Skip to contentSkip to conversation

cover-letter: Allow MMIO regions to be exported through dmabuf

PATCH In review Wei Lin Guay 7 participants 19 messages Dec 16, 2024 – Jan 8, 2025

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover cover-letter: Allow MMIO regions to be exported through dmabuf

From: Wei Lin Guay <[email protected]>

This is another attempt to revive the patches posted by Jason Gunthorpe and Vivek Kasireddy, at
https://patchwork.kernel.org/project/linux-media/cover/[email protected]/
https://lwn.net/Articles/970751/

In addition to the initial proposal by Jason, another promising application is exposing memory from an AI accelerator (bound to VFIO) to an RDMA device. This would allow the RDMA device to directly access the accelerator's memory, thereby facilitating direct data transactions between the RDMA device and the accelerator.

Below is from the text/motivation from the orginal cover letter.

dma-buf has become a way to safely acquire a handle to non-struct page memory that can still have lifetime controlled by the exporter. Notably RDMA can now import dma-buf FDs and build them into MRs which allows for PCI P2P operations. Extend this to allow vfio-pci to export MMIO memory from PCI device BARs.

This series supports a use case for SPDK where a NVMe device will be owned by SPDK through VFIO but interacting with a RDMA device. The RDMA device may directly access the NVMe CMB or directly manipulate the NVMe device's doorbell using PCI P2P.

However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

v2:
- Name the new file dma_buf.c
- Restore orig_nents before freeing
- Fix reversed logic around priv->revoked
- Set priv->index
- Rebased on v2 "Break up ioctl dispatch functions"
v1: https://lore.kernel.org/r/[email protected]

Cc Oded Gabbay <[email protected]>
Cc Christian König <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Leon Romanovsky <[email protected]>
Cc Maor Gottlieb <[email protected]>

Jason Gunthorpe (3):
vfio: Add vfio_device_get()
dma-buf: Add dma_buf_try_get()
vfio/pci: Allow MMIO regions to be exported through dma-buf

Wei Lin Guay (1):
vfio/pci: Allow export dmabuf without move_notify from importer

 drivers/vfio/pci/Makefile          |   1 +
 drivers/vfio/pci/dma_buf.c         | 291 +++++++++++++++++++++++++++++
 drivers/vfio/pci/vfio_pci_config.c |   8 +-
 drivers/vfio/pci/vfio_pci_core.c   |  44 ++++-
 drivers/vfio/pci/vfio_pci_priv.h   |  30 +++
 drivers/vfio/vfio_main.c           |   1 +
 include/linux/dma-buf.h            |  13 ++
 include/linux/vfio.h               |   6 +
 include/linux/vfio_pci_core.h      |   1 +
 include/uapi/linux/vfio.h          |  18 ++
 10 files changed, 405 insertions(+), 8 deletions(-)
 create mode 100644 drivers/vfio/pci/dma_buf.c

--
2.43.5

#1 lore raw

PATCH 1/4 vfio: Add vfio_device_get()

From: Jason Gunthorpe <[email protected]>

Summary:
To increment a reference the caller already holds. Export vfio_device_put() to pair with it.

---
 drivers/vfio/vfio_main.c | 1 +
 include/linux/vfio.h     | 6 ++++++
 2 files changed, 7 insertions(+)
drivers/vfio/vfio_main.c +1 -0
index a5a62d9d963f..7e318e15abd5 100644
@@ -171,6 +171,7 @@ void vfio_device_put_registration(struct vfio_device *device)
 	if (refcount_dec_and_test(&device->refcount))
 		complete(&device->comp);
 }
+EXPORT_SYMBOL_GPL(vfio_device_put_registration);
 
 bool vfio_device_try_get_registration(struct vfio_device *device)
 {
include/linux/vfio.h +6 -0
index 000a6cab2d31..d7c790be4bbc 100644
@@ -279,6 +279,12 @@ static inline void vfio_put_device(struct vfio_device *device)
 int vfio_register_group_dev(struct vfio_device *device);
 int vfio_register_emulated_iommu_dev(struct vfio_device *device);
 void vfio_unregister_group_dev(struct vfio_device *device);
+void vfio_device_put_registration(struct vfio_device *device);
+
+static inline void vfio_device_get(struct vfio_device *device)
+{
+	refcount_inc(&device->refcount);
+}
 
 int vfio_assign_device_set(struct vfio_device *device, void *set_id);
 unsigned int vfio_device_set_open_count(struct vfio_device_set *dev_set);

--
2.43.5

#2 lore raw

PATCH 2/4 dma-buf: Add dma_buf_try_get()

In reply to Wei Lin Guay

From: Jason Gunthorpe <[email protected]>

Summary:
Used to increment the refcount of the dma buf's struct file, only if the refcount is not zero. Useful to allow the struct file's lifetime to control the lifetime of the dmabuf while still letting the driver to keep track of created dmabufs.

---
 include/linux/dma-buf.h | 13 +++++++++++++
 1 file changed, 13 insertions(+)
include/linux/dma-buf.h +13 -0
index 36216d28d8bd..9854578afecd 100644
@@ -614,6 +614,19 @@ int dma_buf_fd(struct dma_buf *dmabuf, int flags);
 struct dma_buf *dma_buf_get(int fd);
 void dma_buf_put(struct dma_buf *dmabuf);
 
+/**
+ * dma_buf_try_get - try to get a reference on a dmabuf
+ * @dmabuf - the dmabuf to get
+ *
+ * Returns true if a reference was successfully obtained. The caller must
+ * interlock with the dmabuf's release function in some way, such as RCU, to
+ * ensure that this is not called on freed memory.
+ */
+static inline bool dma_buf_try_get(struct dma_buf *dmabuf)
+{
+	return get_file_rcu(&dmabuf->file);
+}
+
 struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *,
 					enum dma_data_direction);
 void dma_buf_unmap_attachment(struct dma_buf_attachment *, struct sg_table *,

--
2.43.5

#3 lore raw

PATCH 3/4 vfio/pci: Allow MMIO regions to be exported through dma-buf

In reply to Wei Lin Guay

From: Jason Gunthorpe <[email protected]>

Summary:
dma-buf has become a way to safely acquire a handle to non-struct page memory that can still have lifetime controlled by the exporter. Notably RDMA can now import dma-buf FDs and build them into MRs which allows for PCI P2P operations. Extend this to allow vfio-pci to export MMIO memory from PCI device BARs.

The patch design loosely follows the pattern in commit db1a8dd916aa ("habanalabs: add support for dma-buf exporter") except this does not support pinning.

Instead, this implements what, in the past, we've called a revocable attachment using move. In normal situations the attachment is pinned, as a BAR does not change physical address. However when the VFIO device is closed, or a PCI reset is issued, access to the MMIO memory is revoked.

