Skip to contentSkip to conversation

Introduce generic context device bus for IOMMU context isolation

PATCH In review Ekansh Gupta 6 participants 15 messages Apr 14 – Apr 22

Subsystems (8)
lore
Share & tools
lore

Ekansh Gupta

author reviewer
#0 lore raw

cover Introduce generic context device bus for IOMMU context isolation

This series introduces a generic context device bus to replace driver-specific implementations and enable proper IOMMU handling of synthetic context bank devices used by accelerator and GPU drivers.

Background:
-----------
During review of the fastrpc driver changes [1], it was pointed out that using a platform driver for compute bank (CB) devices was incorrect since these are synthetic IOMMU context banks, not real platform devices. The platform driver approach also had a race condition where device nodes were created before channel resources were initialized, and probe was async, allowing applications to open devices before sessions were available.

The reviewer suggested following the pattern used in host1x_memory_context_list_init() and creating devices manually. During review of the QDA driver [2], it was further suggested to create a generic bus type that could be shared across multiple drivers rather than having each driver implement its own bus.

This series implements that suggestion by:
1. Creating a generic context_device_bus_type in drivers/base/
2. Migrating the existing host1x driver to use it
3. Converting fastrpc to use it, fixing the race condition in the process

References:
-----------
[1] Fastrpc platform driver review:
https://lore.kernel.org/all/golcrcr6voafr3fqsnihyjyut36sii55vzws4josfhkjjg3nie@ur43qq2kvlsv/

[2] QDA driver review requesting generic bus:
https://lore.kernel.org/all/en5tdgemre7vq5qihe6wnkbrro24vtzwfxwqrpnrrmugxoszs7@x2ox64zeznvd/

---
Ekansh Gupta (3):
drivers: base: Add generic context device bus
gpu: host1x: Migrate to generic context device bus
misc: fastrpc: Use context device bus for compute banks

 drivers/base/Kconfig             |   3 +
 drivers/base/Makefile            |   1 +
 drivers/base/context_bus.c       |  24 ++++++
 drivers/gpu/host1x/Kconfig       |   5 +-
 drivers/gpu/host1x/Makefile      |   1 -
 drivers/gpu/host1x/context.c     |   2 +-
 drivers/gpu/host1x/context.h     |   3 +-
 drivers/gpu/host1x/context_bus.c |  26 ------
 drivers/iommu/iommu.c            |   6 +-
 drivers/misc/Kconfig             |   1 +
 drivers/misc/fastrpc.c           | 180 +++++++++++++++++++++++++++------------
 include/linux/context_bus.h      |  15 ++++
 12 files changed, 174 insertions(+), 93 deletions(-)

---
base-commit: 1c7cc4904160c6fc6377564140062d68a3dc93a0
change-id: 20260414-computebus-e852f994ee25

Best regards,

Signature · 1 line
Ekansh Gupta <[email protected]>
author reviewer
#1 lore raw

PATCH 1/3 drivers: base: Add generic context device bus

Introduce a new generic bus type for synthetic context bank devices that require IOMMU context isolation. This bus provides a shared infrastructure for accelerator and GPU drivers that create virtual devices representing IOMMU context banks.

Currently, drivers like host1x implement their own bus types for context devices. This generic implementation allows multiple drivers to share the same bus infrastructure, simplifying the IOMMU subsystem integration and reducing code duplication.

---
 drivers/base/Kconfig        |  3 +++
 drivers/base/Makefile       |  1 +
 drivers/base/context_bus.c  | 24 ++++++++++++++++++++++++
 include/linux/context_bus.h | 15 +++++++++++++++
 4 files changed, 43 insertions(+)
drivers/base/Kconfig +3 -0
index f7d385cbd3ba..479bc4bb442b 100644
@@ -4,6 +4,9 @@ menu "Generic Driver Options"
 config AUXILIARY_BUS
 	bool
 
+config CONTEXT_DEVICE_BUS
+	bool
+
 config UEVENT_HELPER
 	bool "Support for uevent helper"
 	help
drivers/base/Makefile +1 -0
index 8074a10183dc..ab9a0b2dc73b 100644
@@ -8,6 +8,7 @@ obj-y			:= component.o core.o bus.o dd.o syscore.o \
 			   topology.o container.o property.o cacheinfo.o \
 			   swnode.o faux.o
 obj-$(CONFIG_AUXILIARY_BUS) += auxiliary.o
+obj-$(CONFIG_CONTEXT_DEVICE_BUS) += context_bus.o
 obj-$(CONFIG_DEVTMPFS)	+= devtmpfs.o
 obj-y			+= power/
 obj-$(CONFIG_ISA_BUS_API)	+= isa.o
drivers/base/context_bus.c new file +24 -0
new file mode 100644
index 000000000000..6ddb6c27bf69
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: GPL-2.0-only
+// Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+
+#include <linux/context_bus.h>
+#include <linux/init.h>
+
+const struct bus_type context_device_bus_type = {
+	.name = "context-device",
+};
+EXPORT_SYMBOL_GPL(context_device_bus_type);
+
+static int __init context_device_bus_init(void)
+{
+	int err;
+
+	err = bus_register(&context_device_bus_type);
+	if (err < 0) {
+		pr_err("context-device bus registration failed: %d\n", err);
+		return err;
+	}
+
+	return 0;
+}
+postcore_initcall(context_device_bus_init);
include/linux/context_bus.h new file +15 -0
new file mode 100644
index 000000000000..0cd44cb5b147
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+ */
+
+#ifndef __LINUX_CONTEXT_BUS_H
+#define __LINUX_CONTEXT_BUS_H
+
+#include <linux/device.h>
+
+#ifdef CONFIG_CONTEXT_DEVICE_BUS
+extern const struct bus_type context_device_bus_type;
+#endif
+
+#endif /* __LINUX_CONTEXT_BUS_H */
Signature · 1 line
2.34.1
#2 lore raw

Re: [PATCH 1/3] drivers: base: Add generic context device bus

Quoted 25 lines · On Tue, Apr 14, 2026 at 10:01:15PM +0530, Ekansh Gupta via B4 Relay wrote:

From: Ekansh Gupta <[email protected]>

Introduce a new generic bus type for synthetic context bank devices that require IOMMU context isolation. This bus provides a shared infrastructure for accelerator and GPU drivers that create virtual devices representing IOMMU context banks.

Currently, drivers like host1x implement their own bus types for context devices. This generic implementation allows multiple drivers to share the same bus infrastructure, simplifying the IOMMU subsystem integration and reducing code duplication.

---
 drivers/base/Kconfig        |  3 +++
 drivers/base/Makefile       |  1 +
 drivers/base/context_bus.c  | 24 ++++++++++++++++++++++++
 include/linux/context_bus.h | 15 +++++++++++++++
 4 files changed, 43 insertions(+)
drivers/base/Kconfig +0 -0
index f7d385cbd3ba..479bc4bb442b 100644

drivers/base/Kconfig menu "Generic Driver Options" · line 8

Quoted 5 lines
drivers/base/Kconfig +2 -0
@@ -4,6 +4,9 @@ menu "Generic Driver Options"
 config AUXILIARY_BUS
 	bool
 
+config CONTEXT_DEVICE_BUS
+	bool

So this is a context-free bus? I.e. no documentation?

:)

This feels really odd, this "bus" isn't doing anything at all. So why do you need it and can't use auxbus or faux?

You say something about an iommu, but yet this has no information about an iommu at all, so why are they somehow special?

