[PATCH v7 0/5] RDMA: Add dma-buf support

[PATCH v7 0/5] RDMA: Add dma-buf support

[Jianxin Xiong]

This is the seventh version of the patch set. Changelog:

v7:
* Rebase on top of latest mlx5 MR patch series
* Slice dma-buf sg list at runtime instead of creating a new list
* Preload the buffer page mapping when the MR is created
* Move the 'dma_virt_ops' check into dma_buf_dynamic_attach()

v6: https://www.spinics.net/lists/linux-rdma/msg96923.html
* Move the dma-buf invalidation callback from the core to the device
  driver
* Move mapping update from work queue to pagefault handler
* Add dma-buf based MRs to the xarray of mmkeys so that the pagefault
  handler can be reached
* Update the new driver method and uverbs command signature by changing
  the paramter 'addr' to 'offset'
* Modify the sg list returned from dma_buf_map_attachment() based on
  the parameters 'offset' and 'length'
* Don't import dma-buf if 'dma_virt_ops' is used by the dma device
* The patch that clarifies dma-buf sg lists alignment has landed at
  https://cgit.freedesktop.org/drm/drm-misc/commit/?id=ac80cd17a615
  and thus is no longer included with this set

v5: https://www.spinics.net/lists/linux-rdma/msg96786.html
* Fix a few warnings reported by kernel test robot:
    - no previous prototype for function 'ib_umem_dmabuf_release' 
    - no previous prototype for function 'ib_umem_dmabuf_map_pages'
    - comparison of distinct pointer types in 'check_add_overflow'
* Add comment for the wait between getting the dma-buf sg tagle and
  updating the NIC page table

v4: https://www.spinics.net/lists/linux-rdma/msg96767.html
* Add a new ib_device method reg_user_mr_dmabuf() instead of expanding
  the existing method reg_user_mr()
* Use a separate code flow for dma-buf instead of adding special cases
  to the ODP memory region code path
* In invalidation callback, new mapping is updated as whole using work
  queue instead of being updated in page granularity in the page fault
  handler
* Use dma_resv_get_excl() and dma_fence_wait() to ensure the content of
  the pages have been moved to the new location before the new mapping
  is programmed into the NIC
* Add code to the ODP page fault handler to check the mapping status
* The new access flag added in v3 is removed.
* The checking for on-demand paging support in the new uverbs command
  is removed because it is implied by implementing the new ib_device
  method
* Clarify that dma-buf sg lists are page aligned

v3: https://www.spinics.net/lists/linux-rdma/msg96330.html
* Use dma_buf_dynamic_attach() instead of dma_buf_attach()
* Use on-demand paging mechanism to avoid pinning the GPU memory
* Instead of adding a new parameter to the device method for memory
  registration, pass all the attributes including the file descriptor
  as a structure
* Define a new access flag for dma-buf based memory region
* Check for on-demand paging support in the new uverbs command

v2: https://www.spinics.net/lists/linux-rdma/msg93643.html
* The Kconfig option is removed. There is no dependence issue since
  dma-buf driver is always enabled.
* The declaration of new data structure and functions is reorganized to
  minimize the visibility of the changes.
* The new uverbs command now goes through ioctl() instead of write().
* The rereg functionality is removed.
* Instead of adding new device method for dma-buf specific registration,
  existing method is extended to accept an extra parameter. 
* The correct function is now used for address range checking. 

v1: https://www.spinics.net/lists/linux-rdma/msg90720.html
* The initial patch set
* Implement core functions for importing and mapping dma-buf
* Use dma-buf static attach interface
* Add two ib_device methods reg_user_mr_fd() and rereg_user_mr_fd()
* Add two uverbs commands via the write() interface
* Add Kconfig option
* Add dma-buf support to mlx5 device

When enabled, an RDMA capable NIC can perform peer-to-peer transactions
over PCIe to access the local memory located on another device. This can
often lead to better performance than using a system memory buffer for
RDMA and copying data between the buffer and device memory.

Current kernel RDMA stack uses get_user_pages() to pin the physical
pages backing the user buffer and uses dma_map_sg_attrs() to get the
dma addresses for memory access. This usually doesn't work for peer
device memory due to the lack of associated page structures.

Several mechanisms exist today to facilitate device memory access.

ZONE_DEVICE is a new zone for device memory in the memory management
subsystem. It allows pages from device memory being described with
specialized page structures, but what can be done with these page
structures may be different from system memory. ZONE_DEVICE is further
specialized into multiple memory types, such as one type for PCI
p2pmem/p2pdma and one type for HMM.

PCI p2pmem/p2pdma uses ZONE_DEVICE to represent device memory residing
in a PCI BAR and provides a set of calls to publish, discover, allocate,
and map such memory for peer-to-peer transactions. One feature of the
API is that the buffer is allocated by the side that does the DMA
transfer. This works well with the storage usage case, but is awkward
with GPU-NIC communication, where typically the buffer is allocated by
the GPU driver rather than the NIC driver.

Heterogeneous Memory Management (HMM) utilizes mmu_interval_notifier
and ZONE_DEVICE to support shared virtual address space and page
migration between system memory and device memory. HMM doesn't support
pinning device memory because pages located on device must be able to
migrate to system memory when accessed by CPU. Peer-to-peer access
is currently not supported by HMM.

Dma-buf is a standard mechanism for sharing buffers among different
device drivers. The buffer to be shared is exported by the owning
driver and imported by the driver that wants to use it. The exporter
provides a set of ops that the importer can call to pin and map the
buffer. In addition, a file descriptor can be associated with a dma-
buf object as the handle that can be passed to user space.

This patch series adds dma-buf importer role to the RDMA driver in
attempt to support RDMA using device memory such as GPU VRAM. Dma-buf is
chosen for a few reasons: first, the API is relatively simple and allows
a lot of flexibility in implementing the buffer manipulation ops.
Second, it doesn't require page structure. Third, dma-buf is already
supported in many GPU drivers. However, we are aware that existing GPU
drivers don't allow pinning device memory via the dma-buf interface.
Pinning would simply cause the backing storage to migrate to system RAM.
True peer-to-peer access is only possible using dynamic attach, which
requires on-demand paging support from the NIC to work. For this reason,
this series only works with ODP capable NICs.

This series consists of five patches. The first patch adds the common
code for importing dma-buf from a file descriptor and mapping the
dma-buf pages. Patch 2 add the new driver method reg_user_mr_dmabuf().
Patch 3 adds a new uverbs command for registering dma-buf based memory
region. Patch 4 adds dma-buf support to the mlx5 driver. Patch 5 adds
code to dma_buf_dynamic_attach() to check the importer device for use
of dma_virt_ops and reject the request if so.

Related user space RDMA library changes will be provided as a separate
patch series.