Revoked means that move occurs, but an attempt to immediately re-map the memory will fail. In the reset case a future move will be triggered when MMIO access returns. As both close and reset are under userspace control it is expected that userspace will suspend use of the dma-buf before doing these operations, the revoke is purely for kernel self-defense against a hostile userspace.

---
 drivers/vfio/pci/Makefile          |   1 +
 drivers/vfio/pci/dma_buf.c         | 269 +++++++++++++++++++++++++++++
 drivers/vfio/pci/vfio_pci_config.c |   8 +-
 drivers/vfio/pci/vfio_pci_core.c   |  28 ++-
 drivers/vfio/pci/vfio_pci_priv.h   |  23 +++
 include/linux/vfio_pci_core.h      |   1 +
 include/uapi/linux/vfio.h          |  18 ++
 7 files changed, 340 insertions(+), 8 deletions(-)
 create mode 100644 drivers/vfio/pci/dma_buf.c
drivers/vfio/pci/Makefile +1 -0
index cf00c0a7e55c..0cfdc9ede82f 100644
@@ -2,6 +2,7 @@
 
 vfio-pci-core-y := vfio_pci_core.o vfio_pci_intrs.o vfio_pci_rdwr.o vfio_pci_config.o
 vfio-pci-core-$(CONFIG_VFIO_PCI_ZDEV_KVM) += vfio_pci_zdev.o
+vfio-pci-core-$(CONFIG_DMA_SHARED_BUFFER) += dma_buf.o
 obj-$(CONFIG_VFIO_PCI_CORE) += vfio-pci-core.o
 
 vfio-pci-y := vfio_pci.o
drivers/vfio/pci/dma_buf.c new file +269 -0

269 lines not on this page · raw message

drivers/vfio/pci/vfio_pci_config.c +6 -2
index 97422aafaa7b..c605c5cb0078 100644
@@ -585,10 +585,12 @@ static int vfio_basic_config_write(struct vfio_pci_core_device *vdev, int pos,
 		virt_mem = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_MEMORY);
 		new_mem = !!(new_cmd & PCI_COMMAND_MEMORY);
 
-		if (!new_mem)
+		if (!new_mem) {
 			vfio_pci_zap_and_down_write_memory_lock(vdev);
-		else
+			vfio_pci_dma_buf_move(vdev, true);
+		} else {
 			down_write(&vdev->memory_lock);
+		}
 
 		/*
 		 * If the user is writing mem/io enable (new_mem/io) and we
@@ -623,6 +625,8 @@ static int vfio_basic_config_write(struct vfio_pci_core_device *vdev, int pos,
 		*virt_cmd &= cpu_to_le16(~mask);
 		*virt_cmd |= cpu_to_le16(new_cmd & mask);
 
+		if (__vfio_pci_memory_enabled(vdev))
+			vfio_pci_dma_buf_move(vdev, false);
 		up_write(&vdev->memory_lock);
 	}
 
drivers/vfio/pci/vfio_pci_core.c +22 -6
index ba0ce0075b2f..bb97b4d94eb7 100644
@@ -700,6 +700,8 @@ void vfio_pci_core_close_device(struct vfio_device *core_vdev)
 #endif
 	vfio_pci_core_disable(vdev);
 
+	vfio_pci_dma_buf_cleanup(vdev);
+
 	mutex_lock(&vdev->igate);
 	if (vdev->err_trigger) {
 		eventfd_ctx_put(vdev->err_trigger);
@@ -1244,7 +1246,10 @@ static int vfio_pci_ioctl_reset(struct vfio_pci_core_device *vdev,
 	 */
 	vfio_pci_set_power_state(vdev, PCI_D0);
 
+	vfio_pci_dma_buf_move(vdev, true);
 	ret = pci_try_reset_function(vdev->pdev);
+	if (__vfio_pci_memory_enabled(vdev))
+		vfio_pci_dma_buf_move(vdev, false);
 	up_write(&vdev->memory_lock);
 
 	return ret;
@@ -1490,11 +1495,10 @@ long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned int cmd,
 }
 EXPORT_SYMBOL_GPL(vfio_pci_core_ioctl);
 
-static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
-				       uuid_t __user *arg, size_t argsz)
+static int vfio_pci_core_feature_token(struct vfio_pci_core_device *vdev,
+				       u32 flags, uuid_t __user *arg,
+				       size_t argsz)
 {
-	struct vfio_pci_core_device *vdev =
-		container_of(device, struct vfio_pci_core_device, vdev);
 	uuid_t uuid;
 	int ret;
 
@@ -1521,6 +1525,9 @@ static int vfio_pci_core_feature_token(struct vfio_device *device, u32 flags,
 int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
 				void __user *arg, size_t argsz)
 {
+	struct vfio_pci_core_device *vdev =
+		container_of(device, struct vfio_pci_core_device, vdev);
+
 	switch (flags & VFIO_DEVICE_FEATURE_MASK) {
 	case VFIO_DEVICE_FEATURE_LOW_POWER_ENTRY:
 		return vfio_pci_core_pm_entry(device, flags, arg, argsz);
@@ -1530,7 +1537,9 @@ int vfio_pci_core_ioctl_feature(struct vfio_device *device, u32 flags,
 	case VFIO_DEVICE_FEATURE_LOW_POWER_EXIT:
 		return vfio_pci_core_pm_exit(device, flags, arg, argsz);
 	case VFIO_DEVICE_FEATURE_PCI_VF_TOKEN:
-		return vfio_pci_core_feature_token(device, flags, arg, argsz);
+		return vfio_pci_core_feature_token(vdev, flags, arg, argsz);
+	case VFIO_DEVICE_FEATURE_DMA_BUF:
+		return vfio_pci_core_feature_dma_buf(vdev, flags, arg, argsz);
 	default:
 		return -ENOTTY;
 	}
@@ -2083,6 +2092,7 @@ int vfio_pci_core_init_dev(struct vfio_device *core_vdev)
 	INIT_LIST_HEAD(&vdev->sriov_pfs_item);
 	init_rwsem(&vdev->memory_lock);
 	xa_init(&vdev->ctx);
+	INIT_LIST_HEAD(&vdev->dmabufs);
 
 	return 0;
 }
@@ -2463,11 +2473,17 @@ static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
 	 * cause the PCI config space reset without restoring the original
 	 * state (saved locally in 'vdev->pm_save').
 	 */
-	list_for_each_entry(vdev, &dev_set->device_list, vdev.dev_set_list)
+	list_for_each_entry(vdev, &dev_set->device_list, vdev.dev_set_list) {
+		vfio_pci_dma_buf_move(vdev, true);
 		vfio_pci_set_power_state(vdev, PCI_D0);
+	}
 
 	ret = pci_reset_bus(pdev);
 