You need to provide a lot more information here, sorry.

greg k-h

#3 lore raw

Re: [PATCH 1/3] drivers: base: Add generic context device bus

Quoted 34 lines · Greg Kroah-Hartman
Quoted 25 lines · On Tue, Apr 14, 2026 at 10:01:15PM +0530, Ekansh Gupta via B4 Relay wrote:

From: Ekansh Gupta <[email protected]>

Introduce a new generic bus type for synthetic context bank devices that require IOMMU context isolation. This bus provides a shared infrastructure for accelerator and GPU drivers that create virtual devices representing IOMMU context banks.

Currently, drivers like host1x implement their own bus types for context devices. This generic implementation allows multiple drivers to share the same bus infrastructure, simplifying the IOMMU subsystem integration and reducing code duplication.

---
  drivers/base/Kconfig        |  3 +++
  drivers/base/Makefile       |  1 +
  drivers/base/context_bus.c  | 24 ++++++++++++++++++++++++
  include/linux/context_bus.h | 15 +++++++++++++++
  4 files changed, 43 insertions(+)
drivers/base/Kconfig +0 -0
index f7d385cbd3ba..479bc4bb442b 100644

drivers/base/Kconfig menu "Generic Driver Options" · line 8

Quoted 5 lines
drivers/base/Kconfig +2 -0
@@ -4,6 +4,9 @@ menu "Generic Driver Options"
  config AUXILIARY_BUS
  	bool
 
+config CONTEXT_DEVICE_BUS
+	bool

So this is a context-free bus? I.e. no documentation?

Quoted 4 lines

:)

This feels really odd, this "bus" isn't doing anything at all. So why do you need it and can't use auxbus or faux?

Looks like the device have specific iommu configuration (via of_dma_configure()), then using the auxbus or faux bus types does not have the bus notifier calls to iommu_release_device(). Please correct me if I misunderstood?

Quoted 9 lines

You say something about an iommu, but yet this has no information about an iommu at all, so why are they somehow special?

You need to provide a lot more information here, sorry.

greg k-h

#4 lore raw

Re: [PATCH 1/3] drivers: base: Add generic context device bus

Quoted 4 lines · Vishnu Reddy

Looks like the device have specific iommu configuration (via of_dma_configure()), then using the auxbus or faux bus types does not have the bus notifier calls to iommu_release_device(). Please correct me if I misunderstood?

I think the general feeling is the open coding of_dma_configure() calls in host1x are not desirable, it should only ever be called from bus code.

If you need to create struct device instances bound to the iommu somehow this needs to be much robustly designed than wrappering the host1x stuff in a new bus.

Jason

#5 lore raw

Re: [PATCH 1/3] drivers: base: Add generic context device bus

In reply to Ekansh Gupta
Quoted 12 lines · On 4/14/26 4:31 PM, Ekansh Gupta via B4 Relay wrote:

From: Ekansh Gupta <[email protected]>

Introduce a new generic bus type for synthetic context bank devices that require IOMMU context isolation. This bus provides a shared infrastructure for accelerator and GPU drivers that create virtual devices representing IOMMU context banks.

Currently, drivers like host1x implement their own bus types for context devices. This generic implementation allows multiple drivers to share the same bus infrastructure, simplifying the IOMMU subsystem integration and reducing code duplication.

Quoted 6 lines
---
 drivers/base/Kconfig        |  3 +++
 drivers/base/Makefile       |  1 +
 drivers/base/context_bus.c  | 24 ++++++++++++++++++++++++
 include/linux/context_bus.h | 15 +++++++++++++++

This patch as it is does not add any details on how the iommu will be binded to this bus, Please consider adding the iommu changes to this patch so that its easy to understand and for completeness.

Also naming it as context bus is bit generic, I think the whole purpose of this is bus is to add this as another supported bus in the iommu.

may be iommu_context bus or something on those lines would be much clear.

Quoted 34 lines
 4 files changed, 43 insertions(+)
drivers/base/Kconfig +3 -0
index f7d385cbd3ba..479bc4bb442b 100644
@@ -4,6 +4,9 @@ menu "Generic Driver Options"
 config AUXILIARY_BUS
 	bool
 
+config CONTEXT_DEVICE_BUS
+	bool
+
 config UEVENT_HELPER
 	bool "Support for uevent helper"
 	help
drivers/base/Makefile +1 -0
index 8074a10183dc..ab9a0b2dc73b 100644
@@ -8,6 +8,7 @@ obj-y			:= component.o core.o bus.o dd.o syscore.o \
 			   topology.o container.o property.o cacheinfo.o \
 			   swnode.o faux.o
 obj-$(CONFIG_AUXILIARY_BUS) += auxiliary.o
+obj-$(CONFIG_CONTEXT_DEVICE_BUS) += context_bus.o
 obj-$(CONFIG_DEVTMPFS)	+= devtmpfs.o
 obj-y			+= power/
 obj-$(CONFIG_ISA_BUS_API)	+= isa.o
drivers/base/context_bus.c new file +0 -0
new file mode 100644
index 000000000000..6ddb6c27bf69

drivers/base/context_bus.c line 2

Quoted 2 lines
drivers/base/context_bus.c new file +2 -0
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: GPL-2.0-only
+// Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.

Some credits to the inspiring code would be nice.

Quoted 11 lines
drivers/base/context_bus.c new file +11 -0
+
+#include <linux/context_bus.h>
+#include <linux/init.h>
+
+const struct bus_type context_device_bus_type = {
+	.name = "context-device",
+};
+EXPORT_SYMBOL_GPL(context_device_bus_type);
+
+static int __init context_device_bus_init(void)
+{

drivers/base/context_bus.c line 23

Quoted 10 lines
drivers/base/context_bus.c new file +10 -0
+	int err;
+
+	err = bus_register(&context_device_bus_type);
+	if (err < 0) {
+		pr_err("context-device bus registration failed: %d\n", err);
+		return err;
+	}
+
+	return 0;
+}

So is the consumer drivers are expected to do the device management for this bus, this certainly is not correct and every consumer would be duplicating it.

Did you consider faux bus?

Quoted 23 lines
drivers/base/context_bus.c new file +1 -0
+postcore_initcall(context_device_bus_init);
include/linux/context_bus.h new file +15 -0
new file mode 100644
index 000000000000..0cd44cb5b147
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+ */
+
+#ifndef __LINUX_CONTEXT_BUS_H
+#define __LINUX_CONTEXT_BUS_H
+
+#include <linux/device.h>
+
+#ifdef CONFIG_CONTEXT_DEVICE_BUS
+extern const struct bus_type context_device_bus_type;
+#endif
+
+#endif /* __LINUX_CONTEXT_BUS_H */
#6 lore raw

Re: [PATCH 1/3] drivers: base: Add generic context device bus

In reply to Ekansh Gupta
Quoted 50 lines · Ekansh Gupta

Introduce a new generic bus type for synthetic context bank devices that require IOMMU context isolation. This bus provides a shared infrastructure for accelerator and GPU drivers that create virtual devices representing IOMMU context banks.

Currently, drivers like host1x implement their own bus types for context devices. This generic implementation allows multiple drivers to share the same bus infrastructure, simplifying the IOMMU subsystem integration and reducing code duplication.