Jianxin Xiong (5):
  RDMA/umem: Support importing dma-buf as user memory region
  RDMA/core: Add device method for registering dma-buf based memory
    region
  RDMA/uverbs: Add uverbs command for dma-buf based MR registration
  RDMA/mlx5: Support dma-buf based userspace memory region
  dma-buf: Reject attach request from importers that use dma_virt_ops

 drivers/dma-buf/dma-buf.c                     |   5 +
 drivers/infiniband/core/Makefile              |   2 +-
 drivers/infiniband/core/device.c              |   1 +
 drivers/infiniband/core/umem.c                |   7 ++
 drivers/infiniband/core/umem_dmabuf.c         | 131 ++++++++++++++++++++++++++
 drivers/infiniband/core/umem_dmabuf.h         |  11 +++
 drivers/infiniband/core/uverbs_std_types_mr.c | 112 ++++++++++++++++++++++
 drivers/infiniband/hw/mlx5/main.c             |   2 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |  18 ++++
 drivers/infiniband/hw/mlx5/mr.c               | 123 +++++++++++++++++++++++-
 drivers/infiniband/hw/mlx5/odp.c              | 105 +++++++++++++++++++--
 include/rdma/ib_umem.h                        |  35 ++++++-
 include/rdma/ib_verbs.h                       |   6 +-
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  14 +++
 14 files changed, 557 insertions(+), 15 deletions(-)
 create mode 100644 drivers/infiniband/core/umem_dmabuf.c
 create mode 100644 drivers/infiniband/core/umem_dmabuf.h

-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v7 1/5] RDMA/umem: Support importing dma-buf as user memory region

[Jianxin Xiong]

Dma-buf is a standard cross-driver buffer sharing mechanism that can be
used to support peer-to-peer access from RDMA devices.

Device memory exported via dma-buf is associated with a file descriptor.
This is passed to the user space as a property associated with the
buffer allocation. When the buffer is registered as a memory region,
the file descriptor is passed to the RDMA driver along with other
parameters.

Implement the common code for importing dma-buf object and mapping
dma-buf pages.

Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
---
 drivers/infiniband/core/Makefile      |   2 +-
 drivers/infiniband/core/umem.c        |   7 ++
 drivers/infiniband/core/umem_dmabuf.c | 131 ++++++++++++++++++++++++++++++++++
 drivers/infiniband/core/umem_dmabuf.h |  11 +++
 include/rdma/ib_umem.h                |  35 ++++++++-
 5 files changed, 184 insertions(+), 2 deletions(-)
 create mode 100644 drivers/infiniband/core/umem_dmabuf.c
 create mode 100644 drivers/infiniband/core/umem_dmabuf.h

diff --git a/drivers/infiniband/core/Makefile b/drivers/infiniband/core/Makefile
index ccf2670..8ab4eea 100644
--- a/drivers/infiniband/core/Makefile
+++ b/drivers/infiniband/core/Makefile
@@ -40,5 +40,5 @@ ib_uverbs-y :=			uverbs_main.o uverbs_cmd.o uverbs_marshall.o \
 				uverbs_std_types_srq.o \
 				uverbs_std_types_wq.o \
 				uverbs_std_types_qp.o
-ib_uverbs-$(CONFIG_INFINIBAND_USER_MEM) += umem.o
+ib_uverbs-$(CONFIG_INFINIBAND_USER_MEM) += umem.o umem_dmabuf.o
 ib_uverbs-$(CONFIG_INFINIBAND_ON_DEMAND_PAGING) += umem_odp.o
diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c
index f1fc7e3..1a8a5b5 100644
--- a/drivers/infiniband/core/umem.c
+++ b/drivers/infiniband/core/umem.c
@@ -2,6 +2,7 @@
  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2020 Intel Corporation. All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -43,6 +44,7 @@
 #include <rdma/ib_umem_odp.h>
 
 #include "uverbs.h"
+#include "umem_dmabuf.h"
 
 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
 {
@@ -93,6 +95,9 @@ unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 		return page_size;
 	}
 
+	if (umem->is_dmabuf)
+		return PAGE_SIZE;
+
 	/* rdma_for_each_block() has a bug if the page size is smaller than the
 	 * page size used to build the umem. For now prevent smaller page sizes
 	 * from being returned.
@@ -278,6 +283,8 @@ void ib_umem_release(struct ib_umem *umem)
 {
 	if (!umem)
 		return;
+	if (umem->is_dmabuf)
+		return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem));
 	if (umem->is_odp)
 		return ib_umem_odp_release(to_ib_umem_odp(umem));
 
diff --git a/drivers/infiniband/core/umem_dmabuf.c b/drivers/infiniband/core/umem_dmabuf.c
new file mode 100644
index 0000000..816e73d
--- /dev/null
+++ b/drivers/infiniband/core/umem_dmabuf.c
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
+/*
+ * Copyright (c) 2020 Intel Corporation. All rights reserved.
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/dma-resv.h>
+#include <linux/dma-mapping.h>
+
+#include "uverbs.h"
+#include "umem_dmabuf.h"
+
+int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf)
+{
+	struct sg_table *sgt;
+	struct dma_fence *fence;
+
+	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
+
+	sgt = dma_buf_map_attachment(umem_dmabuf->attach,
+				     DMA_BIDIRECTIONAL);
+
+	if (IS_ERR(sgt))
+		return PTR_ERR(sgt);
+
+	umem_dmabuf->umem.sg_head = *sgt;
+	umem_dmabuf->umem.nmap = sgt->nents;
+	umem_dmabuf->sgt = sgt;
+
+	/*
+	 * Although the sg list is valid now, the content of the pages
+	 * may be not up-to-date. Wait for the exporter to finish
+	 * the migration.
+	 */
+	fence = dma_resv_get_excl(umem_dmabuf->attach->dmabuf->resv);
+	if (fence)
+		dma_fence_wait(fence, false);
+
+	return 0;
+}
+EXPORT_SYMBOL(ib_umem_dmabuf_map_pages);
+
+void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf)
+{
+	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
+
+	if (!umem_dmabuf->sgt)
+		return;
+
+	dma_buf_unmap_attachment(umem_dmabuf->attach, umem_dmabuf->sgt,
+				 DMA_BIDIRECTIONAL);
+	umem_dmabuf->sgt = NULL;
+}
+EXPORT_SYMBOL(ib_umem_dmabuf_unmap_pages);
+
+struct ib_umem *ib_umem_dmabuf_get(struct ib_device *device,
+				   unsigned long offset, size_t size,
+				   int fd, int access,
+				   const struct dma_buf_attach_ops *ops)
+{
+	struct dma_buf *dmabuf;
+	struct ib_umem_dmabuf *umem_dmabuf;
+	struct ib_umem *umem;
+	unsigned long end;
+	long ret;
+
+	if (check_add_overflow(offset, (unsigned long)size, &end))
+		return ERR_PTR(-EINVAL);
+
+	if (unlikely(PAGE_ALIGN(end) < PAGE_SIZE))
+		return ERR_PTR(-EINVAL);
+
+	if (unlikely(!ops || !ops->move_notify))
+		return ERR_PTR(-EINVAL);
+
+	umem_dmabuf = kzalloc(sizeof(*umem_dmabuf), GFP_KERNEL);
+	if (!umem_dmabuf)
+		return ERR_PTR(-ENOMEM);
+
+	umem = &umem_dmabuf->umem;
+	umem->ibdev = device;
+	umem->length = size;
+	umem->address = offset;
+	umem->iova = offset;
+	umem->writable = ib_access_writable(access);
+	umem->is_dmabuf = 1;
+
+	if (unlikely(!ib_umem_num_pages(umem))) {
+		ret = -EINVAL;
+		goto out_free_umem;
+	}
+
+	dmabuf = dma_buf_get(fd);
+	if (IS_ERR(dmabuf)) {
+		ret = PTR_ERR(dmabuf);
+		goto out_free_umem;
+	}
+
+	umem_dmabuf->attach = dma_buf_dynamic_attach(
+					dmabuf,
+					device->dma_device,
+					ops,
+					umem_dmabuf);
+	if (IS_ERR(umem_dmabuf->attach)) {
+		ret = PTR_ERR(umem_dmabuf->attach);
+		goto out_release_dmabuf;
+	}
+
+	return umem;
+
+out_release_dmabuf:
+	dma_buf_put(dmabuf);
+
+out_free_umem:
+	kfree(umem_dmabuf);
+	return ERR_PTR(ret);
+}
+EXPORT_SYMBOL(ib_umem_dmabuf_get);
+
+void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf)
+{
+	struct dma_buf *dmabuf = umem_dmabuf->attach->dmabuf;
+
+	dma_resv_lock(dmabuf->resv, NULL);
+	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
+	dma_resv_unlock(dmabuf->resv);
+
+	dma_buf_detach(dmabuf, umem_dmabuf->attach);
+	dma_buf_put(dmabuf);
+	kfree(umem_dmabuf);
+}
diff --git a/drivers/infiniband/core/umem_dmabuf.h b/drivers/infiniband/core/umem_dmabuf.h
new file mode 100644
index 0000000..13acf55
--- /dev/null
+++ b/drivers/infiniband/core/umem_dmabuf.h
@@ -0,0 +1,11 @@
+/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */
+/*
+ * Copyright (c) 2020 Intel Corporation. All rights reserved.
+ */
+
+#ifndef UMEM_DMABUF_H
+#define UMEM_DMABUF_H
+
+void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf);
+
+#endif /* UMEM_DMABUF_H */
diff --git a/include/rdma/ib_umem.h b/include/rdma/ib_umem.h
index 7059750..cf46b15 100644
--- a/include/rdma/ib_umem.h
+++ b/include/rdma/ib_umem.h
@@ -1,6 +1,7 @@
 /* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
 /*
  * Copyright (c) 2007 Cisco Systems.  All rights reserved.
+ * Copyright (c) 2020 Intel Corporation.  All rights reserved.
  */
 