+	list_for_each_entry(vdev, &dev_set->device_list, vdev.dev_set_list)
+		if (__vfio_pci_memory_enabled(vdev))
+			vfio_pci_dma_buf_move(vdev, false);
+
 	vdev = list_last_entry(&dev_set->device_list,
 			       struct vfio_pci_core_device, vdev.dev_set_list);
 
drivers/vfio/pci/vfio_pci_priv.h +23 -0
index 5e4fa69aee16..09d3c300918c 100644
@@ -101,4 +101,27 @@ static inline bool vfio_pci_is_vga(struct pci_dev *pdev)
 	return (pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA;
 }
 
+#ifdef CONFIG_DMA_SHARED_BUFFER
+int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
+				  struct vfio_device_feature_dma_buf __user *arg,
+				  size_t argsz);
+void vfio_pci_dma_buf_cleanup(struct vfio_pci_core_device *vdev);
+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool revoked);
+#else
+static int
+vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
+			      struct vfio_device_feature_dma_buf __user *arg,
+			      size_t argsz)
+{
+	return -ENOTTY;
+}
+static inline void vfio_pci_dma_buf_cleanup(struct vfio_pci_core_device *vdev)
+{
+}
+static inline void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev,
+					 bool revoked)
+{
+}
+#endif
+
 #endif
include/linux/vfio_pci_core.h +1 -0
index fbb472dd99b3..da5d8955ae56 100644
@@ -94,6 +94,7 @@ struct vfio_pci_core_device {
 	struct vfio_pci_core_device	*sriov_pf_core_dev;
 	struct notifier_block	nb;
 	struct rw_semaphore	memory_lock;
+	struct list_head	dmabufs;
 };
 
 /* Will be exported for vfio pci drivers usage */
include/uapi/linux/vfio.h +18 -0
index 2b68e6cdf190..8812b4750cc5 100644
@@ -1458,6 +1458,24 @@ struct vfio_device_feature_bus_master {
 };
 #define VFIO_DEVICE_FEATURE_BUS_MASTER 10
 
+/**
+ * Upon VFIO_DEVICE_FEATURE_GET create a dma_buf fd for the
+ * region selected.
+ *
+ * open_flags are the typical flags passed to open(2), eg O_RDWR, O_CLOEXEC,
+ * etc. offset/length specify a slice of the region to create the dmabuf from.
+ * If both are 0 then the whole region is used.
+ *
+ * Return: The fd number on success, -1 and errno is set on failure.
+ */
+struct vfio_device_feature_dma_buf {
+	__u32 region_index;
+	__u32 open_flags;
+	__u32 offset;
+	__u64 length;
+};
+#define VFIO_DEVICE_FEATURE_DMA_BUF 11
+
 /* -------- API for Type1 VFIO IOMMU -------- */
 
 /**

--
2.43.5

#4 lore raw

PATCH 4/4 vfio/pci: Allow export dmabuf without move_notify from importer

In reply to Wei Lin Guay

From: Wei Lin Guay <[email protected]>

Summary:
Support vfio to export dmabuf to importer such as RDMA NIC that does not support move_notify callback, since not all RDMA driver support on-demand-paging (ODP).

There are some use-cases such as bind accelerator that always pinned the device memory via vfio and export it to RDMA NIC such as EFA, BNXT_RE or IRDMA that does not support ODP.

---
 drivers/vfio/pci/dma_buf.c       | 32 +++++++++++++++++++++++++++-----
 drivers/vfio/pci/vfio_pci_core.c | 16 ++++++++++++++++
 drivers/vfio/pci/vfio_pci_priv.h |  7 +++++++
 3 files changed, 50 insertions(+), 5 deletions(-)
drivers/vfio/pci/dma_buf.c +27 -5
index fd772b520cd7..8017f48296cb 100644
@@ -17,6 +17,7 @@ struct vfio_pci_dma_buf {
 	unsigned int orig_nents;
 	size_t offset;
 	bool revoked;
+	bool pinned;
 };
 
 static int vfio_pci_dma_buf_attach(struct dma_buf *dmabuf,
@@ -32,17 +33,38 @@ static int vfio_pci_dma_buf_attach(struct dma_buf *dmabuf,
 	return 0;
 }
 
+bool vfio_pci_dma_buf_pinned(struct vfio_pci_core_device *vdev)
+{
+	struct vfio_pci_dma_buf *priv;
+	struct vfio_pci_dma_buf *tmp;
+	bool pinned = false;
+
+	down_write(&vdev->memory_lock);
+	list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) {
+		if (!dma_buf_try_get(priv->dmabuf))
+			continue;
+		if (priv->pinned) {
+			pinned = true;
+			break;
+		}
+	}
+	up_write(&vdev->memory_lock);
+	return pinned;
+}
+
 static void vfio_pci_dma_buf_unpin(struct dma_buf_attachment *attachment)
 {
+	struct vfio_pci_dma_buf *priv = attachment->dmabuf->priv;
+
+	priv->pinned = false;
 }
 
 static int vfio_pci_dma_buf_pin(struct dma_buf_attachment *attachment)
 {
-	/*
-	 * Uses the dynamic interface but must always allow for
-	 * dma_buf_move_notify() to do revoke
-	 */
-	return -EINVAL;
+	struct vfio_pci_dma_buf *priv = attachment->dmabuf->priv;
+
+	priv->pinned = true;
+	return 0;
 }
 
 static struct sg_table *
drivers/vfio/pci/vfio_pci_core.c +16 -0
index bb97b4d94eb7..db28fa2cc9a8 100644
@@ -1246,6 +1246,13 @@ static int vfio_pci_ioctl_reset(struct vfio_pci_core_device *vdev,
 	 */
 	vfio_pci_set_power_state(vdev, PCI_D0);
 
+	/*
+	 * prevent reset if dma_buf is pinned to avoid stale pinned
+	 * expose to the dmabuf exporter.
+	 */
+	if (vfio_pci_dma_buf_pinned(vdev))
+		return -EINVAL;
+
 	vfio_pci_dma_buf_move(vdev, true);
 	ret = pci_try_reset_function(vdev->pdev);
 	if (__vfio_pci_memory_enabled(vdev))
@@ -2444,6 +2451,15 @@ static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set,
 			break;
 		}
 
+		/*
+		 * prevent reset if dma_buf is pinned to avoid stale pinned
+		 * expose to the dmabuf exporter.
+		 */
+		if (vfio_pci_dma_buf_pinned(vdev)) {
+			ret = -EINVAL;
+			break;
+		}
+
 		/*
 		 * Take the memory write lock for each device and zap BAR
 		 * mappings to prevent the user accessing the device while in
drivers/vfio/pci/vfio_pci_priv.h +7 -0
index 09d3c300918c..43c40dc4751c 100644
@@ -107,6 +107,7 @@ int vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
 				  size_t argsz);
 void vfio_pci_dma_buf_cleanup(struct vfio_pci_core_device *vdev);
 void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool revoked);
+bool vfio_pci_dma_buf_pinned(struct vfio_pci_core_device *vdev);
 #else
 static int
 vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
@@ -115,6 +116,12 @@ vfio_pci_core_feature_dma_buf(struct vfio_pci_core_device *vdev, u32 flags,
 {
 	return -ENOTTY;
 }
+
+static inline bool vfio_pci_dma_buf_pinned(struct vfio_pci_core_device *vdev)
+{
+	return false;
+}
+
 static inline void vfio_pci_dma_buf_cleanup(struct vfio_pci_core_device *vdev)
 {
 }

--
2.43.5

maintainer
#5 lore raw
In reply to Wei Lin Guay

Am 16.12.24 um 10:54 schrieb Wei Lin Guay:

Quoted 26 lines

From: Wei Lin Guay <[email protected]>

This is another attempt to revive the patches posted by Jason Gunthorpe and Vivek Kasireddy, at
https://patchwork.kernel.org/project/linux-media/cover/[email protected]/
https://lwn.net/Articles/970751/

In addition to the initial proposal by Jason, another promising application is exposing memory from an AI accelerator (bound to VFIO) to an RDMA device. This would allow the RDMA device to directly access the accelerator's memory, thereby facilitating direct data transactions between the RDMA device and the accelerator.