---
  drivers/base/Kconfig        |  3 +++
  drivers/base/Makefile       |  1 +
  drivers/base/context_bus.c  | 24 ++++++++++++++++++++++++
  include/linux/context_bus.h | 15 +++++++++++++++
  4 files changed, 43 insertions(+)
drivers/base/Kconfig +3 -0
index f7d385cbd3ba..479bc4bb442b 100644
@@ -4,6 +4,9 @@ menu "Generic Driver Options"
  config AUXILIARY_BUS
  	bool
 
+config CONTEXT_DEVICE_BUS
+	bool
+
  config UEVENT_HELPER
  	bool "Support for uevent helper"
  	help
drivers/base/Makefile +1 -0
index 8074a10183dc..ab9a0b2dc73b 100644
@@ -8,6 +8,7 @@ obj-y			:= component.o core.o bus.o dd.o syscore.o \
  			   topology.o container.o property.o cacheinfo.o \
  			   swnode.o faux.o
  obj-$(CONFIG_AUXILIARY_BUS) += auxiliary.o
+obj-$(CONFIG_CONTEXT_DEVICE_BUS) += context_bus.o
  obj-$(CONFIG_DEVTMPFS)	+= devtmpfs.o
  obj-y			+= power/
  obj-$(CONFIG_ISA_BUS_API)	+= isa.o
drivers/base/context_bus.c new file +0 -0
new file mode 100644
index 000000000000..6ddb6c27bf69

drivers/base/context_bus.c line 8

Quoted 8 lines
drivers/base/context_bus.c new file +8 -0
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: GPL-2.0-only
+// Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+
+#include <linux/context_bus.h>
+#include <linux/init.h>
+
+const struct bus_type context_device_bus_type = {
+	.name = "context-device",

add the .dma_configure callback instead of calling from the client driver.

Quoted 38 lines
drivers/base/context_bus.c new file +16 -0
+};
+EXPORT_SYMBOL_GPL(context_device_bus_type);
+
+static int __init context_device_bus_init(void)
+{
+	int err;
+
+	err = bus_register(&context_device_bus_type);
+	if (err < 0) {
+		pr_err("context-device bus registration failed: %d\n", err);
+		return err;
+	}
+
+	return 0;
+}
+postcore_initcall(context_device_bus_init);
include/linux/context_bus.h new file +15 -0
new file mode 100644
index 000000000000..0cd44cb5b147
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+ */
+
+#ifndef __LINUX_CONTEXT_BUS_H
+#define __LINUX_CONTEXT_BUS_H
+
+#include <linux/device.h>
+
+#ifdef CONFIG_CONTEXT_DEVICE_BUS
+extern const struct bus_type context_device_bus_type;
+#endif
+
+#endif /* __LINUX_CONTEXT_BUS_H */
author reviewer
#7 lore raw

PATCH 2/3 gpu: host1x: Migrate to generic context device bus

In reply to Ekansh Gupta

Replace the host1x-specific context device bus with the generic context_device_bus_type. This eliminates driver-specific bus infrastructure in favor of shared code in the driver core.

The host1x driver creates synthetic context bank devices to represent IOMMU contexts for memory isolation. These devices were previously registered on a host1x-specific bus ("host1x-context"), but this functionality is now provided by the generic "context-device" bus.

The IOMMU subsystem is updated to recognize the generic bus instead of the host1x-specific one, allowing proper IOMMU operations on context devices.

This change maintains functional equivalence - context devices still work the same way, just on a different bus. The device names remain "host1x-ctx.N" to preserve any userspace dependencies.

---
 drivers/gpu/host1x/Kconfig       |  5 +----
 drivers/gpu/host1x/Makefile      |  1 -
 drivers/gpu/host1x/context.c     |  2 +-
 drivers/gpu/host1x/context.h     |  3 +--
 drivers/gpu/host1x/context_bus.c | 26 --------------------------
 drivers/iommu/iommu.c            |  6 +++---
 6 files changed, 6 insertions(+), 37 deletions(-)
drivers/gpu/host1x/Kconfig +1 -4
index e6c78ae2003a..e3e3896f4d71 100644
@@ -1,13 +1,10 @@
 # SPDX-License-Identifier: GPL-2.0-only
 
-config TEGRA_HOST1X_CONTEXT_BUS
-	bool
-
 config TEGRA_HOST1X
 	tristate "NVIDIA Tegra host1x driver"
 	depends on ARCH_TEGRA || COMPILE_TEST
 	select DMA_SHARED_BUFFER
-	select TEGRA_HOST1X_CONTEXT_BUS
+	select CONTEXT_DEVICE_BUS
 	select IOMMU_IOVA
 	help
 	  Driver for the NVIDIA Tegra host1x hardware.
drivers/gpu/host1x/Makefile +0 -1
index fead483af0b4..2ccd9a5f1c65 100644
@@ -23,4 +23,3 @@ host1x-$(CONFIG_IOMMU_API) += \
 	context.o
 
 obj-$(CONFIG_TEGRA_HOST1X) += host1x.o
-obj-$(CONFIG_TEGRA_HOST1X_CONTEXT_BUS) += context_bus.o
drivers/gpu/host1x/context.c +1 -1
index d50d41c20561..6411c17cc060 100644
@@ -54,7 +54,7 @@ int host1x_memory_context_list_init(struct host1x *host1x)
 		ctx->dev.dma_mask = &ctx->dma_mask;
 		ctx->dev.coherent_dma_mask = ctx->dma_mask;
 		dev_set_name(&ctx->dev, "host1x-ctx.%d", i);
-		ctx->dev.bus = &host1x_context_device_bus_type;
+		ctx->dev.bus = &context_device_bus_type;
 		ctx->dev.parent = host1x->dev;
 		ctx->dev.release = host1x_memory_context_release;
 
drivers/gpu/host1x/context.h +1 -2
index 3e03bc1d3bac..87ae522fafc7 100644
@@ -8,13 +8,12 @@
 #ifndef __HOST1X_CONTEXT_H
 #define __HOST1X_CONTEXT_H
 
+#include <linux/context_bus.h>
 #include <linux/mutex.h>
 #include <linux/refcount.h>
 
 struct host1x;
 
-extern struct bus_type host1x_context_device_bus_type;
-
 struct host1x_memory_context_list {
 	struct mutex lock;
 	struct host1x_memory_context *devs;
drivers/gpu/host1x/context_bus.c deleted +0 -26
deleted file mode 100644
index 7cd0e1a5edd1..000000000000
@@ -1,26 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0-only
-/*
- * Copyright (c) 2021, NVIDIA Corporation.
- */
-
-#include <linux/device.h>
-#include <linux/of.h>
-
-const struct bus_type host1x_context_device_bus_type = {
-	.name = "host1x-context",
-};
-EXPORT_SYMBOL_GPL(host1x_context_device_bus_type);
-
-static int __init host1x_context_device_bus_init(void)
-{
-	int err;
-
-	err = bus_register(&host1x_context_device_bus_type);
-	if (err < 0) {
-		pr_err("bus type registration failed: %d\n", err);
-		return err;
-	}
-
-	return 0;
-}
-postcore_initcall(host1x_context_device_bus_init);
drivers/iommu/iommu.c +3 -3
index 61c12ba78206..f01a13e2e634 100644
@@ -16,7 +16,7 @@
 #include <linux/export.h>
 #include <linux/slab.h>
 #include <linux/errno.h>
-#include <linux/host1x_context_bus.h>
+#include <linux/context_bus.h>
 #include <linux/iommu.h>
 #include <linux/iommufd.h>
 #include <linux/idr.h>
@@ -173,8 +173,8 @@ static const struct bus_type * const iommu_buses[] = {
 #ifdef CONFIG_FSL_MC_BUS
 	&fsl_mc_bus_type,
 #endif
-#ifdef CONFIG_TEGRA_HOST1X_CONTEXT_BUS
-	&host1x_context_device_bus_type,
+#ifdef CONFIG_CONTEXT_DEVICE_BUS
+	&context_device_bus_type,
 #endif
 #ifdef CONFIG_CDX_BUS
 	&cdx_bus_type,
Signature · 1 line
2.34.1
author reviewer
#8 lore raw