 #ifndef IB_UMEM_H
@@ -13,6 +14,7 @@
 
 struct ib_ucontext;
 struct ib_umem_odp;
+struct dma_buf_attach_ops;
 
 struct ib_umem {
 	struct ib_device       *ibdev;
@@ -22,12 +24,25 @@ struct ib_umem {
 	unsigned long		address;
 	u32 writable : 1;
 	u32 is_odp : 1;
+	u32 is_dmabuf : 1;
 	struct work_struct	work;
 	struct sg_table sg_head;
 	int             nmap;
 	unsigned int    sg_nents;
 };
 
+struct ib_umem_dmabuf {
+	struct ib_umem umem;
+	struct dma_buf_attachment *attach;
+	struct sg_table *sgt;
+	void *private;
+};
+
+static inline struct ib_umem_dmabuf *to_ib_umem_dmabuf(struct ib_umem *umem)
+{
+	return container_of(umem, struct ib_umem_dmabuf, umem);
+}
+
 /* Returns the offset of the umem start relative to the first page. */
 static inline int ib_umem_offset(struct ib_umem *umem)
 {
@@ -79,6 +94,12 @@ int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 				     unsigned long pgsz_bitmap,
 				     unsigned long virt);
+struct ib_umem *ib_umem_dmabuf_get(struct ib_device *device,
+				   unsigned long offset, size_t size,
+				   int fd, int access,
+				   const struct dma_buf_attach_ops *ops);
+int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf);
+void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf);
 
 #else /* CONFIG_INFINIBAND_USER_MEM */
 
@@ -101,7 +122,19 @@ static inline unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 {
 	return 0;
 }
+static inline struct ib_umem *ib_umem_dmabuf_get(struct ib_device *device,
+						 unsigned long offset,
+						 size_t size, int fd,
+						 int access,
+						 struct dma_buf_attach_ops *ops)
+{
+	return ERR_PTR(-EINVAL);
+}
+static inline int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf)
+{
+	return -EINVAL;
+}
+static inline void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf) { }
 
 #endif /* CONFIG_INFINIBAND_USER_MEM */
-
 #endif /* IB_UMEM_H */
-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v7 2/5] RDMA/core: Add device method for registering dma-buf based memory region

[Jianxin Xiong]

Dma-buf based memory region requires one extra parameter and is processed
quite differently. Adding a separate method allows clean separation from
regular memory regions.

Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
---
 drivers/infiniband/core/device.c | 1 +
 include/rdma/ib_verbs.h          | 6 +++++-
 2 files changed, 6 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
index ce26564..6768a19 100644
--- a/drivers/infiniband/core/device.c
+++ b/drivers/infiniband/core/device.c
@@ -2689,6 +2689,7 @@ void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
 	SET_DEVICE_OP(dev_ops, read_counters);
 	SET_DEVICE_OP(dev_ops, reg_dm_mr);
 	SET_DEVICE_OP(dev_ops, reg_user_mr);
+	SET_DEVICE_OP(dev_ops, reg_user_mr_dmabuf);
 	SET_DEVICE_OP(dev_ops, req_ncomp_notif);
 	SET_DEVICE_OP(dev_ops, req_notify_cq);
 	SET_DEVICE_OP(dev_ops, rereg_user_mr);
diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index 9420827..3d1d098 100644
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h
@@ -2,7 +2,7 @@
 /*
  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
- * Copyright (c) 2004 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2004, 2020 Intel Corporation.  All rights reserved.
  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
@@ -2433,6 +2433,10 @@ struct ib_device_ops {
 	struct ib_mr *(*reg_user_mr)(struct ib_pd *pd, u64 start, u64 length,
 				     u64 virt_addr, int mr_access_flags,
 				     struct ib_udata *udata);
+	struct ib_mr *(*reg_user_mr_dmabuf)(struct ib_pd *pd, u64 offset,
+				     u64 length, u64 virt_addr, int fd,
+				     int mr_access_flags,
+				     struct ib_udata *udata);
 	int (*rereg_user_mr)(struct ib_mr *mr, int flags, u64 start, u64 length,
 			     u64 virt_addr, int mr_access_flags,
 			     struct ib_pd *pd, struct ib_udata *udata);
-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v7 3/5] RDMA/uverbs: Add uverbs command for dma-buf based MR registration

[Jianxin Xiong]

Implement a new uverbs ioctl method for memory registration with file
descriptor as an extra parameter.

Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
---
 drivers/infiniband/core/uverbs_std_types_mr.c | 112 ++++++++++++++++++++++++++
 include/uapi/rdma/ib_user_ioctl_cmds.h        |  14 ++++
 2 files changed, 126 insertions(+)

diff --git a/drivers/infiniband/core/uverbs_std_types_mr.c b/drivers/infiniband/core/uverbs_std_types_mr.c
index 9b22bb5..f22e58b 100644
--- a/drivers/infiniband/core/uverbs_std_types_mr.c
+++ b/drivers/infiniband/core/uverbs_std_types_mr.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2018, Mellanox Technologies inc.  All rights reserved.
+ * Copyright (c) 2020, Intel Corporation.  All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -178,6 +179,85 @@ static int UVERBS_HANDLER(UVERBS_METHOD_QUERY_MR)(
 	return IS_UVERBS_COPY_ERR(ret) ? ret : 0;
 }
 
+static int UVERBS_HANDLER(UVERBS_METHOD_REG_DMABUF_MR)(
+	struct uverbs_attr_bundle *attrs)
+{
+	struct ib_uobject *uobj =
+		uverbs_attr_get_uobject(attrs, UVERBS_ATTR_REG_DMABUF_MR_HANDLE);
+	struct ib_pd *pd =
+		uverbs_attr_get_obj(attrs, UVERBS_ATTR_REG_DMABUF_MR_PD_HANDLE);
+	struct ib_device *ib_dev = pd->device;
+
+	u64 offset, length, virt_addr;
+	u32 fd, access_flags;
+	struct ib_mr *mr;
+	int ret;
+
+	if (!ib_dev->ops.reg_user_mr_dmabuf)
+		return -EOPNOTSUPP;
+
+	ret = uverbs_copy_from(&offset, attrs,
+			       UVERBS_ATTR_REG_DMABUF_MR_OFFSET);
+	if (ret)
+		return ret;
+
+	ret = uverbs_copy_from(&length, attrs,
+			       UVERBS_ATTR_REG_DMABUF_MR_LENGTH);
+	if (ret)
+		return ret;
+
+	ret = uverbs_copy_from(&virt_addr, attrs,
+			       UVERBS_ATTR_REG_DMABUF_MR_IOVA);
+	if (ret)
+		return ret;
+
+	ret = uverbs_copy_from(&fd, attrs,
+			       UVERBS_ATTR_REG_DMABUF_MR_FD);
+	if (ret)
+		return ret;
+
+	ret = uverbs_get_flags32(&access_flags, attrs,
+				 UVERBS_ATTR_REG_DMABUF_MR_ACCESS_FLAGS,
+				 IB_ACCESS_SUPPORTED);
+	if (ret)
+		return ret;
+
+	ret = ib_check_mr_access(access_flags);
+	if (ret)
+		return ret;
+
+	mr = pd->device->ops.reg_user_mr_dmabuf(pd, offset, length, virt_addr,
+						fd, access_flags,
+						&attrs->driver_udata);
+	if (IS_ERR(mr))
+		return PTR_ERR(mr);
+
+	mr->device  = pd->device;
+	mr->pd      = pd;
+	mr->type    = IB_MR_TYPE_USER;
+	mr->uobject = uobj;
+	atomic_inc(&pd->usecnt);
+
+	uobj->object = mr;
+
+	ret = uverbs_copy_to(attrs, UVERBS_ATTR_REG_DMABUF_MR_RESP_LKEY,
+			     &mr->lkey, sizeof(mr->lkey));
+	if (ret)
+		goto err_dereg;
+
+	ret = uverbs_copy_to(attrs, UVERBS_ATTR_REG_DMABUF_MR_RESP_RKEY,
+			     &mr->rkey, sizeof(mr->rkey));
+	if (ret)
+		goto err_dereg;
+
+	return 0;
+
+err_dereg:
+	ib_dereg_mr_user(mr, uverbs_get_cleared_udata(attrs));
+
+	return ret;
+}
+
 DECLARE_UVERBS_NAMED_METHOD(
 	UVERBS_METHOD_ADVISE_MR,
 	UVERBS_ATTR_IDR(UVERBS_ATTR_ADVISE_MR_PD_HANDLE,
@@ -243,6 +323,37 @@ static int UVERBS_HANDLER(UVERBS_METHOD_QUERY_MR)(
 			    UVERBS_ATTR_TYPE(u32),
 			    UA_MANDATORY));
 
+DECLARE_UVERBS_NAMED_METHOD(
+	UVERBS_METHOD_REG_DMABUF_MR,
+	UVERBS_ATTR_IDR(UVERBS_ATTR_REG_DMABUF_MR_HANDLE,
+			UVERBS_OBJECT_MR,
+			UVERBS_ACCESS_NEW,
+			UA_MANDATORY),
+	UVERBS_ATTR_IDR(UVERBS_ATTR_REG_DMABUF_MR_PD_HANDLE,
+			UVERBS_OBJECT_PD,
+			UVERBS_ACCESS_READ,
+			UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_REG_DMABUF_MR_OFFSET,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_REG_DMABUF_MR_LENGTH,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_REG_DMABUF_MR_IOVA,
+			   UVERBS_ATTR_TYPE(u64),
+			   UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(UVERBS_ATTR_REG_DMABUF_MR_FD,
+			   UVERBS_ATTR_TYPE(u32),
+			   UA_MANDATORY),
+	UVERBS_ATTR_FLAGS_IN(UVERBS_ATTR_REG_DMABUF_MR_ACCESS_FLAGS,
+			     enum ib_access_flags),
+	UVERBS_ATTR_PTR_OUT(UVERBS_ATTR_REG_DMABUF_MR_RESP_LKEY,
+			    UVERBS_ATTR_TYPE(u32),
+			    UA_MANDATORY),
+	UVERBS_ATTR_PTR_OUT(UVERBS_ATTR_REG_DMABUF_MR_RESP_RKEY,
+			    UVERBS_ATTR_TYPE(u32),
+			    UA_MANDATORY));
+
 DECLARE_UVERBS_NAMED_METHOD_DESTROY(
 	UVERBS_METHOD_MR_DESTROY,
 	UVERBS_ATTR_IDR(UVERBS_ATTR_DESTROY_MR_HANDLE,
@@ -253,6 +364,7 @@ static int UVERBS_HANDLER(UVERBS_METHOD_QUERY_MR)(
 DECLARE_UVERBS_NAMED_OBJECT(
 	UVERBS_OBJECT_MR,
 	UVERBS_TYPE_ALLOC_IDR(uverbs_free_mr),
+	&UVERBS_METHOD(UVERBS_METHOD_REG_DMABUF_MR),
 	&UVERBS_METHOD(UVERBS_METHOD_DM_MR_REG),
 	&UVERBS_METHOD(UVERBS_METHOD_MR_DESTROY),
 	&UVERBS_METHOD(UVERBS_METHOD_ADVISE_MR),
diff --git a/include/uapi/rdma/ib_user_ioctl_cmds.h b/include/uapi/rdma/ib_user_ioctl_cmds.h
index 7968a18..dafc7eb 100644
--- a/include/uapi/rdma/ib_user_ioctl_cmds.h
+++ b/include/uapi/rdma/ib_user_ioctl_cmds.h
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2018, Mellanox Technologies inc.  All rights reserved.
+ * Copyright (c) 2020, Intel Corporation. All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -251,6 +252,7 @@ enum uverbs_methods_mr {
 	UVERBS_METHOD_MR_DESTROY,
 	UVERBS_METHOD_ADVISE_MR,
 	UVERBS_METHOD_QUERY_MR,
+	UVERBS_METHOD_REG_DMABUF_MR,
 };
 
 enum uverbs_attrs_mr_destroy_ids {
@@ -272,6 +274,18 @@ enum uverbs_attrs_query_mr_cmd_attr_ids {
 	UVERBS_ATTR_QUERY_MR_RESP_IOVA,
 };
 
+enum uverbs_attrs_reg_dmabuf_mr_cmd_attr_ids {
+	UVERBS_ATTR_REG_DMABUF_MR_HANDLE,
+	UVERBS_ATTR_REG_DMABUF_MR_PD_HANDLE,
+	UVERBS_ATTR_REG_DMABUF_MR_OFFSET,
+	UVERBS_ATTR_REG_DMABUF_MR_LENGTH,
+	UVERBS_ATTR_REG_DMABUF_MR_IOVA,
+	UVERBS_ATTR_REG_DMABUF_MR_FD,
+	UVERBS_ATTR_REG_DMABUF_MR_ACCESS_FLAGS,
+	UVERBS_ATTR_REG_DMABUF_MR_RESP_LKEY,
+	UVERBS_ATTR_REG_DMABUF_MR_RESP_RKEY,
+};
+
 enum uverbs_attrs_create_counters_cmd_attr_ids {
 	UVERBS_ATTR_CREATE_COUNTERS_HANDLE,
 };
-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region

[Jianxin Xiong]

Implement the new driver method 'reg_user_mr_dmabuf'.  Utilize the core
functions to import dma-buf based memory region and update the mappings.