Below is from the text/motivation from the orginal cover letter.

dma-buf has become a way to safely acquire a handle to non-struct page memory that can still have lifetime controlled by the exporter. Notably RDMA can now import dma-buf FDs and build them into MRs which allows for PCI P2P operations. Extend this to allow vfio-pci to export MMIO memory from PCI device BARs.

This series supports a use case for SPDK where a NVMe device will be owned by SPDK through VFIO but interacting with a RDMA device. The RDMA device may directly access the NVMe CMB or directly manipulate the NVMe device's doorbell using PCI P2P.

Quoted 6 lines

However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

Yeah that sounds like it should work.

But where is the rest of the series, I only see the cover letter?

Quoted 17 lines

v2:
- Name the new file dma_buf.c
- Restore orig_nents before freeing
- Fix reversed logic around priv->revoked
- Set priv->index
- Rebased on v2 "Break up ioctl dispatch functions"
v1: https://lore.kernel.org/r/[email protected]

Cc Oded Gabbay <[email protected]>
Cc Christian König <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Leon Romanovsky <[email protected]>
Cc Maor Gottlieb <[email protected]>
Quoted 3 lines

Jason Gunthorpe (3):
vfio: Add vfio_device_get()
dma-buf: Add dma_buf_try_get()

That is usually a no-go. We have rejected adding dma_buf_try_get() multiple times.

Please explain *exactly* what you need that for and how you protect against races with tearndown.

Regards,
Christian.

Quoted 20 lines

vfio/pci: Allow MMIO regions to be exported through dma-buf

Wei Lin Guay (1):
vfio/pci: Allow export dmabuf without move_notify from importer

  drivers/vfio/pci/Makefile          |   1 +
  drivers/vfio/pci/dma_buf.c         | 291 +++++++++++++++++++++++++++++
  drivers/vfio/pci/vfio_pci_config.c |   8 +-
  drivers/vfio/pci/vfio_pci_core.c   |  44 ++++-
  drivers/vfio/pci/vfio_pci_priv.h   |  30 +++
  drivers/vfio/vfio_main.c           |   1 +
  include/linux/dma-buf.h            |  13 ++
  include/linux/vfio.h               |   6 +
  include/linux/vfio_pci_core.h      |   1 +
  include/uapi/linux/vfio.h          |  18 ++
  10 files changed, 405 insertions(+), 8 deletions(-)

create mode 100644 drivers/vfio/pci/dma_buf.c

--
2.43.5

#6 lore raw
Quoted 6 lines · Christian König

Am 16.12.24 um 10:54 schrieb Wei Lin Guay:

Quoted 5 lines

From: Wei Lin Guay <[email protected]>
However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

Quoted 6 lines
Quoted 2 lines

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

Yeah that sounds like it should work.

But where is the rest of the series, I only see the cover letter?

Should be here:

https://lore.kernel.org/linux-rdma/[email protected]/T/#u

maintainer
#7 lore raw

Am 16.12.24 um 17:54 schrieb Keith Busch:

Quoted 7 lines · Christian König
Quoted 6 lines

Am 16.12.24 um 10:54 schrieb Wei Lin Guay:

Quoted 5 lines

From: Wei Lin Guay <[email protected]>
However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

Yeah that sounds like it should work.

But where is the rest of the series, I only see the cover letter?

Should be here:

   https://lore.kernel.org/linux-rdma/[email protected]/T/#u

Please send that out once more with me on explicit CC.

Apart from that I have to reject the adding of dma_buf_try_get(), that is clearly not something we should do.

Thanks,
Christian.

#8 lore raw

Hi Christian,

Thanks again for your prompt response/review.

Quoted 13 lines · Christian König
Quoted 2 lines

Am 16.12.24 um 17:54 schrieb Keith Busch:

Quoted 9 lines · Christian König
Quoted 8 lines

Am 16.12.24 um 10:54 schrieb Wei Lin Guay:

Quoted 7 lines

From: Wei Lin Guay <[email protected]>
However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

Yeah that sounds like it should work.

But where is the rest of the series, I only see the cover letter?

Should be here:

https://lore.kernel.org/linux-rdma/[email protected]/T/#u

Please send that out once more with me on explicit CC.

I will resend the patch series. I was experiencing issues with my email client, which inadvertently split the series into two separate emails.

Quoted 2 lines

Apart from that I have to reject the adding of dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his updated patch series

Thanks,
Wei Lin

Quoted 3 lines

Thanks,
Christian.

maintainer
#9 lore raw

Am 17.12.24 um 12:06 schrieb Wei Lin Guay:

Quoted 3 lines

[SNIP]

Quoted 1 line

Please send that out once more with me on explicit CC.

I will resend the patch series. I was experiencing issues with my email client, which inadvertently split the series into two separate emails.

Alternatively I can also copy the code from the list archive and explain why that doesn't work:

+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool revoked) +{
+    struct vfio_pci_dma_buf *priv;
+    struct vfio_pci_dma_buf *tmp;
+
+    lockdep_assert_held_write(&vdev->memory_lock);

This makes sure that the caller is holding vdev->memory_lock.

+
+    list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) { +        if (!dma_buf_try_get(priv->dmabuf))

This here only works because vfio_pci_dma_buf_release() also grabs vdev->memory_lock and so we are protected against concurrent releases.

+            continue;
+        if (priv->revoked != revoked) {
+            dma_resv_lock(priv->dmabuf->resv, NULL); +            priv->revoked = revoked;
+            dma_buf_move_notify(priv->dmabuf); +            dma_resv_unlock(priv->dmabuf->resv); +        }
+        dma_buf_put(priv->dmabuf);

The problem is now that this here might drop the last reference which in turn calls vfio_pci_dma_buf_release() which also tries to grab vdev->memory_lock and so results in a deadlock.

+    }
+}

This pattern was suggested multiple times and I had to rejected it every time because the whole idea is just fundamentally broken.