PATCH 3/3 misc: fastrpc: Use context device bus for compute banks

In reply to Ekansh Gupta

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
 drivers/misc/Kconfig   |   1 +
 drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
 2 files changed, 125 insertions(+), 56 deletions(-)
drivers/misc/Kconfig +1 -0
index 00683bf06258..b501462a4548 100644
@@ -304,6 +304,7 @@ config QCOM_FASTRPC
 	depends on RPMSG
 	select DMA_SHARED_BUFFER
 	select QCOM_SCM
+	select CONTEXT_DEVICE_BUS
 	help
 	  Provides a communication mechanism that allows for clients to
 	  make remote method invocations across processor boundary to
drivers/misc/fastrpc.c +124 -56

268 lines not on this page · raw message

Signature · 1 line
2.34.1
#9 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

Quoted 139 lines · On Tue, Apr 14, 2026 at 10:01:17PM +0530, Ekansh Gupta via B4 Relay wrote:

From: Ekansh Gupta <[email protected]>

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
 drivers/misc/Kconfig   |   1 +
 drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
 2 files changed, 125 insertions(+), 56 deletions(-)
drivers/misc/Kconfig +1 -0
index 00683bf06258..b501462a4548 100644
@@ -304,6 +304,7 @@ config QCOM_FASTRPC
 	depends on RPMSG
 	select DMA_SHARED_BUFFER
 	select QCOM_SCM
+	select CONTEXT_DEVICE_BUS
 	help
 	  Provides a communication mechanism that allows for clients to
 	  make remote method invocations across processor boundary to
drivers/misc/fastrpc.c +41 -20
index 1080f9acf70a..f66fd3eea5fa 100644
@@ -13,9 +13,9 @@
 #include <linux/module.h>
 #include <linux/of_address.h>
 #include <linux/of.h>
-#include <linux/platform_device.h>
+#include <linux/of_device.h>
 #include <linux/sort.h>
-#include <linux/of_platform.h>
+#include <linux/context_bus.h>
 #include <linux/rpmsg.h>
 #include <linux/scatterlist.h>
 #include <linux/slab.h>
@@ -250,6 +250,18 @@ struct fastrpc_invoke_ctx {
 	struct fastrpc_channel_ctx *cctx;
 };
 
+/**
+ * struct fastrpc_cb_device - Compute bank device wrapper
+ * @dev: Device structure
+ * @sess: Back-pointer to the session context
+ */
+struct fastrpc_cb_device {
+	struct device dev;
+	struct fastrpc_session_ctx *sess;
+};
+
+#define to_fastrpc_cb_device(d) container_of(d, struct fastrpc_cb_device, dev)
+
 struct fastrpc_session_ctx {
 	struct device *dev;
 	int sid;
@@ -2190,92 +2202,156 @@ static const struct file_operations fastrpc_fops = {
 	.compat_ioctl = fastrpc_device_ioctl,
 };
 
-static int fastrpc_cb_probe(struct platform_device *pdev)
+static void fastrpc_cb_dev_release(struct device *dev)
+{
+	struct fastrpc_cb_device *cb_dev = to_fastrpc_cb_device(dev);
+
+	of_node_put(dev->of_node);
+	kfree(cb_dev);
+}
+
+static int fastrpc_create_cb_device(struct fastrpc_channel_ctx *cctx,
+				    struct device *parent,
+				    struct device_node *cb_node)
 {
-	struct fastrpc_channel_ctx *cctx;
 	struct fastrpc_session_ctx *sess;
-	struct device *dev = &pdev->dev;
-	int i, sessions = 0;
+	struct fastrpc_cb_device *cb_dev;
 	unsigned long flags;
-	int rc;
-	u32 dma_bits;
-
-	cctx = dev_get_drvdata(dev->parent);
-	if (!cctx)
-		return -EINVAL;
+	int i, sessions = 0, rc;
+	u32 dma_bits, sid = 0;
 
-	of_property_read_u32(dev->of_node, "qcom,nsessions", &sessions);
+	/* Read SID early so it can be used in the device name */
+	of_property_read_u32(cb_node, "reg", &sid);
+	of_property_read_u32(cb_node, "qcom,nsessions", &sessions);
 
 	spin_lock_irqsave(&cctx->lock, flags);
 	if (cctx->sesscount >= FASTRPC_MAX_SESSIONS) {
-		dev_err(&pdev->dev, "too many sessions\n");
+		dev_err(parent, "too many sessions\n");
 		spin_unlock_irqrestore(&cctx->lock, flags);
 		return -ENOSPC;
 	}
 	dma_bits = cctx->soc_data->dma_addr_bits_default;
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
 	sess = &cctx->session[cctx->sesscount++];
 	sess->used = false;
 	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
 
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
 
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
 

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2256

Quoted 10 lines
drivers/misc/fastrpc.c +8 -2
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
+	cb_dev->dev.release = fastrpc_cb_dev_release;
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);

Please extract this code to a helper function inside the context bus and use it for both host1x and fastrpc.

Quoted 173 lines
drivers/misc/fastrpc.c +75 -34
+	rc = device_add(&cb_dev->dev);
+	if (rc) {
+		dev_err(parent, "failed to add CB device: %d\n", rc);
+		goto err_put;
+	}
+
+	rc = of_dma_configure(&cb_dev->dev, cb_node, true);
+	if (rc) {
+		dev_err(parent, "of_dma_configure failed for CB device: %d\n", rc);
+		goto err_del;
+	}
+
+	rc = dma_set_mask(&cb_dev->dev, DMA_BIT_MASK(dma_bits));
+	if (rc) {
+		dev_err(parent, "%u-bit DMA enable failed\n", dma_bits);
+		goto err_del;
+	}
+
+	sess->dev = &cb_dev->dev;
+
+	if (sessions > 0) {
+		spin_lock_irqsave(&cctx->lock, flags);
 		for (i = 1; i < sessions; i++) {
+			struct fastrpc_session_ctx *dup_sess;
+
 			if (cctx->sesscount >= FASTRPC_MAX_SESSIONS)
 				break;
 			dup_sess = &cctx->session[cctx->sesscount++];
 			memcpy(dup_sess, sess, sizeof(*dup_sess));
 		}
-	}
-	spin_unlock_irqrestore(&cctx->lock, flags);
-	rc = dma_set_mask(dev, DMA_BIT_MASK(dma_bits));
-	if (rc) {
-		dev_err(dev, "%u-bit DMA enable failed\n", dma_bits);
-		return rc;
+		spin_unlock_irqrestore(&cctx->lock, flags);
 	}
 
 	return 0;
+
+err_del:
+	device_del(&cb_dev->dev);
+err_put:
+	of_node_put(cb_dev->dev.of_node);
+	put_device(&cb_dev->dev);
+	return rc;
 }
 
-static void fastrpc_cb_remove(struct platform_device *pdev)
+static void fastrpc_depopulate_cb_devices(struct fastrpc_channel_ctx *cctx)
 {
-	struct fastrpc_channel_ctx *cctx = dev_get_drvdata(pdev->dev.parent);
-	struct fastrpc_session_ctx *sess = dev_get_drvdata(&pdev->dev);
 	unsigned long flags;
-	int i;
+	int i, j;
 
 	spin_lock_irqsave(&cctx->lock, flags);
 	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
-		if (cctx->session[i].sid == sess->sid) {
+		if (cctx->session[i].valid) {
 			cctx->session[i].valid = false;
 			cctx->sesscount--;
 		}
 	}
 	spin_unlock_irqrestore(&cctx->lock, flags);
+
+	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
+		struct device *dev = cctx->session[i].dev;
+
+		if (!dev)
+			continue;
+
+		/* Unregister the device once */
+		device_unregister(dev);
+
+		/* Clear this dev pointer from all sessions that share it */
+		for (j = i; j < FASTRPC_MAX_SESSIONS; j++) {
+			if (cctx->session[j].dev == dev)
+				cctx->session[j].dev = NULL;
+		}
+	}
 }
 
-static const struct of_device_id fastrpc_match_table[] = {
-	{ .compatible = "qcom,fastrpc-compute-cb", },
-	{}
-};
+static int fastrpc_populate_cb_devices(struct fastrpc_channel_ctx *cctx,
+					struct device *parent,
+					struct device_node *parent_node)
+{
+	struct device_node *child;
+	int ret = 0;
 
-static struct platform_driver fastrpc_cb_driver = {
-	.probe = fastrpc_cb_probe,
-	.remove = fastrpc_cb_remove,
-	.driver = {
-		.name = "qcom,fastrpc-cb",
-		.of_match_table = fastrpc_match_table,
-		.suppress_bind_attrs = true,
-	},
-};
+	for_each_child_of_node(parent_node, child) {
+		if (!of_device_is_compatible(child, "qcom,fastrpc-compute-cb"))
+			continue;
+
+		ret = fastrpc_create_cb_device(cctx, parent, child);
+		if (ret) {
+			dev_err(parent, "failed to create CB device for %s: %d\n",
+				child->name, ret);
+			of_node_put(child);
+			fastrpc_depopulate_cb_devices(cctx);
+			return ret;
+		}
+	}
+
+	return 0;
+}
 
 static int fastrpc_device_register(struct device *dev, struct fastrpc_channel_ctx *cctx,
 				   bool is_secured, const char *domain)
@@ -2441,7 +2517,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
 	data->domain_id = domain_id;
 	data->rpdev = rpdev;
 
-	err = of_platform_populate(rdev->of_node, NULL, NULL, rdev);
+	err = fastrpc_populate_cb_devices(data, rdev, rdev->of_node);
 	if (err)
 		goto err_deregister_fdev;
 
@@ -2496,7 +2572,7 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
 	if (cctx->remote_heap)
 		fastrpc_buf_free(cctx->remote_heap);
 
-	of_platform_depopulate(&rpdev->dev);
+	fastrpc_depopulate_cb_devices(cctx);
 
 	fastrpc_channel_ctx_put(cctx);
 }
@@ -2558,16 +2634,9 @@ static int fastrpc_init(void)
 {
 	int ret;
 
-	ret = platform_driver_register(&fastrpc_cb_driver);
-	if (ret < 0) {
-		pr_err("fastrpc: failed to register cb driver\n");
-		return ret;
-	}
-
 	ret = register_rpmsg_driver(&fastrpc_driver);
 	if (ret < 0) {
 		pr_err("fastrpc: failed to register rpmsg driver\n");
-		platform_driver_unregister(&fastrpc_cb_driver);
 		return ret;
 	}
 
@@ -2577,7 +2646,6 @@ module_init(fastrpc_init);
 
 static void fastrpc_exit(void)
 {
-	platform_driver_unregister(&fastrpc_cb_driver);
 	unregister_rpmsg_driver(&fastrpc_driver);
 }
 module_exit(fastrpc_exit);
Signature · 2 lines
-- 
2.34.1
Signature · 2 lines
With best wishes
Dmitry
#10 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

In reply to Ekansh Gupta
Quoted 57 lines · On 4/14/26 4:31 PM, Ekansh Gupta via B4 Relay wrote:

From: Ekansh Gupta <[email protected]>

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
 drivers/misc/Kconfig   |   1 +
 drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
 2 files changed, 125 insertions(+), 56 deletions(-)
drivers/misc/Kconfig +1 -0
index 00683bf06258..b501462a4548 100644
@@ -304,6 +304,7 @@ config QCOM_FASTRPC
 	depends on RPMSG
 	select DMA_SHARED_BUFFER
 	select QCOM_SCM
+	select CONTEXT_DEVICE_BUS
 	help
 	  Provides a communication mechanism that allows for clients to
 	  make remote method invocations across processor boundary to
drivers/misc/fastrpc.c +2 -2
index 1080f9acf70a..f66fd3eea5fa 100644
@@ -13,9 +13,9 @@
 #include <linux/module.h>
 #include <linux/of_address.h>
 #include <linux/of.h>
-#include <linux/platform_device.h>
+#include <linux/of_device.h>
 #include <linux/sort.h>
-#include <linux/of_platform.h>
+#include <linux/context_bus.h>
 #include <linux/rpmsg.h>
 #include <linux/scatterlist.h>
 #include <linux/slab.h>
@@ -250,6 +250,18 @@ struct fastrpc_invoke_ctx {
 	struct fastrpc_channel_ctx *cctx;
 };

drivers/misc/fastrpc.c struct fastrpc_invoke_ctx { · line 261

Quoted 10 lines
drivers/misc/fastrpc.c +9 -0
 
+/**
+ * struct fastrpc_cb_device - Compute bank device wrapper
+ * @dev: Device structure
+ * @sess: Back-pointer to the session context
+ */
+struct fastrpc_cb_device {
+	struct device dev;
+	struct fastrpc_session_ctx *sess;
+};

As part of this migration are we saying that driver and device model should be done in the consumer driver?

May be the right question is why can not we add other bus types into iommu supported buses.

There has been some discussion on this in the past
https://www.mail-archive.com/[email protected]/msg22139.html

This is one of the restrictions that is forcing the fake the context devices as platform devices.

for me this looks like half baked solution and very much specific to deal with the iommu driver situation

--srini

Quoted 251 lines
drivers/misc/fastrpc.c +113 -54
+
+#define to_fastrpc_cb_device(d) container_of(d, struct fastrpc_cb_device, dev)
+
 struct fastrpc_session_ctx {
 	struct device *dev;
 	int sid;
@@ -2190,92 +2202,156 @@ static const struct file_operations fastrpc_fops = {
 	.compat_ioctl = fastrpc_device_ioctl,
 };
 
-static int fastrpc_cb_probe(struct platform_device *pdev)
+static void fastrpc_cb_dev_release(struct device *dev)
+{
+	struct fastrpc_cb_device *cb_dev = to_fastrpc_cb_device(dev);
+
+	of_node_put(dev->of_node);
+	kfree(cb_dev);
+}
+
+static int fastrpc_create_cb_device(struct fastrpc_channel_ctx *cctx,
+				    struct device *parent,
+				    struct device_node *cb_node)
 {
-	struct fastrpc_channel_ctx *cctx;
 	struct fastrpc_session_ctx *sess;
-	struct device *dev = &pdev->dev;
-	int i, sessions = 0;
+	struct fastrpc_cb_device *cb_dev;
 	unsigned long flags;
-	int rc;
-	u32 dma_bits;
-
-	cctx = dev_get_drvdata(dev->parent);
-	if (!cctx)
-		return -EINVAL;
+	int i, sessions = 0, rc;
+	u32 dma_bits, sid = 0;
 
-	of_property_read_u32(dev->of_node, "qcom,nsessions", &sessions);
+	/* Read SID early so it can be used in the device name */
+	of_property_read_u32(cb_node, "reg", &sid);
+	of_property_read_u32(cb_node, "qcom,nsessions", &sessions);
 
 	spin_lock_irqsave(&cctx->lock, flags);
 	if (cctx->sesscount >= FASTRPC_MAX_SESSIONS) {
-		dev_err(&pdev->dev, "too many sessions\n");
+		dev_err(parent, "too many sessions\n");
 		spin_unlock_irqrestore(&cctx->lock, flags);
 		return -ENOSPC;
 	}
 	dma_bits = cctx->soc_data->dma_addr_bits_default;
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
 	sess = &cctx->session[cctx->sesscount++];
 	sess->used = false;
 	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
 