Add code to handle dma-buf related page fault.

Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
---
 drivers/infiniband/hw/mlx5/main.c    |   2 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  18 +++++
 drivers/infiniband/hw/mlx5/mr.c      | 123 ++++++++++++++++++++++++++++++++++-
 drivers/infiniband/hw/mlx5/odp.c     | 105 +++++++++++++++++++++++++++---
 4 files changed, 236 insertions(+), 12 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 36b15a0..e647ea4 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
 /*
  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
+ * Copyright (c) 2020, Intel Corporation. All rights reserved.
  */
 
 #include <linux/debugfs.h>
@@ -4055,6 +4056,7 @@ static int mlx5_ib_enable_driver(struct ib_device *dev)
 	.query_srq = mlx5_ib_query_srq,
 	.query_ucontext = mlx5_ib_query_ucontext,
 	.reg_user_mr = mlx5_ib_reg_user_mr,
+	.reg_user_mr_dmabuf = mlx5_ib_reg_user_mr_dmabuf,
 	.req_notify_cq = mlx5_ib_arm_cq,
 	.rereg_user_mr = mlx5_ib_rereg_user_mr,
 	.resize_cq = mlx5_ib_resize_cq,
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index bb44080..3ef6872 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -1,6 +1,7 @@
 /* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
 /*
  * Copyright (c) 2013-2020, Mellanox Technologies inc. All rights reserved.
+ * Copyright (c) 2020, Intel Corporation. All rights reserved.
  */
 
 #ifndef MLX5_IB_H
@@ -665,6 +666,12 @@ static inline bool is_odp_mr(struct mlx5_ib_mr *mr)
 	       mr->umem->is_odp;
 }
 
+static inline bool is_dmabuf_mr(struct mlx5_ib_mr *mr)
+{
+	return IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && mr->umem &&
+	       mr->umem->is_dmabuf;
+}
+
 struct mlx5_ib_mw {
 	struct ib_mw		ibmw;
 	struct mlx5_core_mkey	mmkey;
@@ -1200,6 +1207,10 @@ int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 				  u64 virt_addr, int access_flags,
 				  struct ib_udata *udata);
+struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start,
+					 u64 length, u64 virt_addr,
+					 int fd, int access_flags,
+					 struct ib_udata *udata);
 int mlx5_ib_advise_mr(struct ib_pd *pd,
 		      enum ib_uverbs_advise_mr_advice advice,
 		      u32 flags,
@@ -1210,11 +1221,13 @@ int mlx5_ib_advise_mr(struct ib_pd *pd,
 int mlx5_ib_dealloc_mw(struct ib_mw *mw);
 int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
 		       int page_shift, int flags);
+int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags);
 struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
 					     struct ib_udata *udata,
 					     int access_flags);
 void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *mr);
 void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr);
+void mlx5_ib_fence_dmabuf_mr(struct mlx5_ib_mr *mr);
 int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
 			  u64 length, u64 virt_addr, int access_flags,
 			  struct ib_pd *pd, struct ib_udata *udata);
@@ -1306,6 +1319,7 @@ int mlx5_ib_advise_mr_prefetch(struct ib_pd *pd,
 			       enum ib_uverbs_advise_mr_advice advice,
 			       u32 flags, struct ib_sge *sg_list, u32 num_sge);
 int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr, bool enable);
+int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr);
 #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 static inline void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
 {
@@ -1331,6 +1345,10 @@ static inline int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr, bool enable)
 {
 	return -EOPNOTSUPP;
 }
+static inline int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
+{
+	return -EOPNOTSUPP;
+}
 #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 
 extern const struct mmu_interval_notifier_ops mlx5_mn_ops;
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 9f653b4..f39127b 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
+ * Copyright (c) 2020, Intel Corporation. All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -36,6 +37,8 @@
 #include <linux/debugfs.h>
 #include <linux/export.h>
 #include <linux/delay.h>
+#include <linux/dma-buf.h>
+#include <linux/dma-resv.h>
 #include <rdma/ib_umem.h>
 #include <rdma/ib_umem_odp.h>
 #include <rdma/ib_verbs.h>
@@ -1212,8 +1215,11 @@ int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
 
 /*
  * Send the DMA list to the HW for a normal MR using UMR.
+ * DMABUF MR is handled in a similar way, with two exceptions:
+ * (1) The MR may only cover a sub-range of the scatterlist;
+ * (2) The MLX5_IB_UPD_XLT_ZAP flag can be used.
  */
-static int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
+int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 {
 	struct mlx5_ib_dev *dev = mr->dev;
 	struct device *ddev = dev->ib_dev.dev.parent;
@@ -1225,6 +1231,8 @@ static int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 	size_t final_size;
 	struct ib_sge sg;
 	int err = 0;
+	unsigned int skip;
+	unsigned int todo;
 
 	if (WARN_ON(mr->umem->is_odp))
 		return -EINVAL;
@@ -1237,6 +1245,11 @@ static int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 		return -ENOMEM;
 	orig_sg_length = sg.length;
 
+	if (mr->umem->is_dmabuf) {
+		skip = mr->umem->address >> mr->page_shift;
+		todo = ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift);
+	}
+
 	cur_mtt = mtt;
 	rdma_for_each_block (mr->umem->sg_head.sgl, &biter, mr->umem->nmap,
 			     BIT(mr->page_shift)) {
@@ -1255,6 +1268,19 @@ static int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 		cur_mtt->ptag =
 			cpu_to_be64(rdma_block_iter_dma_address(&biter) |
 				    MLX5_IB_MTT_PRESENT);
+
+		if (mr->umem->is_dmabuf) {
+			if (flags & MLX5_IB_UPD_XLT_ZAP)
+				cur_mtt->ptag = 0;
+			if (skip) {
+				skip--;
+				continue;
+			}
+			if (!todo)
+				break;
+			todo--;
+		}
+
 		cur_mtt++;
 	}
 
@@ -1581,6 +1607,97 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	return ERR_PTR(err);
 }
 