It's really astonishing how people always come up with the same broken pattern.

Regards,
Christian.

Quoted 10 lines
Quoted 2 lines

Apart from that I have to reject the adding of dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his updated patch series

Thanks,
Wei Lin

Quoted 2 lines

Thanks,
Christian.

#10 lore raw

Hi Christian,

Quoted 36 lines

Subject: Re: [PATCH 0/4] cover-letter: Allow MMIO regions to be exported through dmabuf

Quoted 2 lines

I will resend the patch series. I was experiencing issues with my email
client, which inadvertently split the series into two separate emails.

Alternatively I can also copy the code from the list archive and explain why that doesn't work:

(quoted diff) +6 -0
+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool
+revoked) {
+    struct vfio_pci_dma_buf *priv;
+    struct vfio_pci_dma_buf *tmp;
+
+    lockdep_assert_held_write(&vdev->memory_lock);

This makes sure that the caller is holding vdev->memory_lock.

(quoted diff) +3 -0
+
+    list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) {
+        if (!dma_buf_try_get(priv->dmabuf))

This here only works because vfio_pci_dma_buf_release() also grabs vdev-

Quoted 1 line

memory_lock and so we are protected against concurrent releases.

(quoted diff) +8 -0
+            continue;
+        if (priv->revoked != revoked) {
+            dma_resv_lock(priv->dmabuf->resv, NULL);
+            priv->revoked = revoked;
+            dma_buf_move_notify(priv->dmabuf);
+            dma_resv_unlock(priv->dmabuf->resv);
+        }
+        dma_buf_put(priv->dmabuf);
Quoted 3 lines

The problem is now that this here might drop the last reference which in turn calls vfio_pci_dma_buf_release() which also tries to grab vdev-

Quoted 1 line

memory_lock and so results in a deadlock.

AFAICS, vfio_pci_dma_buf_release() would not be called synchronously after the last reference is dropped by dma_buf_put(). This is because fput(), which is called by dma_buf_put() triggers f_op->release() asynchronously; therefore, a deadlock is unlikely to occur in this scenario, unless I am overlooking something.

Thanks,
Vivek

Quoted 39 lines
(quoted diff) +2 -0
+    }
+}

This pattern was suggested multiple times and I had to rejected it every time because the whole idea is just fundamentally broken.

It's really astonishing how people always come up with the same broken pattern.

Regards,
Christian.

Apart from that I have to reject the adding of
dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has
resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his updated patch series

Thanks,
Wei Lin

Thanks,
Christian.

maintainer
#11 lore raw

Am 18.12.24 um 07:16 schrieb Kasireddy, Vivek:

Quoted 36 lines

Hi Christian,

Quoted 34 lines

Subject: Re: [PATCH 0/4] cover-letter: Allow MMIO regions to be exported through dmabuf

Quoted 2 lines

I will resend the patch series. I was experiencing issues with my email
client, which inadvertently split the series into two separate emails.

Alternatively I can also copy the code from the list archive and explain why that doesn't work:

(quoted diff) +6 -0
+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool
+revoked) {
+    struct vfio_pci_dma_buf *priv;
+    struct vfio_pci_dma_buf *tmp;
+
+    lockdep_assert_held_write(&vdev->memory_lock);

This makes sure that the caller is holding vdev->memory_lock.

(quoted diff) +3 -0
+
+    list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) {
+        if (!dma_buf_try_get(priv->dmabuf))

This here only works because vfio_pci_dma_buf_release() also grabs vdev-

Quoted 1 line

memory_lock and so we are protected against concurrent releases.

(quoted diff) +8 -0
+            continue;
+        if (priv->revoked != revoked) {
+            dma_resv_lock(priv->dmabuf->resv, NULL);
+            priv->revoked = revoked;
+            dma_buf_move_notify(priv->dmabuf);
+            dma_resv_unlock(priv->dmabuf->resv);
+        }
+        dma_buf_put(priv->dmabuf);
Quoted 7 lines
Quoted 3 lines

The problem is now that this here might drop the last reference which in turn calls vfio_pci_dma_buf_release() which also tries to grab vdev-

Quoted 1 line

memory_lock and so results in a deadlock.

AFAICS, vfio_pci_dma_buf_release() would not be called synchronously after the last reference is dropped by dma_buf_put(). This is because fput(), which is called by dma_buf_put() triggers f_op->release() asynchronously; therefore, a deadlock is unlikely to occur in this scenario, unless I am overlooking something.

My recollection is that the f_op->release handler is only called asynchronously if fput() was issued in interrupt context.

But could be that this information is outdated.

Regards,
Christian.

Quoted 42 lines

Thanks,
Vivek

Quoted 38 lines
(quoted diff) +2 -0
+    }
+}

This pattern was suggested multiple times and I had to rejected it every time because the whole idea is just fundamentally broken.

It's really astonishing how people always come up with the same broken pattern.

Regards,
Christian.

Apart from that I have to reject the adding of
dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has
resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his updated patch series

Thanks,
Wei Lin

Thanks,
Christian.

#12 lore raw

Hi Christian,

Quoted 41 lines

Subject: Re: [PATCH 0/4] cover-letter: Allow MMIO regions to be exported through dmabuf

Quoted 14 lines
Quoted 14 lines
Quoted 3 lines

I will resend the patch series. I was experiencing issues with my email
client, which inadvertently split the series into two separate emails.

Alternatively I can also copy the code from the list archive and explain why that doesn't work:

(quoted diff) +6 -0
+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool
+revoked) {
+    struct vfio_pci_dma_buf *priv;
+    struct vfio_pci_dma_buf *tmp;
+
+    lockdep_assert_held_write(&vdev->memory_lock);
Quoted 7 lines
Quoted 7 lines

This makes sure that the caller is holding vdev->memory_lock.

(quoted diff) +3 -0
+
+    list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) {
+        if (!dma_buf_try_get(priv->dmabuf))

This here only works because vfio_pci_dma_buf_release() also grabs

vdev-

Quoted 11 lines
Quoted 11 lines
Quoted 1 line

memory_lock and so we are protected against concurrent releases.

(quoted diff) +8 -0
+            continue;
+        if (priv->revoked != revoked) {
+            dma_resv_lock(priv->dmabuf->resv, NULL);
+            priv->revoked = revoked;
+            dma_buf_move_notify(priv->dmabuf);
+            dma_resv_unlock(priv->dmabuf->resv);
+        }
+        dma_buf_put(priv->dmabuf);

The problem is now that this here might drop the last reference which in

turn

Quoted 3 lines
Quoted 2 lines

calls vfio_pci_dma_buf_release() which also tries to grab vdev-

Quoted 1 line

memory_lock and so results in a deadlock.

AFAICS, vfio_pci_dma_buf_release() would not be called synchronously

after the

Quoted 8 lines
Quoted 1 line

last reference is dropped by dma_buf_put(). This is because fput(), which is

called

Quoted 1 line

by dma_buf_put() triggers f_op->release() asynchronously; therefore, a

deadlock

Quoted 1 line

is unlikely to occur in this scenario, unless I am overlooking something.

My recollection is that the f_op->release handler is only called asynchronously if fput() was issued in interrupt context.

Here is the code of fput() from the current master:
void fput(struct file *file)

{
        if (file_ref_put(&file->f_ref)) {
                struct task_struct *task = current;

                if (unlikely(!(file->f_mode & (FMODE_BACKING | FMODE_OPENED)))) {
                        file_free(file);
                        return;
                }
                if (likely(!in_interrupt() && !(task->flags & PF_KTHREAD))) {
                        init_task_work(&file->f_task_work, ____fput);
                        if (!task_work_add(task, &file->f_task_work, TWA_RESUME))
                                return;
                        /*
                         * After this task has run exit_task_work(),
                         * task_work_add() will fail.  Fall through to delayed
                         * fput to avoid leaking *file.
                         */
                }

                if (llist_add(&file->f_llist, &delayed_fput_list))
                        schedule_delayed_work(&delayed_fput_work, 1);
        }
}

IIUC, based on the above code, it looks like there are two ways in which the f_op->release() handler is triggered from fput():
- via delayed_fput() for kernel threads and code in interrupt context
- via task_work_run() just before the task/process returns to the user-mode

The first scenario above is definitely asynchronous as the release() handler is called from a worker thread. But I think the second case (which is the most common and relevant for our use-case) can also be considered asynchronous, because the execution of the system call or ioctl that led to the context in which dma_buf_put() was called is completed.

Here is a trace from my light testing with the udmabuf driver, where you can see the release() handler being called by syscall_exit_to_user_mode() :

[  158.464203] Call Trace:
[  158.466681]  <TASK>
[  158.468815]  dump_stack_lvl+0x60/0x80
[  158.472507]  dump_stack+0x14/0x16
[  158.475853]  release_udmabuf+0x2f/0x9f
[  158.479631]  dma_buf_release+0x3c/0x90
[  158.483408]  __dentry_kill+0x8f/0x180
[  158.487098]  dput+0xe7/0x1a0
[  158.490013]  __fput+0x131/0x2b0
[  158.493178]  ____fput+0x19/0x20
[  158.496352]  task_work_run+0x61/0x90
[  158.499959]  syscall_exit_to_user_mode+0x1a4/0x1b0
[  158.504769]  do_syscall_64+0x5b/0x110
[  158.508458]  entry_SYSCALL_64_after_hwframe+0x4b/0x53

And, here is the relevant syscall code (from arch/x86/entry/common.c):
__visible noinstr bool do_syscall_64(struct pt_regs *regs, int nr)

{
        add_random_kstack_offset();
        nr = syscall_enter_from_user_mode(regs, nr);

        instrumentation_begin();

        if (!do_syscall_x64(regs, nr) && !do_syscall_x32(regs, nr) && nr != -1) {
                /* Invalid system call, but still a system call. */
                regs->ax = __x64_sys_ni_syscall(regs);
        }

        instrumentation_end();
        syscall_exit_to_user_mode(regs);

I also confirmed that the release() handler is indeed called after dma_buf_put() (and not by dma_buf_put()) by adding debug prints before and after dma_buf_put() and one in the release() handler. Furthermore, I also found that calling close() on the dmabuf fd in the user-space is one scenario in which fput() is called synchronously. Here is the relevant trace:

[  302.770910] Call Trace:
[  302.773389]  <TASK>
[  302.775516]  dump_stack_lvl+0x60/0x80
[  302.779209]  dump_stack+0x14/0x16
[  302.782549]  release_udmabuf+0x2f/0x9f
[  302.786329]  dma_buf_release+0x3c/0x90
[  302.790105]  __dentry_kill+0x8f/0x180
[  302.793789]  dput+0xe7/0x1a0
[  302.796703]  __fput+0x131/0x2b0
[  302.799866]  __fput_sync+0x53/0x70
[  302.803299]  __x64_sys_close+0x58/0xc0
[  302.807076]  x64_sys_call+0x126a/0x17d0
[  302.810938]  do_syscall_64+0x4f/0x110
[  302.814622]  entry_SYSCALL_64_after_hwframe+0x4b/0x53

As you can see above, there is indeed a synchronous version of fput() defined just below fput():
/*
* synchronous analog of fput(); for kernel threads that might be needed
* in some umount() (and thus can't use flush_delayed_fput() without
* risking deadlocks), need to wait for completion of __fput() and know
* for this specific struct file it won't involve anything that would
* need them. Use only if you really need it - at the very least,
* don't blindly convert fput() by kernel thread to that.
*/
void __fput_sync(struct file *file)

{
	if (file_ref_put(&file->f_ref))
		__fput(file);
}

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Thanks,
Vivek

Quoted 50 lines

But could be that this information is outdated.

Regards,
Christian.

Quoted 8 lines

Thanks,
Vivek

Quoted 4 lines
(quoted diff) +2 -0
+    }
+}

This pattern was suggested multiple times and I had to rejected it every

time

Quoted 14 lines
Quoted 14 lines

because the whole idea is just fundamentally broken.

It's really astonishing how people always come up with the same broken pattern.

Regards,
Christian.

Quoted 8 lines
Quoted 8 lines

Apart from that I have to reject the adding of
dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has
resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his

updated

Quoted 12 lines
Quoted 12 lines

patch series

Thanks,
Wei Lin

Thanks,
Christian.

maintainer
#13 lore raw

Am 19.12.24 um 08:02 schrieb Kasireddy, Vivek:

Quoted 149 lines

Hi Christian,

Quoted 38 lines

Subject: Re: [PATCH 0/4] cover-letter: Allow MMIO regions to be exported through dmabuf

Quoted 12 lines
Quoted 12 lines
Quoted 2 lines

I will resend the patch series. I was experiencing issues with my email
client, which inadvertently split the series into two separate emails.

Alternatively I can also copy the code from the list archive and explain why that doesn't work:

(quoted diff) +6 -0
+void vfio_pci_dma_buf_move(struct vfio_pci_core_device *vdev, bool
+revoked) {
+    struct vfio_pci_dma_buf *priv;
+    struct vfio_pci_dma_buf *tmp;
+
+    lockdep_assert_held_write(&vdev->memory_lock);
Quoted 7 lines
Quoted 7 lines

This makes sure that the caller is holding vdev->memory_lock.

(quoted diff) +3 -0
+
+    list_for_each_entry_safe(priv, tmp, &vdev->dmabufs, dmabufs_elm) {
+        if (!dma_buf_try_get(priv->dmabuf))

This here only works because vfio_pci_dma_buf_release() also grabs

vdev-

Quoted 11 lines
Quoted 11 lines
Quoted 1 line

memory_lock and so we are protected against concurrent releases.

(quoted diff) +8 -0
+            continue;
+        if (priv->revoked != revoked) {
+            dma_resv_lock(priv->dmabuf->resv, NULL);
+            priv->revoked = revoked;
+            dma_buf_move_notify(priv->dmabuf);
+            dma_resv_unlock(priv->dmabuf->resv);
+        }
+        dma_buf_put(priv->dmabuf);

The problem is now that this here might drop the last reference which in

turn

Quoted 3 lines
Quoted 2 lines

calls vfio_pci_dma_buf_release() which also tries to grab vdev-

Quoted 1 line

memory_lock and so results in a deadlock.

AFAICS, vfio_pci_dma_buf_release() would not be called synchronously

Quoted 8 lines

after the

Quoted 1 line

last reference is dropped by dma_buf_put(). This is because fput(), which is

called

Quoted 1 line

by dma_buf_put() triggers f_op->release() asynchronously; therefore, a

deadlock

Quoted 1 line

is unlikely to occur in this scenario, unless I am overlooking something.

My recollection is that the f_op->release handler is only called asynchronously if fput() was issued in interrupt context.

Here is the code of fput() from the current master:
void fput(struct file *file)

{
         if (file_ref_put(&file->f_ref)) {
                 struct task_struct *task = current;

                 if (unlikely(!(file->f_mode & (FMODE_BACKING | FMODE_OPENED)))) {
                         file_free(file);
                         return;
                 }
                 if (likely(!in_interrupt() && !(task->flags & PF_KTHREAD))) {
                         init_task_work(&file->f_task_work, ____fput);
                         if (!task_work_add(task, &file->f_task_work, TWA_RESUME))
                                 return;
                         /*
                          * After this task has run exit_task_work(),
                          * task_work_add() will fail.  Fall through to delayed
                          * fput to avoid leaking *file.
                          */
                 }

                 if (llist_add(&file->f_llist, &delayed_fput_list))
                         schedule_delayed_work(&delayed_fput_work, 1);
         }
}

IIUC, based on the above code, it looks like there are two ways in which the f_op->release() handler is triggered from fput():
- via delayed_fput() for kernel threads and code in interrupt context
- via task_work_run() just before the task/process returns to the user-mode

The first scenario above is definitely asynchronous as the release() handler is called from a worker thread. But I think the second case (which is the most common and relevant for our use-case) can also be considered asynchronous, because the execution of the system call or ioctl that led to the context in which dma_buf_put() was called is completed.

Here is a trace from my light testing with the udmabuf driver, where you can see the release() handler being called by syscall_exit_to_user_mode() :

[  158.464203] Call Trace:
[  158.466681]  <TASK>
[  158.468815]  dump_stack_lvl+0x60/0x80
[  158.472507]  dump_stack+0x14/0x16
[  158.475853]  release_udmabuf+0x2f/0x9f
[  158.479631]  dma_buf_release+0x3c/0x90
[  158.483408]  __dentry_kill+0x8f/0x180
[  158.487098]  dput+0xe7/0x1a0
[  158.490013]  __fput+0x131/0x2b0
[  158.493178]  ____fput+0x19/0x20
[  158.496352]  task_work_run+0x61/0x90
[  158.499959]  syscall_exit_to_user_mode+0x1a4/0x1b0
[  158.504769]  do_syscall_64+0x5b/0x110
[  158.508458]  entry_SYSCALL_64_after_hwframe+0x4b/0x53

And, here is the relevant syscall code (from arch/x86/entry/common.c):
__visible noinstr bool do_syscall_64(struct pt_regs *regs, int nr)

{
         add_random_kstack_offset();
         nr = syscall_enter_from_user_mode(regs, nr);

         instrumentation_begin();

         if (!do_syscall_x64(regs, nr) && !do_syscall_x32(regs, nr) && nr != -1) {
                 /* Invalid system call, but still a system call. */
                 regs->ax = __x64_sys_ni_syscall(regs);
         }

         instrumentation_end();
         syscall_exit_to_user_mode(regs);

I also confirmed that the release() handler is indeed called after dma_buf_put() (and not by dma_buf_put()) by adding debug prints before and after dma_buf_put() and one in the release() handler. Furthermore, I also found that calling close() on the dmabuf fd in the user-space is one scenario in which fput() is called synchronously. Here is the relevant trace:

[  302.770910] Call Trace:
[  302.773389]  <TASK>
[  302.775516]  dump_stack_lvl+0x60/0x80
[  302.779209]  dump_stack+0x14/0x16
[  302.782549]  release_udmabuf+0x2f/0x9f
[  302.786329]  dma_buf_release+0x3c/0x90
[  302.790105]  __dentry_kill+0x8f/0x180
[  302.793789]  dput+0xe7/0x1a0
[  302.796703]  __fput+0x131/0x2b0
[  302.799866]  __fput_sync+0x53/0x70
[  302.803299]  __x64_sys_close+0x58/0xc0
[  302.807076]  x64_sys_call+0x126a/0x17d0
[  302.810938]  do_syscall_64+0x4f/0x110
[  302.814622]  entry_SYSCALL_64_after_hwframe+0x4b/0x53

As you can see above, there is indeed a synchronous version of fput() defined just below fput():
/*
* synchronous analog of fput(); for kernel threads that might be needed
* in some umount() (and thus can't use flush_delayed_fput() without
* risking deadlocks), need to wait for completion of __fput() and know
* for this specific struct file it won't involve anything that would
* need them. Use only if you really need it - at the very least,
* don't blindly convert fput() by kernel thread to that.
*/
void __fput_sync(struct file *file)

Quoted 8 lines
{
	if (file_ref_put(&file->f_ref))
		__fput(file);
}

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Yeah, I agree.

Interesting to know, I wasn't aware that the task_work functionality exists for that use case.

Thanks,
Christian.

Quoted 52 lines

Thanks,
Vivek

Quoted 48 lines

But could be that this information is outdated.

Regards,
Christian.

Quoted 7 lines

Thanks,
Vivek

Quoted 4 lines
(quoted diff) +2 -0
+    }
+}

This pattern was suggested multiple times and I had to rejected it every

time

Quoted 14 lines
Quoted 14 lines

because the whole idea is just fundamentally broken.

It's really astonishing how people always come up with the same broken pattern.

Regards,
Christian.

Quoted 8 lines
Quoted 8 lines

Apart from that I have to reject the adding of
dma_buf_try_get(), that is clearly not something we should do.

Understood. It appears that Vivek has confirmed that his v2 has
resolved the issue. I will follow up with him to determine if he plans to resume his patch, and if so, I will apply my last patch on top of his

updated

Quoted 12 lines
Quoted 12 lines

patch series

Thanks,
Wei Lin

Thanks,
Christian.

maintainer
#14 lore raw
Quoted 3 lines · Christian König
Quoted 3 lines

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Right, there is no deadlock here, and there is nothing inhernetly wrong with using try_get like this. The locking here is ugly ugly ugly, I forget why, but this was the best I came up with to untangle it without deadlocks or races.

Quoted 4 lines

Yeah, I agree.

Interesting to know, I wasn't aware that the task_work functionality exists for that use case.

IIRC it was changed a while back because call chains were getting kind of crazy long with the file release functions and stack overflow was a problem in some cases.

Jason

#15 lore raw
Quoted 10 lines · Jason Gunthorpe
Quoted 3 lines · Christian König
Quoted 3 lines

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Right, there is no deadlock here, and there is nothing inhernetly wrong with using try_get like this. The locking here is ugly ugly ugly, I forget why, but this was the best I came up with to untangle it without deadlocks or races.

Yeah, imo try_get is perfectly fine pattern. With that sorted my only request is to make the try_get specific to the dma_ops, because it only works if both ->release and the calling context of try_get follow the same rules, which by necessity are exporter specific.

In full generality as a dma_buf.c interface it's just busted and there's no way to make it work, unless we inflict that locking ugliness on absolutely every exporter.

Quoted 8 lines
Quoted 4 lines

Yeah, I agree.

Interesting to know, I wasn't aware that the task_work functionality exists for that use case.

IIRC it was changed a while back because call chains were getting kind of crazy long with the file release functions and stack overflow was a problem in some cases.

Hm I thought it was also a "oh fuck deadlocks" moment, because usually most of the very deep fput calls are for temporary reference and not the final one. Unless userspace is sneaky and drops its own fd reference in a separate thread with close(), then everything goes boom. And untangling all these was impossible.
-Sima

Signature · 3 lines
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
maintainer
#16 lore raw
Quoted 12 lines · Simona Vetter
Quoted 10 lines · Jason Gunthorpe
Quoted 3 lines · Christian König
Quoted 3 lines

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Right, there is no deadlock here, and there is nothing inhernetly wrong with using try_get like this. The locking here is ugly ugly ugly, I forget why, but this was the best I came up with to untangle it without deadlocks or races.

Quoted 4 lines

Yeah, imo try_get is perfectly fine pattern. With that sorted my only request is to make the try_get specific to the dma_ops, because it only works if both ->release and the calling context of try_get follow the same rules, which by necessity are exporter specific.

We already know the fd is a dma_ops one because it is on an internal list and it could not get there otherwise.

The pointer cannot become invalid and freed back to the type safe RCU while on the list, meaning the try_get is safe as is.

I think Christian's original advice to just open code it is the best option.

Quoted 3 lines

In full generality as a dma_buf.c interface it's just busted and there's no way to make it work, unless we inflict that locking ugliness on absolutely every exporter.

I'm not sure about that, the struct file code has special logic to accommodate the type safe RCU trick. I didn't try to digest it fully, but I expect there are ways to use it safely without the locking on release.

Quoted 7 lines
Quoted 3 lines

IIRC it was changed a while back because call chains were getting kind of crazy long with the file release functions and stack overflow was a problem in some cases.

Hm I thought it was also a "oh fuck deadlocks" moment, because usually most of the very deep fput calls are for temporary reference and not the final one.

That sounds motivating too, but I've also seen cases of being careful to unlock before fputting things..

Jason

#17 lore raw
Quoted 18 lines · Jason Gunthorpe
Quoted 12 lines · Simona Vetter
Quoted 10 lines · Jason Gunthorpe
Quoted 3 lines · Christian König
Quoted 3 lines

Based on all the above, I think we can conclude that since dma_buf_put() does not directly (or synchronously) call the f_op->release() handler, a deadlock is unlikely to occur in the scenario you described.

Right, there is no deadlock here, and there is nothing inhernetly wrong with using try_get like this. The locking here is ugly ugly ugly, I forget why, but this was the best I came up with to untangle it without deadlocks or races.

Quoted 4 lines

Yeah, imo try_get is perfectly fine pattern. With that sorted my only request is to make the try_get specific to the dma_ops, because it only works if both ->release and the calling context of try_get follow the same rules, which by necessity are exporter specific.

Quoted 8 lines

We already know the fd is a dma_ops one because it is on an internal list and it could not get there otherwise.

The pointer cannot become invalid and freed back to the type safe RCU while on the list, meaning the try_get is safe as is.

I think Christian's original advice to just open code it is the best option.

Yeah open coding in vfio is imo best, agreed on that.

Quoted 8 lines
Quoted 3 lines

In full generality as a dma_buf.c interface it's just busted and there's no way to make it work, unless we inflict that locking ugliness on absolutely every exporter.

I'm not sure about that, the struct file code has special logic to accommodate the type safe RCU trick. I didn't try to digest it fully, but I expect there are ways to use it safely without the locking on release.

Hm yes, if you use a write barrier when set your file pointer and clear it in release, then you can use get_file_rcu with just rcu_read_lock on the read side. But you have to directly use your own struct file * pointer since it needs to reload that in a loop, you can't use dma_buf->file.

At that point you're already massively peeking behind the dma_buf abstraction that just directly using get_file_rcu is imo better.

And for anything else, whether it's rcu or plain locks, it's all subsystem specific anyway that I think a dma_buf.c wrapper

Not entirely sure, but for the dma_buf_try_get wrapper you have to put a full lock acquisition or rcu_sychnronize into your ->release callback, otherwise I don't think it's safe to use.

Quoted 10 lines
Quoted 7 lines
Quoted 3 lines

IIRC it was changed a while back because call chains were getting kind of crazy long with the file release functions and stack overflow was a problem in some cases.

Hm I thought it was also a "oh fuck deadlocks" moment, because usually most of the very deep fput calls are for temporary reference and not the final one.

That sounds motivating too, but I've also seen cases of being careful to unlock before fputting things..

Yeah I think knowledge about this issue was very uneven in different subsystems.
-Sima

Signature · 3 lines
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#18 lore raw
In reply to Christian König
Quoted 13 lines · Christian König

Am 16.12.24 um 17:54 schrieb Keith Busch:

Quoted 12 lines · Christian König
Quoted 11 lines

Am 16.12.24 um 10:54 schrieb Wei Lin Guay:

Quoted 7 lines

From: Wei Lin Guay <[email protected]>
However, as a general mechanism, it can support many other scenarios with VFIO. I imagine this dmabuf approach to be usable by iommufd as well for generic and safe P2P mappings.

This series goes after the "Break up ioctl dispatch functions to one function per ioctl" series.

Yeah that sounds like it should work.

But where is the rest of the series, I only see the cover letter?

Quoted 8 lines
Quoted 3 lines

Should be here:

   https://lore.kernel.org/linux-rdma/[email protected]/T/#u

Please send that out once more with me on explicit CC.

Apart from that I have to reject the adding of dma_buf_try_get(), that is clearly not something we should do.

Yeah if we do try_get it would need to be at least on a specific dma_buf type (so checking for dma_buf->ops), since in general this does not work (unless we add the general code in dma_buf.c somehow to make it work). Aside from any other design concerns.
-Sima

Signature · 3 lines
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch