
  * @rq_umem: RQ buffer umem (optional)
+ * @sq_dbr_umem: SQ doorbell record umem (optional)
+ * @rq_dbr_umem: RQ doorbell record umem (optional)
  * @udata: User data
  * @uobj: uverbs object
  * @caller: caller's build-time module name
@@ -1357,6 +1361,8 @@ struct ib_qp *ib_create_qp_user_umem(struct ib_device *dev, struct ib_pd *pd,
 				     struct ib_qp_init_attr *attr,
 				     struct ib_umem *sq_umem,
 				     struct ib_umem *rq_umem,
+				     struct ib_umem *sq_dbr_umem,
+				     struct ib_umem *rq_dbr_umem,
 				     struct ib_udata *udata,
 				     struct ib_uqp_object *uobj,
 				     const char *caller)
@@ -1364,10 +1370,11 @@ struct ib_qp *ib_create_qp_user_umem(struct ib_device *dev, struct ib_pd *pd,
 	struct ib_qp *qp, *xrc_qp;
 
 	if (attr->qp_type == IB_QPT_XRC_TGT)
-		qp = create_qp(dev, pd, attr, sq_umem, rq_umem, NULL, NULL, caller);
+		qp = create_qp(dev, pd, attr, NULL, NULL, NULL, NULL, NULL,
+			       NULL, caller);
 	else
-		qp = create_qp(dev, pd, attr, sq_umem, rq_umem, udata, uobj,
-			       NULL);
+		qp = create_qp(dev, pd, attr, sq_umem, rq_umem, sq_dbr_umem,
+			       rq_dbr_umem, udata, uobj, NULL);
 	if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
 		return qp;
 
@@ -1399,8 +1406,8 @@ struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
 				struct ib_udata *udata,
 				struct ib_uqp_object *uobj, const char *caller)
 {
-	return ib_create_qp_user_umem(dev, pd, attr, NULL, NULL, udata, uobj,
-				      caller);
+	return ib_create_qp_user_umem(dev, pd, attr, NULL, NULL, NULL, NULL,
+				      udata, uobj, caller);
 }
 EXPORT_SYMBOL(ib_create_qp_user);
 
@@ -1452,7 +1459,7 @@ struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
 		rdma_rw_init_qp(device, qp_init_attr);
 
 	qp = create_qp(device, pd, qp_init_attr, NULL, NULL, NULL, NULL,
-		       caller);
+		       NULL, NULL, caller);
 	if (IS_ERR(qp))
 		return qp;
 
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -0
index c1e60f5c1754..f654d0fde3f1 100644
@@ -1358,6 +1358,7 @@ int mlx5_ib_create_qp(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
 		      struct ib_udata *udata);
 int mlx5_ib_create_qp_umem(struct ib_qp *qp, struct ib_qp_init_attr *init_attr,
 			   struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			   struct ib_umem *sq_dbr_umem, struct ib_umem *rq_dbr_umem,
 			   struct ib_udata *udata);
 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		      int attr_mask, struct ib_udata *udata);
drivers/infiniband/hw/mlx5/qp.c +1 -0
index 4b926b6f2461..79d30e68b4cb 100644
@@ -3399,6 +3399,7 @@ int mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 
 int mlx5_ib_create_qp_umem(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 			   struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			   struct ib_umem *sq_dbr_umem, struct ib_umem *rq_dbr_umem,
 			   struct ib_udata *udata)
 {
 	return __mlx5_ib_create_qp(ibqp, attr, udata, sq_umem, rq_umem);
include/rdma/ib_verbs.h +2 -0
index eeafa5358b49..36d59e9d45b5 100644
@@ -2524,6 +2524,8 @@ struct ib_device_ops {
 	int (*create_qp_umem)(struct ib_qp *qp,
 			      struct ib_qp_init_attr *qp_init_attr,
 			      struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			      struct ib_umem *sq_dbr_umem,
+			      struct ib_umem *rq_dbr_umem,
 			      struct ib_udata *udata);
 	int (*modify_qp)(struct ib_qp *qp, struct ib_qp_attr *qp_attr,
 			 int qp_attr_mask, struct ib_udata *udata);
include/uapi/rdma/ib_user_ioctl_cmds.h +8 -0
index ef33b96511a8..5e9be458e990 100644
@@ -169,6 +169,14 @@ enum uverbs_attrs_create_qp_cmd_attr_ids {
 	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_LENGTH,
 	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_FD,
 	UVERBS_ATTR_CREATE_QP_RQ_BUFFER_OFFSET,
+	UVERBS_ATTR_CREATE_QP_SQ_DBR_VA,
+	UVERBS_ATTR_CREATE_QP_SQ_DBR_LENGTH,
+	UVERBS_ATTR_CREATE_QP_SQ_DBR_FD,
+	UVERBS_ATTR_CREATE_QP_SQ_DBR_OFFSET,
+	UVERBS_ATTR_CREATE_QP_RQ_DBR_VA,
+	UVERBS_ATTR_CREATE_QP_RQ_DBR_LENGTH,
+	UVERBS_ATTR_CREATE_QP_RQ_DBR_FD,
+	UVERBS_ATTR_CREATE_QP_RQ_DBR_OFFSET,
 };
 
 enum uverbs_attrs_destroy_qp_cmd_attr_ids {
Signature · 1 line
2.51.1
author
#31 lore raw

PATCH 10/10 RDMA/mlx5: Add external doorbell record umem support for QP

In reply to Jiri Pirko

From: Jiri Pirko <[email protected]>

Extend the mlx5 QP creation path to support externally provided doorbell record (DBR) umem buffers. This enables userspace to pass pre-pinned DBR memory for QP creation.

Note that a single doorbell is shared between SQ and RQ for raw-QPs.

Change-Id I641f6412ad99b48f6633da742bf169df2bd1edf3
---
 drivers/infiniband/hw/mlx5/qp.c | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +14 -7
index 79d30e68b4cb..f8c24f66fef5 100644
@@ -944,7 +944,8 @@ static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 			   struct mlx5_ib_create_qp_resp *resp, int *inlen,
 			   struct mlx5_ib_qp_base *base,
 			   struct mlx5_ib_create_qp *ucmd,
-			   struct ib_umem *ext_umem)
+			   struct ib_umem *ext_umem,
+			   struct ib_umem *ext_dbr_umem)
 {
 	struct mlx5_ib_ucontext *context;
 	struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
@@ -1064,7 +1065,7 @@ static int _create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 		resp->bfreg_index = MLX5_IB_INVALID_BFREG;
 	qp->bfregn = bfregn;
 
-	err = mlx5_ib_db_map_user(context, ucmd->db_addr, NULL, &qp->db);
+	err = mlx5_ib_db_map_user(context, ucmd->db_addr, ext_dbr_umem, &qp->db);
 	if (err) {
 		mlx5_ib_dbg(dev, "map failed\n");
 		goto err_free;
@@ -1744,6 +1745,7 @@ struct mlx5_create_qp_params {
 	struct mlx5_ib_create_qp_resp resp;
 	struct ib_umem *sq_ext_umem;
 	struct ib_umem *rq_ext_umem;
+	struct ib_umem *rq_dbr_ext_umem;
 };
 
 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct ib_pd *pd,
@@ -2158,7 +2160,7 @@ static int create_dci(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 		return ts_format;
 
 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
-			      &inlen, base, ucmd, NULL);
+			      &inlen, base, ucmd, NULL, NULL);
 	if (err)
 		return err;
 
@@ -2326,7 +2328,8 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 	}
 
 	err = _create_user_qp(dev, pd, qp, udata, init_attr, &in, &params->resp,
-			      &inlen, base, ucmd, params->rq_ext_umem);
+			      &inlen, base, ucmd, params->rq_ext_umem,
+			      params->rq_dbr_ext_umem);
 	if (err)
 		return err;
 
@@ -3290,7 +3293,8 @@ static int check_ucmd_data(struct mlx5_ib_dev *dev,
 
 static int __mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 			       struct ib_udata *udata,
-			       struct ib_umem *sq_umem, struct ib_umem *rq_umem)
+			       struct ib_umem *sq_umem, struct ib_umem *rq_umem,
+			       struct ib_umem *rq_dbr_umem)
 {
 	struct mlx5_create_qp_params params = {};
 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
@@ -3317,6 +3321,7 @@ static int __mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 	params.is_rss_raw = !!attr->rwq_ind_tbl;
 	params.sq_ext_umem = sq_umem;
 	params.rq_ext_umem = rq_umem;
+	params.rq_dbr_ext_umem = rq_dbr_umem;
 
 	if (udata) {
 		err = process_udata_size(dev, &params);
@@ -3394,7 +3399,7 @@ static int __mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 int mlx5_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 		      struct ib_udata *udata)
 {
-	return __mlx5_ib_create_qp(ibqp, attr, udata, NULL, NULL);
+	return __mlx5_ib_create_qp(ibqp, attr, udata, NULL, NULL, NULL);
 }
 
 int mlx5_ib_create_qp_umem(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
@@ -3402,7 +3407,9 @@ int mlx5_ib_create_qp_umem(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
 			   struct ib_umem *sq_dbr_umem, struct ib_umem *rq_dbr_umem,
 			   struct ib_udata *udata)
 {
-	return __mlx5_ib_create_qp(ibqp, attr, udata, sq_umem, rq_umem);
+	/* Single DBR umem for both SQ and RQ. */
+	return __mlx5_ib_create_qp(ibqp, attr, udata, sq_umem, rq_umem,
+				   rq_dbr_umem);
 }
 
 int mlx5_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
Signature · 1 line
2.51.1
#32 lore raw
In reply to Jiri Pirko
Quoted 8 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

This patchset extends the existing CQ umem buffer infrastructure to QP creation and adds doorbell record (DBR) umem support for both CQ and QP.

The kernel already supports CQ creation with externally provided umem buffers. This patchset:

Quoted 8 lines

1. Adds umem reference counting to simplify memory lifecycle management.
2. Adds mlx5 driver implementation for the existing CQ buffer umem support.
3. Extends QP creation to support external SQ/RQ buffer umems.
4. Adds doorbell record umem support for both CQ and QP.
5. Implements all extensions in mlx5 driver.

This enables integration with dmabuf exporters for specialized memory types, such as GPU memory for GPU-direct RDMA scenarios.

What do you mean by that?

Are you referring to the RDMA dma‑buf exporter from Yishai? The GPU dma‑buf exporters already work correctly with the existing RDMA importer.

Thanks

author
#33 lore raw
Quoted 15 lines · Tue, Feb 03, 2026 at 10:59:33AM +0100, [email protected] wrote:
Quoted 14 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

This patchset extends the existing CQ umem buffer infrastructure to QP creation and adds doorbell record (DBR) umem support for both CQ and QP.

The kernel already supports CQ creation with externally provided umem buffers. This patchset:

1. Adds umem reference counting to simplify memory lifecycle management.
2. Adds mlx5 driver implementation for the existing CQ buffer umem support.
3. Extends QP creation to support external SQ/RQ buffer umems.
4. Adds doorbell record umem support for both CQ and QP.
5. Implements all extensions in mlx5 driver.

Quoted 8 lines
Quoted 2 lines

This enables integration with dmabuf exporters for specialized memory types, such as GPU memory for GPU-direct RDMA scenarios.

What do you mean by that?

Are you referring to the RDMA dma‑buf exporter from Yishai? The GPU dma‑buf exporters already work correctly with the existing RDMA importer.

Basically it allows you to use any DMA-buf exported memory for CQ/QP/DBR purposes. Currently, only MR works with DMA-buf exported memory.