+static void mlx5_ib_dmabuf_invalidate_cb(struct dma_buf_attachment *attach)
+{
+	struct ib_umem_dmabuf *umem_dmabuf = attach->importer_priv;
+	struct mlx5_ib_mr *mr = umem_dmabuf->private;
+
+	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
+
+	/* mr could have been destroyed. see mlx5_ib_fence_dmabuf_mr(). */
+	if (!mr)
+		return;
+
+	mlx5_ib_update_mr_pas(mr, MLX5_IB_UPD_XLT_ZAP);
+	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
+}
+
+static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = {
+	.allow_peer2peer = 1,
+	.move_notify = mlx5_ib_dmabuf_invalidate_cb,
+};
+
+struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
+					 u64 length, u64 virt_addr,
+					 int fd, int access_flags,
+					 struct ib_udata *udata)
+{
+	struct mlx5_ib_dev *dev = to_mdev(pd->device);
+	struct mlx5_ib_mr *mr = NULL;
+	struct ib_umem *umem;
+	int err;
+
+	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
+		return ERR_PTR(-EOPNOTSUPP);
+
+	mlx5_ib_dbg(dev,
+		    "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x\n",
+		    offset, virt_addr, length, fd, access_flags);
+
+	if (!mlx5_ib_can_load_pas_with_umr(dev, length))
+		return ERR_PTR(-EINVAL);
+
+	umem = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd, access_flags,
+				  &mlx5_ib_dmabuf_attach_ops);
+	if (IS_ERR(umem)) {
+		mlx5_ib_dbg(dev, "umem get failed (%ld)\n", PTR_ERR(umem));
+		return ERR_PTR(PTR_ERR(umem));
+	}
+
+	mr = alloc_mr_from_cache(pd, umem, virt_addr, access_flags);
+	if (IS_ERR(mr))
+		mr = NULL;
+
+	if (!mr) {
+		mutex_lock(&dev->slow_path_mutex);
+		mr = reg_create(NULL, pd, umem, virt_addr, access_flags,
+				false);
+		mutex_unlock(&dev->slow_path_mutex);
+	}
+
+	if (IS_ERR(mr)) {
+		err = PTR_ERR(mr);
+		goto error;
+	}
+
+	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
+
+	mr->umem = umem;
+	atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages);
+	set_mr_fields(dev, mr, length, access_flags);
+
+	to_ib_umem_dmabuf(umem)->private = mr;
+	init_waitqueue_head(&mr->q_deferred_work);
+	atomic_set(&mr->num_deferred_work, 0);
+	err = xa_err(xa_store(&dev->odp_mkeys,
+			      mlx5_base_mkey(mr->mmkey.key), &mr->mmkey,
+			      GFP_KERNEL));
+	if (err) {
+		dereg_mr(dev, mr);
+		return ERR_PTR(err);
+	}
+
+	err = mlx5_ib_init_dmabuf_mr(mr);
+	if (err) {
+		dereg_mr(dev, mr);
+		return ERR_PTR(err);
+	}
+	return &mr->ibmr;
+error:
+	ib_umem_release(umem);
+	return ERR_PTR(err);
+}
+
 /**
  * mlx5_mr_cache_invalidate - Fence all DMA on the MR
  * @mr: The MR to fence
@@ -1649,7 +1766,7 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
 	if (!mr->umem)
 		return -EINVAL;
 
-	if (is_odp_mr(mr))
+	if (is_odp_mr(mr) || is_dmabuf_mr(mr))
 		return -EOPNOTSUPP;
 
 	if (flags & IB_MR_REREG_TRANS) {
@@ -1812,6 +1929,8 @@ static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 	/* Stop all DMA */
 	if (is_odp)
 		mlx5_ib_fence_odp_mr(mr);
+	else if (is_dmabuf_mr(mr))
+		mlx5_ib_fence_dmabuf_mr(mr);
 	else
 		clean_mr(dev, mr);
 
diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 5c853ec..546edc6 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -33,6 +33,8 @@
 #include <rdma/ib_umem.h>
 #include <rdma/ib_umem_odp.h>
 #include <linux/kernel.h>
+#include <linux/dma-buf.h>
+#include <linux/dma-resv.h>
 
 #include "mlx5_ib.h"
 #include "cmd.h"
@@ -664,6 +666,36 @@ void mlx5_ib_fence_odp_mr(struct mlx5_ib_mr *mr)
 	dma_fence_odp_mr(mr);
 }
 
+/**
+ * mlx5_ib_fence_dmabuf_mr - Stop all access to the dmabuf MR
+ * @mr: to fence
+ *
+ * On return no parallel threads will be touching this MR and no DMA will be
+ * active.
+ */
+void mlx5_ib_fence_dmabuf_mr(struct mlx5_ib_mr *mr)
+{
+	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
+
+	/* Prevent new page faults and prefetch requests from succeeding */
+	xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
+
+	/* Wait for all running page-fault handlers to finish. */
+	synchronize_srcu(&mr->dev->odp_srcu);
+
+	wait_event(mr->q_deferred_work, !atomic_read(&mr->num_deferred_work));
+
+	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
+	mlx5_mr_cache_invalidate(mr);
+	umem_dmabuf->private = NULL;
+	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
+
+	if (!mr->cache_ent) {
+		mlx5_core_destroy_mkey(mr->dev->mdev, &mr->mmkey);
+		WARN_ON(mr->descs);
+	}
+}
+
 #define MLX5_PF_FLAGS_DOWNGRADE BIT(1)
 #define MLX5_PF_FLAGS_SNAPSHOT BIT(2)
 #define MLX5_PF_FLAGS_ENABLE BIT(3)
@@ -801,6 +833,43 @@ static int pagefault_implicit_mr(struct mlx5_ib_mr *imr,
  * Returns:
  *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
  *           not accessible, or the MR is no longer valid.
+ *  -EAGAIN: The operation should be retried
+ *
+ *  >0: Number of pages mapped
+ */
+static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr,
+			       struct ib_umem_dmabuf *umem_dmabuf,
+			       u64 user_va, size_t bcnt, u32 *bytes_mapped,
+			       u32 flags)
+{
+	int npages;
+	int err;
+	u32 xlt_flags = 0;
+
+	if (flags & MLX5_PF_FLAGS_ENABLE)
+		xlt_flags |= MLX5_IB_UPD_XLT_ENABLE;
+
+	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
+	err = ib_umem_dmabuf_map_pages(umem_dmabuf);
+	if (!err)
+		err = mlx5_ib_update_mr_pas(mr, xlt_flags);
+	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
+
+	if (err)
+		return err;
+
+	if (bytes_mapped)
+		*bytes_mapped += bcnt;
+
+	npages = (ALIGN(user_va + bcnt, PAGE_SIZE) -
+		 ALIGN_DOWN(user_va, PAGE_SIZE)) >> PAGE_SHIFT;
+	return npages;
+}
+
+/*
+ * Returns:
+ *  -EFAULT: The io_virt->bcnt is not within the MR, it covers pages that are
+ *           not accessible, or the MR is no longer valid.
  *  -EAGAIN/-ENOMEM: The operation should be retried
  *
  *  -EINVAL/others: General internal malfunction
@@ -810,22 +879,31 @@ static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
 			u32 *bytes_mapped, u32 flags)
 {
 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
+	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
 
 	lockdep_assert_held(&mr->dev->odp_srcu);
 	if (unlikely(io_virt < mr->mmkey.iova))
 		return -EFAULT;
 