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
 
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
 
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
+	cb_dev->dev.release = fastrpc_cb_dev_release;
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);
 
+	rc = device_add(&cb_dev->dev);
+	if (rc) {
+		dev_err(parent, "failed to add CB device: %d\n", rc);
+		goto err_put;
+	}
+
+	rc = of_dma_configure(&cb_dev->dev, cb_node, true);
+	if (rc) {
+		dev_err(parent, "of_dma_configure failed for CB device: %d\n", rc);
+		goto err_del;
+	}
+
+	rc = dma_set_mask(&cb_dev->dev, DMA_BIT_MASK(dma_bits));
+	if (rc) {
+		dev_err(parent, "%u-bit DMA enable failed\n", dma_bits);
+		goto err_del;
+	}
+
+	sess->dev = &cb_dev->dev;
+
+	if (sessions > 0) {
+		spin_lock_irqsave(&cctx->lock, flags);
 		for (i = 1; i < sessions; i++) {
+			struct fastrpc_session_ctx *dup_sess;
+
 			if (cctx->sesscount >= FASTRPC_MAX_SESSIONS)
 				break;
 			dup_sess = &cctx->session[cctx->sesscount++];
 			memcpy(dup_sess, sess, sizeof(*dup_sess));
 		}
-	}
-	spin_unlock_irqrestore(&cctx->lock, flags);
-	rc = dma_set_mask(dev, DMA_BIT_MASK(dma_bits));
-	if (rc) {
-		dev_err(dev, "%u-bit DMA enable failed\n", dma_bits);
-		return rc;
+		spin_unlock_irqrestore(&cctx->lock, flags);
 	}
 
 	return 0;
+
+err_del:
+	device_del(&cb_dev->dev);
+err_put:
+	of_node_put(cb_dev->dev.of_node);
+	put_device(&cb_dev->dev);
+	return rc;
 }
 
-static void fastrpc_cb_remove(struct platform_device *pdev)
+static void fastrpc_depopulate_cb_devices(struct fastrpc_channel_ctx *cctx)
 {
-	struct fastrpc_channel_ctx *cctx = dev_get_drvdata(pdev->dev.parent);
-	struct fastrpc_session_ctx *sess = dev_get_drvdata(&pdev->dev);
 	unsigned long flags;
-	int i;
+	int i, j;
 
 	spin_lock_irqsave(&cctx->lock, flags);
 	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
-		if (cctx->session[i].sid == sess->sid) {
+		if (cctx->session[i].valid) {
 			cctx->session[i].valid = false;
 			cctx->sesscount--;
 		}
 	}
 	spin_unlock_irqrestore(&cctx->lock, flags);
+
+	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
+		struct device *dev = cctx->session[i].dev;
+
+		if (!dev)
+			continue;
+
+		/* Unregister the device once */
+		device_unregister(dev);
+
+		/* Clear this dev pointer from all sessions that share it */
+		for (j = i; j < FASTRPC_MAX_SESSIONS; j++) {
+			if (cctx->session[j].dev == dev)
+				cctx->session[j].dev = NULL;
+		}
+	}
 }
 
-static const struct of_device_id fastrpc_match_table[] = {
-	{ .compatible = "qcom,fastrpc-compute-cb", },
-	{}
-};
+static int fastrpc_populate_cb_devices(struct fastrpc_channel_ctx *cctx,
+					struct device *parent,
+					struct device_node *parent_node)
+{
+	struct device_node *child;
+	int ret = 0;
 
-static struct platform_driver fastrpc_cb_driver = {
-	.probe = fastrpc_cb_probe,
-	.remove = fastrpc_cb_remove,
-	.driver = {
-		.name = "qcom,fastrpc-cb",
-		.of_match_table = fastrpc_match_table,
-		.suppress_bind_attrs = true,
-	},
-};
+	for_each_child_of_node(parent_node, child) {
+		if (!of_device_is_compatible(child, "qcom,fastrpc-compute-cb"))
+			continue;
+
+		ret = fastrpc_create_cb_device(cctx, parent, child);
+		if (ret) {
+			dev_err(parent, "failed to create CB device for %s: %d\n",
+				child->name, ret);
+			of_node_put(child);
+			fastrpc_depopulate_cb_devices(cctx);
+			return ret;
+		}
+	}
+
+	return 0;
+}
 
 static int fastrpc_device_register(struct device *dev, struct fastrpc_channel_ctx *cctx,
 				   bool is_secured, const char *domain)
@@ -2441,7 +2517,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
 	data->domain_id = domain_id;
 	data->rpdev = rpdev;
 
-	err = of_platform_populate(rdev->of_node, NULL, NULL, rdev);
+	err = fastrpc_populate_cb_devices(data, rdev, rdev->of_node);
 	if (err)
 		goto err_deregister_fdev;
 
@@ -2496,7 +2572,7 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
 	if (cctx->remote_heap)
 		fastrpc_buf_free(cctx->remote_heap);
 
-	of_platform_depopulate(&rpdev->dev);
+	fastrpc_depopulate_cb_devices(cctx);
 
 	fastrpc_channel_ctx_put(cctx);
 }
@@ -2558,16 +2634,9 @@ static int fastrpc_init(void)
 {
 	int ret;
 
-	ret = platform_driver_register(&fastrpc_cb_driver);
-	if (ret < 0) {
-		pr_err("fastrpc: failed to register cb driver\n");
-		return ret;
-	}
-
 	ret = register_rpmsg_driver(&fastrpc_driver);
 	if (ret < 0) {
 		pr_err("fastrpc: failed to register rpmsg driver\n");
-		platform_driver_unregister(&fastrpc_cb_driver);
 		return ret;
 	}
 
@@ -2577,7 +2646,6 @@ module_init(fastrpc_init);
 
 static void fastrpc_exit(void)
 {
-	platform_driver_unregister(&fastrpc_cb_driver);
 	unregister_rpmsg_driver(&fastrpc_driver);
 }
 module_exit(fastrpc_exit);
#11 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

In reply to Ekansh Gupta
Quoted 139 lines · Ekansh Gupta

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
  drivers/misc/Kconfig   |   1 +
  drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
  2 files changed, 125 insertions(+), 56 deletions(-)
drivers/misc/Kconfig +1 -0
index 00683bf06258..b501462a4548 100644
@@ -304,6 +304,7 @@ config QCOM_FASTRPC
  	depends on RPMSG
  	select DMA_SHARED_BUFFER
  	select QCOM_SCM
+	select CONTEXT_DEVICE_BUS
  	help
  	  Provides a communication mechanism that allows for clients to
  	  make remote method invocations across processor boundary to
drivers/misc/fastrpc.c +41 -22
index 1080f9acf70a..f66fd3eea5fa 100644
@@ -13,9 +13,9 @@
  #include <linux/module.h>
  #include <linux/of_address.h>
  #include <linux/of.h>
-#include <linux/platform_device.h>
+#include <linux/of_device.h>
  #include <linux/sort.h>
-#include <linux/of_platform.h>
+#include <linux/context_bus.h>
  #include <linux/rpmsg.h>
  #include <linux/scatterlist.h>
  #include <linux/slab.h>
@@ -250,6 +250,18 @@ struct fastrpc_invoke_ctx {
  	struct fastrpc_channel_ctx *cctx;
  };
 
+/**
+ * struct fastrpc_cb_device - Compute bank device wrapper
+ * @dev: Device structure
+ * @sess: Back-pointer to the session context
+ */
+struct fastrpc_cb_device {
+	struct device dev;
+	struct fastrpc_session_ctx *sess;
+};
+
+#define to_fastrpc_cb_device(d) container_of(d, struct fastrpc_cb_device, dev)
+
  struct fastrpc_session_ctx {
  	struct device *dev;
  	int sid;
@@ -2190,92 +2202,156 @@ static const struct file_operations fastrpc_fops = {
  	.compat_ioctl = fastrpc_device_ioctl,
  };
 
-static int fastrpc_cb_probe(struct platform_device *pdev)
+static void fastrpc_cb_dev_release(struct device *dev)
+{
+	struct fastrpc_cb_device *cb_dev = to_fastrpc_cb_device(dev);
+
+	of_node_put(dev->of_node);
+	kfree(cb_dev);
+}
+
+static int fastrpc_create_cb_device(struct fastrpc_channel_ctx *cctx,
+				    struct device *parent,
+				    struct device_node *cb_node)
  {
-	struct fastrpc_channel_ctx *cctx;
  	struct fastrpc_session_ctx *sess;
-	struct device *dev = &pdev->dev;
-	int i, sessions = 0;
+	struct fastrpc_cb_device *cb_dev;
  	unsigned long flags;
-	int rc;
-	u32 dma_bits;
-
-	cctx = dev_get_drvdata(dev->parent);
-	if (!cctx)
-		return -EINVAL;
+	int i, sessions = 0, rc;
+	u32 dma_bits, sid = 0;
 
-	of_property_read_u32(dev->of_node, "qcom,nsessions", &sessions);
+	/* Read SID early so it can be used in the device name */
+	of_property_read_u32(cb_node, "reg", &sid);
+	of_property_read_u32(cb_node, "qcom,nsessions", &sessions);
 
  	spin_lock_irqsave(&cctx->lock, flags);
  	if (cctx->sesscount >= FASTRPC_MAX_SESSIONS) {
-		dev_err(&pdev->dev, "too many sessions\n");
+		dev_err(parent, "too many sessions\n");
  		spin_unlock_irqrestore(&cctx->lock, flags);
  		return -ENOSPC;
  	}
  	dma_bits = cctx->soc_data->dma_addr_bits_default;
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
  	sess = &cctx->session[cctx->sesscount++];
  	sess->used = false;
  	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
 
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
 
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
 
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2258

Quoted 10 lines
drivers/misc/fastrpc.c +9 -0
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
+	cb_dev->dev.release = fastrpc_cb_dev_release;
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);
 
+	rc = device_add(&cb_dev->dev);

device_initialize() and device_add() can be replaced with single device_register() call. You can refer the below patch,

https://lore.kernel.org/all/[email protected]/

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2264

Quoted 6 lines
drivers/misc/fastrpc.c +6 -0
+	if (rc) {
+		dev_err(parent, "failed to add CB device: %d\n", rc);
+		goto err_put;
+	}
+
+	rc = of_dma_configure(&cb_dev->dev, cb_node, true);

Specific dma configuration should be done via .dma_configure callback, as suggested by Robin earlier in the discussion,

https://lore.kernel.org/all/[email protected]/

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2270

Quoted 6 lines
drivers/misc/fastrpc.c +6 -0
+	if (rc) {
+		dev_err(parent, "of_dma_configure failed for CB device: %d\n", rc);
+		goto err_del;
+	}
+
+	rc = dma_set_mask(&cb_dev->dev, DMA_BIT_MASK(dma_bits));

Can this be set during device configuration and avoid explict dma_set_mask call?

Quoted 155 lines
drivers/misc/fastrpc.c +62 -34
+	if (rc) {
+		dev_err(parent, "%u-bit DMA enable failed\n", dma_bits);
+		goto err_del;
+	}
+
+	sess->dev = &cb_dev->dev;
+
+	if (sessions > 0) {
+		spin_lock_irqsave(&cctx->lock, flags);
  		for (i = 1; i < sessions; i++) {
+			struct fastrpc_session_ctx *dup_sess;
+
  			if (cctx->sesscount >= FASTRPC_MAX_SESSIONS)
  				break;
  			dup_sess = &cctx->session[cctx->sesscount++];
  			memcpy(dup_sess, sess, sizeof(*dup_sess));
  		}
-	}
-	spin_unlock_irqrestore(&cctx->lock, flags);
-	rc = dma_set_mask(dev, DMA_BIT_MASK(dma_bits));
-	if (rc) {
-		dev_err(dev, "%u-bit DMA enable failed\n", dma_bits);
-		return rc;
+		spin_unlock_irqrestore(&cctx->lock, flags);
  	}
 
  	return 0;
+
+err_del:
+	device_del(&cb_dev->dev);
+err_put:
+	of_node_put(cb_dev->dev.of_node);
+	put_device(&cb_dev->dev);
+	return rc;
  }
 
-static void fastrpc_cb_remove(struct platform_device *pdev)
+static void fastrpc_depopulate_cb_devices(struct fastrpc_channel_ctx *cctx)
  {
-	struct fastrpc_channel_ctx *cctx = dev_get_drvdata(pdev->dev.parent);
-	struct fastrpc_session_ctx *sess = dev_get_drvdata(&pdev->dev);
  	unsigned long flags;
-	int i;
+	int i, j;
 
  	spin_lock_irqsave(&cctx->lock, flags);
  	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
-		if (cctx->session[i].sid == sess->sid) {
+		if (cctx->session[i].valid) {
  			cctx->session[i].valid = false;
  			cctx->sesscount--;
  		}
  	}
  	spin_unlock_irqrestore(&cctx->lock, flags);
+
+	for (i = 0; i < FASTRPC_MAX_SESSIONS; i++) {
+		struct device *dev = cctx->session[i].dev;
+
+		if (!dev)
+			continue;
+
+		/* Unregister the device once */
+		device_unregister(dev);
+
+		/* Clear this dev pointer from all sessions that share it */
+		for (j = i; j < FASTRPC_MAX_SESSIONS; j++) {
+			if (cctx->session[j].dev == dev)
+				cctx->session[j].dev = NULL;
+		}
+	}
  }
 
-static const struct of_device_id fastrpc_match_table[] = {
-	{ .compatible = "qcom,fastrpc-compute-cb", },
-	{}
-};
+static int fastrpc_populate_cb_devices(struct fastrpc_channel_ctx *cctx,
+					struct device *parent,
+					struct device_node *parent_node)
+{
+	struct device_node *child;
+	int ret = 0;
 
-static struct platform_driver fastrpc_cb_driver = {
-	.probe = fastrpc_cb_probe,
-	.remove = fastrpc_cb_remove,
-	.driver = {
-		.name = "qcom,fastrpc-cb",
-		.of_match_table = fastrpc_match_table,
-		.suppress_bind_attrs = true,
-	},
-};
+	for_each_child_of_node(parent_node, child) {
+		if (!of_device_is_compatible(child, "qcom,fastrpc-compute-cb"))
+			continue;
+
+		ret = fastrpc_create_cb_device(cctx, parent, child);
+		if (ret) {
+			dev_err(parent, "failed to create CB device for %s: %d\n",
+				child->name, ret);
+			of_node_put(child);
+			fastrpc_depopulate_cb_devices(cctx);
+			return ret;
+		}
+	}
+
+	return 0;
+}
 
  static int fastrpc_device_register(struct device *dev, struct fastrpc_channel_ctx *cctx,
  				   bool is_secured, const char *domain)
@@ -2441,7 +2517,7 @@ static int fastrpc_rpmsg_probe(struct rpmsg_device *rpdev)
  	data->domain_id = domain_id;
  	data->rpdev = rpdev;
 
-	err = of_platform_populate(rdev->of_node, NULL, NULL, rdev);
+	err = fastrpc_populate_cb_devices(data, rdev, rdev->of_node);
  	if (err)
  		goto err_deregister_fdev;
 
@@ -2496,7 +2572,7 @@ static void fastrpc_rpmsg_remove(struct rpmsg_device *rpdev)
  	if (cctx->remote_heap)
  		fastrpc_buf_free(cctx->remote_heap);
 
-	of_platform_depopulate(&rpdev->dev);
+	fastrpc_depopulate_cb_devices(cctx);
 
  	fastrpc_channel_ctx_put(cctx);
  }
@@ -2558,16 +2634,9 @@ static int fastrpc_init(void)
  {
  	int ret;
 
-	ret = platform_driver_register(&fastrpc_cb_driver);
-	if (ret < 0) {
-		pr_err("fastrpc: failed to register cb driver\n");
-		return ret;
-	}
-
  	ret = register_rpmsg_driver(&fastrpc_driver);
  	if (ret < 0) {
  		pr_err("fastrpc: failed to register rpmsg driver\n");
-		platform_driver_unregister(&fastrpc_cb_driver);
  		return ret;
  	}
 
@@ -2577,7 +2646,6 @@ module_init(fastrpc_init);
 
  static void fastrpc_exit(void)
  {
-	platform_driver_unregister(&fastrpc_cb_driver);
  	unregister_rpmsg_driver(&fastrpc_driver);
  }
  module_exit(fastrpc_exit);
#12 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

Quoted 53 lines · Vishnu Reddy
Quoted 50 lines · Ekansh Gupta

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
  drivers/misc/Kconfig   |   1 +
  drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
  2 files changed, 125 insertions(+), 56 deletions(-)
  	}
  	dma_bits = cctx->soc_data->dma_addr_bits_default;
drivers/misc/fastrpc.c +13 -8
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
  	sess = &cctx->session[cctx->sesscount++];
  	sess->used = false;
  	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
+	cb_dev->dev.release = fastrpc_cb_dev_release;
Quoted 8 lines

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2258

Quoted 5 lines
drivers/misc/fastrpc.c +5 -0
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);
+	rc = device_add(&cb_dev->dev);

device_initialize() and device_add() can be replaced with single device_register() call. You can refer the below patch,

Keep in mind that for several arches device_initialize() sets the coherent_dma_mask. So one should set coherent_dma_mask (as it's done here) after calling device_initialize().

Quoted 3 lines

https://lore.kernel.org/all/[email protected]/

Signature · 2 lines
With best wishes
Dmitry
#13 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

Quoted 60 lines · Dmitry Baryshkov
Quoted 51 lines · Vishnu Reddy
Quoted 49 lines · Ekansh Gupta

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
  drivers/misc/Kconfig   |   1 +
  drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
  2 files changed, 125 insertions(+), 56 deletions(-)
  	}
  	dma_bits = cctx->soc_data->dma_addr_bits_default;
drivers/misc/fastrpc.c +12 -8
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
  	sess = &cctx->session[cctx->sesscount++];
  	sess->used = false;
  	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
Quoted 8 lines

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2258

Quoted 6 lines
drivers/misc/fastrpc.c +6 -0
+	cb_dev->dev.release = fastrpc_cb_dev_release;
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);
+	rc = device_add(&cb_dev->dev);

device_initialize() and device_add() can be replaced with single device_register() call. You can refer the below patch,

Quoted 3 lines

Keep in mind that for several arches device_initialize() sets the coherent_dma_mask. So one should set coherent_dma_mask (as it's done here) after calling device_initialize().

I can see that device_initialize() sets dma_coherent flag for several arches, not the coherent_dma_mask. The final dma_coherent flag value is updated later in of_dma_configure_id() based on the dt dma-coherent property.

Setting the coherent_dma_mask here and calling device_register() which will call the device_initialize() won't change the coherent_dma_mask value. Please correct me if I'm wrong?

Quoted 2 lines
Quoted 2 lines

https://lore.kernel.org/all/[email protected]/

#14 lore raw

Re: [PATCH 3/3] misc: fastrpc: Use context device bus for compute banks

Quoted 66 lines · Vishnu Reddy
Quoted 60 lines · Dmitry Baryshkov
Quoted 51 lines · Vishnu Reddy
Quoted 49 lines · Ekansh Gupta

Replace the platform driver approach for compute bank (CB) devices with the generic context_device_bus_type. Compute bank devices are synthetic IOMMU context banks, not real platform devices, so using the context device bus provides a more accurate representation in the device model.

Currently, fastrpc used of_platform_populate() to create platform devices for each "qcom,fastrpc-compute-cb" DT node, with a platform driver (fastrpc_cb_driver) to handle probe/remove. This approach had a race condition: device nodes were created before channel resources (like spin_lock) were initialized, and probe was async, so applications could open the device before sessions were available.

This patch addresses the race by manually creating and configuring CB devices synchronously during fastrpc_rpmsg_probe(), after all channel resources are initialized. The approach follows the pattern used in host1x_memory_context_list_init().

---
  drivers/misc/Kconfig   |   1 +
  drivers/misc/fastrpc.c | 180 ++++++++++++++++++++++++++++++++++---------------
  2 files changed, 125 insertions(+), 56 deletions(-)
  	}
  	dma_bits = cctx->soc_data->dma_addr_bits_default;
drivers/misc/fastrpc.c +12 -8
+	if (cctx->domain_id == CDSP_DOMAIN_ID)
+		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+
  	sess = &cctx->session[cctx->sesscount++];
  	sess->used = false;
  	sess->valid = true;
-	sess->dev = dev;
-	dev_set_drvdata(dev, sess);
+	sess->sid = sid;
+	spin_unlock_irqrestore(&cctx->lock, flags);
-	if (cctx->domain_id == CDSP_DOMAIN_ID)
-		dma_bits = cctx->soc_data->dma_addr_bits_cdsp;
+	cb_dev = kzalloc_obj(*cb_dev);
+	if (!cb_dev)
+		return -ENOMEM;
-	if (of_property_read_u32(dev->of_node, "reg", &sess->sid))
-		dev_info(dev, "FastRPC Session ID not specified in DT\n");
+	cb_dev->sess = sess;
-	if (sessions > 0) {
-		struct fastrpc_session_ctx *dup_sess;
+	device_initialize(&cb_dev->dev);
+	cb_dev->dev.parent = parent;
+	cb_dev->dev.bus = &context_device_bus_type;
Quoted 8 lines

drivers/misc/fastrpc.c static const struct file_operations fastrpc_fops = { · line 2258

Quoted 6 lines
drivers/misc/fastrpc.c +6 -0
+	cb_dev->dev.release = fastrpc_cb_dev_release;
+	cb_dev->dev.of_node = of_node_get(cb_node);
+	cb_dev->dev.dma_mask = &cb_dev->dev.coherent_dma_mask;
+	cb_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32);
+	dev_set_name(&cb_dev->dev, "%s:compute-cb@%u", dev_name(parent), sid);
+	rc = device_add(&cb_dev->dev);

device_initialize() and device_add() can be replaced with single device_register() call. You can refer the below patch,

Quoted 3 lines

Keep in mind that for several arches device_initialize() sets the coherent_dma_mask. So one should set coherent_dma_mask (as it's done here) after calling device_initialize().

Quoted 7 lines

I can see that device_initialize() sets dma_coherent flag for several arches, not the coherent_dma_mask. The final dma_coherent flag value is updated later in of_dma_configure_id() based on the dt dma-coherent property.

Setting the coherent_dma_mask here and calling device_register() which will call the device_initialize() won't change the coherent_dma_mask value. Please correct me if I'm wrong?

I should get better glasses. I agree with you here.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

https://lore.kernel.org/all/[email protected]/

Signature · 2 lines
With best wishes
Dmitry