-	if (!odp->is_implicit_odp) {
+	if (is_dmabuf_mr(mr) || !odp->is_implicit_odp) {
 		u64 user_va;
+		u64 end;
 
 		if (check_add_overflow(io_virt - mr->mmkey.iova,
-				       (u64)odp->umem.address, &user_va))
+				       (u64)mr->umem->address, &user_va))
 			return -EFAULT;
-		if (unlikely(user_va >= ib_umem_end(odp) ||
-			     ib_umem_end(odp) - user_va < bcnt))
+		if (is_dmabuf_mr(mr))
+			end = mr->umem->address + mr->umem->length;
+		else
+			end = ib_umem_end(odp);
+		if (unlikely(user_va >= end || end - user_va < bcnt))
 			return -EFAULT;
-		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
-					 flags);
+		if (is_dmabuf_mr(mr))
+			return pagefault_dmabuf_mr(mr, umem_dmabuf, user_va,
+						   bcnt, bytes_mapped, flags);
+		else
+			return pagefault_real_mr(mr, odp, user_va, bcnt,
+						 bytes_mapped, flags);
 	}
 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,
 				     flags);
@@ -845,6 +923,16 @@ int mlx5_ib_init_odp_mr(struct mlx5_ib_mr *mr, bool enable)
 	return ret >= 0 ? 0 : ret;
 }
 
+int mlx5_ib_init_dmabuf_mr(struct mlx5_ib_mr *mr)
+{
+	int ret;
+
+	ret = pagefault_dmabuf_mr(mr, to_ib_umem_dmabuf(mr->umem),
+				  mr->umem->address, mr->umem->length, NULL,
+				  MLX5_PF_FLAGS_ENABLE);
+	return ret >= 0 ? 0 : ret;
+}
+
 struct pf_frame {
 	struct pf_frame *next;
 	u32 key;
@@ -1747,7 +1835,6 @@ static void destroy_prefetch_work(struct prefetch_mr_work *work)
 {
 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
 	struct mlx5_core_mkey *mmkey;
-	struct ib_umem_odp *odp;
 	struct mlx5_ib_mr *mr;
 
 	lockdep_assert_held(&dev->odp_srcu);
@@ -1761,11 +1848,9 @@ static void destroy_prefetch_work(struct prefetch_mr_work *work)
 	if (mr->ibmr.pd != pd)
 		return NULL;
 
-	odp = to_ib_umem_odp(mr->umem);
-
 	/* prefetch with write-access must be supported by the MR */
 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
-	    !odp->umem.writable)
+	    !mr->umem->writable)
 		return NULL;
 
 	return mr;
-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v7 5/5] dma-buf: Reject attach request from importers that use dma_virt_ops

[Jianxin Xiong]

dma_virt_ops is used by virtual devices that map pages / scatterlists to
virtual addresses for CPU access instead of performing DMA. This is not
the intended use of dma_buf_attach() and dma_buf_map_attachment(). CPU
access of dma-buf should use dma_buf_vmap() and related functions.

Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
---
 drivers/dma-buf/dma-buf.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
index 9a054fb5..ba2b877 100644
--- a/drivers/dma-buf/dma-buf.c
+++ b/drivers/dma-buf/dma-buf.c
@@ -686,6 +686,11 @@ struct dma_buf_attachment *
 	if (WARN_ON(importer_ops && !importer_ops->move_notify))
 		return ERR_PTR(-EINVAL);
 
+#ifdef CONFIG_DMA_VIRT_OPS
+	if (dev->dma_ops == &dma_virt_ops)
+		return ERR_PTR(-EINVAL);
+#endif
+
 	attach = kzalloc(sizeof(*attach), GFP_KERNEL);
 	if (!attach)
 		return ERR_PTR(-ENOMEM);
-- 
1.8.3.1

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region

[Jason Gunthorpe]

On Wed, Nov 04, 2020 at 02:06:34PM -0800, Jianxin Xiong wrote:
> +	umem = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd, access_flags,
> +				  &mlx5_ib_dmabuf_attach_ops);
> +	if (IS_ERR(umem)) {
> +		mlx5_ib_dbg(dev, "umem get failed (%ld)\n", PTR_ERR(umem));
> +		return ERR_PTR(PTR_ERR(umem));
> +	}
> +
> +	mr = alloc_mr_from_cache(pd, umem, virt_addr, access_flags);

It is very subtle, but this calls mlx5_umem_find_best_pgsz() which
calls ib_umem_find_best_pgsz() which goes over the SGL to determine
the page size to use.

As part of this it does validation of the IOVA vs first page offset vs
first page dma address. These little details come into play if the
IOVA and offset are not PAGE_SIZE aligned, which is very possible if
the dma buf exporter or system PAGE_SIZE is over 4k.

In other words, the dma_address of the first SGL must be the page
aligned starting point of the MR. Since the 'skip' approach is being
done when breaking the SGL into blocks the ib_umem_find_best_pgsz()
sees an invalid page size.

Slicing it has to be done in a way that gives a properly formed
SGL. 

My suggestion is to just change the SGL in place. Iterate to the
starting SGE in the SGL and assign it to the sg table, modify it to
have a offset dma_address and reduced length

Count the number of SGEs to span the remaning range and use that as
the new nmaped

Edit the last SGE to have a reduced length

Upon unmap undo the edits so the exporter doesn't see the mangled SGL.

It would be saner if the exporter could do this, but oh well.

Approximately like this:

	struct ib_umem *umem = &umem_p->umem;
	struct scatterlist *sg;
	int i;

	for_each_sg(umem_p->umem.sg_head.sgl, sg, umem_p->umem.nmap, i) {
		if (cur + sg_dma_len(sg) > ALIGN_DOWN(umem->address, PAGE_SIZE)) {
			unsigned long offset;

			umem_p->first_sg = sg;
			umem_p->first_dma_address = sg->dma_address;
			umem_p->first_dma_length = sg_dma_len(sg);
			umem_p->first_length = sg->length;
			offset = ALIGN_DOWN(umem->addressm PAGE_SIZE) - cur;
			sg->dma_address += offset;
			sg_dma_len(sg) -= offset;
			sg->length -= offset;
		}
		if (ALIGN(umem->address + umem->length, PAGE_SIZE) < cur + sg_dma_len(sg)) {
			unsigned long trim;

			umem_p->last_sg = sg;
			umem_p->last_dma_length = sg_dma_len(sg);
			umem_p->last_length = sg->length;
			trim =  cur + sg_dma_len(sg) - ALIGN(umem->address + umem->length, PAGE_SIZE);
			sg_dma_len(sg) -= trim;
			sg->length -= trim;
			return npages;
		}
                cur += sg_dma_len(sg);
	}
        /* It is essential that the length of the SGL exactly match
  	   the adjusted page aligned length of umem->length */
	return -EINVAL;

Further, this really only works if the umem->page_size is locked to 4k
because this doesn't have code to resize the MKEY, or change the
underlying page size when the SGL changes.

So, I'd say put something like the above in the core code to validate
and properly form the umem->sgl

Then modify the alloc_mr_from_cache to use only PAGE_SIZE:

 if (umem->is_dma_buf)
        page_size = ib_umem_find_best_pgsz(umem, PAGE_SIZE, iova);
 else
    	page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size, 0, iova);

Jason
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

RE: [PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region

[Xiong, Jianxin]

> -----Original Message-----
> From: Jason Gunthorpe <jgg@ziepe.ca>
> Sent: Wednesday, November 04, 2020 4:07 PM
> To: Xiong, Jianxin <jianxin.xiong@intel.com>
> Cc: linux-rdma@vger.kernel.org; dri-devel@lists.freedesktop.org; Doug Ledford <dledford@redhat.com>; Leon Romanovsky
> <leon@kernel.org>; Sumit Semwal <sumit.semwal@linaro.org>; Christian Koenig <christian.koenig@amd.com>; Vetter, Daniel
> <daniel.vetter@intel.com>
> Subject: Re: [PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region
> 
> On Wed, Nov 04, 2020 at 02:06:34PM -0800, Jianxin Xiong wrote:
> > +	umem = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd, access_flags,
> > +				  &mlx5_ib_dmabuf_attach_ops);
> > +	if (IS_ERR(umem)) {
> > +		mlx5_ib_dbg(dev, "umem get failed (%ld)\n", PTR_ERR(umem));
> > +		return ERR_PTR(PTR_ERR(umem));
> > +	}
> > +
> > +	mr = alloc_mr_from_cache(pd, umem, virt_addr, access_flags);
> 
> It is very subtle, but this calls mlx5_umem_find_best_pgsz() which calls ib_umem_find_best_pgsz() which goes over the SGL to determine
> the page size to use.
> 

When this is called here, the umem sglist is still NULL because dma_buf_map_attachment()
is not called until a page fault occurs. In patch 1/5, the function ib_umem_find_best_pgsz()
has been modified to always return PAGE_SIZE for dma-buf based MR.

> As part of this it does validation of the IOVA vs first page offset vs first page dma address. These little details come into play if the IOVA and
> offset are not PAGE_SIZE aligned, which is very possible if the dma buf exporter or system PAGE_SIZE is over 4k.
> 
> In other words, the dma_address of the first SGL must be the page aligned starting point of the MR. Since the 'skip' approach is being done
> when breaking the SGL into blocks the ib_umem_find_best_pgsz() sees an invalid page size.
> 
> Slicing it has to be done in a way that gives a properly formed SGL.
> 
> My suggestion is to just change the SGL in place. Iterate to the starting SGE in the SGL and assign it to the sg table, modify it to have a offset
> dma_address and reduced length
> 
> Count the number of SGEs to span the remaning range and use that as the new nmaped
> 
> Edit the last SGE to have a reduced length

Do you still think modifying the SGL in place needed given the above explanation? I do see
some benefits of doing so -- hiding the discrepancy of sgl and addr/length from the device drivers and avoid special handling in the code that use the sgl. 

> 
> Upon unmap undo the edits so the exporter doesn't see the mangled SGL.
> 
> It would be saner if the exporter could do this, but oh well.
> 
> Approximately like this:
> 
> 	struct ib_umem *umem = &umem_p->umem;
> 	struct scatterlist *sg;
> 	int i;
> 
> 	for_each_sg(umem_p->umem.sg_head.sgl, sg, umem_p->umem.nmap, i) {
> 		if (cur + sg_dma_len(sg) > ALIGN_DOWN(umem->address, PAGE_SIZE)) {
> 			unsigned long offset;
> 
> 			umem_p->first_sg = sg;
> 			umem_p->first_dma_address = sg->dma_address;
> 			umem_p->first_dma_length = sg_dma_len(sg);
> 			umem_p->first_length = sg->length;
> 			offset = ALIGN_DOWN(umem->addressm PAGE_SIZE) - cur;
> 			sg->dma_address += offset;
> 			sg_dma_len(sg) -= offset;
> 			sg->length -= offset;
> 		}
> 		if (ALIGN(umem->address + umem->length, PAGE_SIZE) < cur + sg_dma_len(sg)) {
> 			unsigned long trim;
> 
> 			umem_p->last_sg = sg;
> 			umem_p->last_dma_length = sg_dma_len(sg);
> 			umem_p->last_length = sg->length;
> 			trim =  cur + sg_dma_len(sg) - ALIGN(umem->address + umem->length, PAGE_SIZE);
> 			sg_dma_len(sg) -= trim;
> 			sg->length -= trim;
> 			return npages;
> 		}
>                 cur += sg_dma_len(sg);
> 	}
>         /* It is essential that the length of the SGL exactly match
>   	   the adjusted page aligned length of umem->length */
> 	return -EINVAL;
> 
> Further, this really only works if the umem->page_size is locked to 4k because this doesn't have code to resize the MKEY, or change the
> underlying page size when the SGL changes.

Yes, now it's locked to 4K. 

> 
> So, I'd say put something like the above in the core code to validate and properly form the umem->sgl
> 
> Then modify the alloc_mr_from_cache to use only PAGE_SIZE:
> 
>  if (umem->is_dma_buf)
>         page_size = ib_umem_find_best_pgsz(umem, PAGE_SIZE, iova);  else
>     	page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size, 0, iova);
> 

This should have been addressed in patch 1/5.

Thanks,
Jianxin

> Jason
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region

[Jason Gunthorpe]

On Thu, Nov 05, 2020 at 12:36:25AM +0000, Xiong, Jianxin wrote:
> > From: Jason Gunthorpe <jgg@ziepe.ca>
> > Sent: Wednesday, November 04, 2020 4:07 PM
> > To: Xiong, Jianxin <jianxin.xiong@intel.com>
> > Cc: linux-rdma@vger.kernel.org; dri-devel@lists.freedesktop.org; Doug Ledford <dledford@redhat.com>; Leon Romanovsky
> > <leon@kernel.org>; Sumit Semwal <sumit.semwal@linaro.org>; Christian Koenig <christian.koenig@amd.com>; Vetter, Daniel
> > <daniel.vetter@intel.com>
> > Subject: Re: [PATCH v7 4/5] RDMA/mlx5: Support dma-buf based userspace memory region
> > 
> > On Wed, Nov 04, 2020 at 02:06:34PM -0800, Jianxin Xiong wrote:
> > > +	umem = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd, access_flags,
> > > +				  &mlx5_ib_dmabuf_attach_ops);
> > > +	if (IS_ERR(umem)) {
> > > +		mlx5_ib_dbg(dev, "umem get failed (%ld)\n", PTR_ERR(umem));
> > > +		return ERR_PTR(PTR_ERR(umem));
> > > +	}
> > > +
> > > +	mr = alloc_mr_from_cache(pd, umem, virt_addr, access_flags);
> > 
> > It is very subtle, but this calls mlx5_umem_find_best_pgsz() which calls ib_umem_find_best_pgsz() which goes over the SGL to determine
> > the page size to use.
> > 
> 
> When this is called here, the umem sglist is still NULL because dma_buf_map_attachment()
> is not called until a page fault occurs. In patch 1/5, the function ib_umem_find_best_pgsz()
> has been modified to always return PAGE_SIZE for dma-buf based MR.

Oh.. That isn't a good idea.

ib_umem_find_best_pgsz() must be run on any SGL list to validate it
against the constraints, making it un-runable for the dmabuf case
means we can never support large page size or even validate that the
SGL is properly formed.

So I think this need to change the alloc_mr_from_cache() to early exit
for dma_buf ones

And it still need to call ib_umem_find_best_pgsz() but
just check the page size.

> > Edit the last SGE to have a reduced length
> 
> Do you still think modifying the SGL in place needed given the above
> explanation? I do see some benefits of doing so -- hiding the
> discrepancy of sgl and addr/length from the device drivers and avoid
> special handling in the code that use the sgl.

Yes, a umem SGL should always be properly formed or I will have a
meltdown trying to keep all the drivers working :\

Jason
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel