Skip to contentSkip to conversation

gpu: host1x: Call of_dma_configure after setting bus

PATCH New Mikko Perttunen 6 participants 8 messages Sep 24, 2017 – Oct 11, 2017

Subsystems (7)
DRM drivers PCI Tegra architecture support Driver core +3 more subsystems
lore
Share & tools
lore

Mikko Perttunen

author maintainer
#0 lore raw

PATCH 1/2 gpu: host1x: Call of_dma_configure after setting bus

of_dma_configure now checks the device's bus before configuring it, so we need to set the device's bus before calling.

Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/gpu/host1x/bus.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/host1x/bus.c +2 -1
index f9cde03030fd..66ea5acee820 100644
@@ -403,12 +403,13 @@ static int host1x_device_add(struct host1x *host1x,
 	device->dev.coherent_dma_mask = host1x->dev->coherent_dma_mask;
 	device->dev.dma_mask = &device->dev.coherent_dma_mask;
 	dev_set_name(&device->dev, "%s", driver->driver.name);
-	of_dma_configure(&device->dev, host1x->dev->of_node);
 	device->dev.release = host1x_device_release;
 	device->dev.of_node = host1x->dev->of_node;
 	device->dev.bus = &host1x_bus_type;
 	device->dev.parent = host1x->dev;
 
+	of_dma_configure(&device->dev, host1x->dev->of_node);
+
 	err = host1x_device_parse_dt(device, driver);
 	if (err < 0) {
 		kfree(device);
Signature · 1 line
2.14.1
author maintainer
#1 lore raw

PATCH 2/2 of: configure DMA for host1x devices

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/of/device.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/of/device.c +3 -0
index 64b710265d39..12368418cd33 100644
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		if (!dev_is_pci(dev) &&
 #ifdef CONFIG_ARM_AMBA
 		    dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+		    !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
 #endif
 		    dev->bus != &platform_bus_type)
 			return ret == -ENODEV ? 0 : ret;
Signature · 1 line
2.14.1
#2 lore raw

PATCH 2/2 of: configure DMA for host1x devices

Quoted 17 lines · Mikko Perttunen

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/of/device.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/of/device.c +0 -0
index 64b710265d39..12368418cd33 100644

drivers/of/device.c int of_dma_configure(struct device *dev, struct device_node *np) · line 112

Quoted 9 lines
drivers/of/device.c +3 -0
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		if (!dev_is_pci(dev) &&
 #ifdef CONFIG_ARM_AMBA
 		    dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+		    !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
 #endif
 		    dev->bus != &platform_bus_type)
 			return ret == -ENODEV ? 0 : ret;

Maybe a slightly better solution would be to add a dev_is_host1x() function that returns dev->bus == &host1x_bus_type if TEGRA_HOST1X is enabled and returns false otherwise. That way we can have a "proper" check for the bus type and avoid the #if in this file.

It's slightly more complicated because of the dependencies between the DT and Tegra DRM trees, but I'm sure we can get Rob to either give his Acked-by on the drivers/of/device.c change or carry the Tegra DRM patch in the DT tree.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170925/9c58ca02/attachment.sig>

#3 lore raw

PATCH 2/2 of: configure DMA for host1x devices

Quoted 33 lines · Thierry Reding
Quoted 17 lines · Mikko Perttunen

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/of/device.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/of/device.c +0 -0
index 64b710265d39..12368418cd33 100644

drivers/of/device.c int of_dma_configure(struct device *dev, struct device_node *np) · line 112

Quoted 9 lines
drivers/of/device.c +3 -0
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		if (!dev_is_pci(dev) &&
 #ifdef CONFIG_ARM_AMBA
 		    dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+		    !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
 #endif
 		    dev->bus != &platform_bus_type)
 			return ret == -ENODEV ? 0 : ret;

Maybe a slightly better solution would be to add a dev_is_host1x() function that returns dev->bus == &host1x_bus_type if TEGRA_HOST1X is enabled and returns false otherwise. That way we can have a "proper" check for the bus type and avoid the #if in this file.

Quoted 4 lines

It's slightly more complicated because of the dependencies between the DT and Tegra DRM trees, but I'm sure we can get Rob to either give his Acked-by on the drivers/of/device.c change or carry the Tegra DRM patch in the DT tree.

Alternatively (and modulo Greg's opinion, obviously) it would arguably be neatest to handle this at a slightly lower level as below - I only really considered this idea after writing the existing fix, but now that I have actually written it up I do rather like it.

Robin.

----->8-----
From: Robin Murphy <[email protected]>
Date: Mon, 25 Sep 2017 12:45:58 +0100
Subject: [PATCH] drivers: Flag buses which demand DMA configuration

We do not want the common dma_configure() pathway to apply indiscriminately to all devices, since there are plenty of buses which do not have DMA capability, and if their child devices were used for DMA API calls it would only be indicative of a driver bug. However, there are a number of buses for which DMA is implicitly expected even when not described by firmware, so we currently whitelist those to automatically opt in to dma_configure(), assuming a 1:1 mapping between the DMA address space and the physical address space.

Commit 723288836628 ("of: restrict DMA configuration") introduced a short-term fix by comparing explicit bus types, but this approach is far from pretty, doesn't scale well, and fails to cope at all with bus drivers which may be built as modules. Let's refine things by making that opt-in a property of the bus type, which neatly addresses those problems and lets the decision of whether firmware description of DMA capability should be optional or mandatory stay internal to the bus drivers themselves.

Signed-off-by Robin Murphy <[email protected]>
---
 drivers/amba/bus.c       | 1 +
 drivers/base/platform.c  | 1 +
 drivers/gpu/host1x/bus.c | 1 +
 drivers/of/device.c      | 8 +-------
 drivers/pci/pci-driver.c | 1 +
 include/linux/device.h   | 4 ++++
 6 files changed, 9 insertions(+), 7 deletions(-)
drivers/amba/bus.c +1 -0
index e0f74ddc22b7..594c228d2f02 100644
@@ -195,6 +195,7 @@ struct bus_type amba_bustype = {
 	.match		= amba_match,
 	.uevent		= amba_uevent,
 	.pm		= &amba_pm,
+	.force_dma	= true,
 };
 
 static int __init amba_init(void)
drivers/base/platform.c +1 -0
index d1bd99271066..9c95c8db00f8 100644
@@ -1142,6 +1142,7 @@ struct bus_type platform_bus_type = {
 	.match		= platform_match,
 	.uevent		= platform_uevent,
 	.pm		= &platform_dev_pm_ops,
+	.force_dma	= true,
 };
 EXPORT_SYMBOL_GPL(platform_bus_type);
drivers/gpu/host1x/bus.c +1 -0
index f9cde03030fd..ed03b3243195 100644
@@ -320,6 +320,7 @@ struct bus_type host1x_bus_type = {
 	.name = "host1x",
 	.match = host1x_device_match,
 	.pm = &host1x_device_pm_ops,
+	.force_dma = true,
 };
 
 static void __host1x_device_del(struct host1x_device *device)
drivers/of/device.c +1 -7
index 64b710265d39..25bddf9c9fe1 100644
@@ -9,9 +9,7 @@
 #include <linux/module.h>
 #include <linux/mod_devicetable.h>
 #include <linux/slab.h>
-#include <linux/pci.h>
 #include <linux/platform_device.h>
-#include <linux/amba/bus.h>
 
 #include <asm/errno.h>
 #include "of_private.h"

@@ -101,11 +99,7 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		 * DMA configuration regardless of whether "dma-ranges" is
 		 * correctly specified or not.
 		 */
-		if (!dev_is_pci(dev) &&
-#ifdef CONFIG_ARM_AMBA
-		    dev->bus != &amba_bustype &&
-#endif
-		    dev->bus != &platform_bus_type)
+		if (!dev->bus->force_dma)
 			return ret == -ENODEV ? 0 : ret;
 
 		dma_addr = offset = 0;
drivers/pci/pci-driver.c +1 -0
index 11bd267fc137..38bdb97b6dc6 100644
@@ -1466,6 +1466,7 @@ struct bus_type pci_bus_type = {
 	.drv_groups	= pci_drv_groups,
 	.pm		= PCI_PM_OPS_PTR,
 	.num_vf		= pci_bus_num_vf,
+	.force_dma	= true,
 };
 EXPORT_SYMBOL(pci_bus_type);
include/linux/device.h +4 -0
index c6f27207dbe8..3a1efa56e7c4 100644
@@ -97,6 +97,8 @@ extern void bus_remove_file(struct bus_type *, struct bus_attribute *);
  * @p:		The private data of the driver core, only the driver core can
  *		touch this.
  * @lock_key:	Lock class key for use by the lock validator
+ * @force_dma:	Assume devices on this bus should be set up by dma_configure()
+ * 		even if DMA capability is not explicitly described by firmware.
  *
  * A bus is a channel between the processor and one or more devices. For the
  * purposes of the device model, all devices are connected via a bus, even if

@@ -135,6 +137,8 @@ struct bus_type {
 
 	struct subsys_private *p;
 	struct lock_class_key lock_key;
+
+	bool force_dma;
 };
 
 extern int __must_check bus_register(struct bus_type *bus);
Signature · 1 line
2.13.4.dirty
#4 lore raw

PATCH 2/2 of: configure DMA for host1x devices

Quoted 4 lines · Robin Murphy

Alternatively (and modulo Greg's opinion, obviously) it would arguably be neatest to handle this at a slightly lower level as below - I only really considered this idea after writing the existing fix, but now that I have actually written it up I do rather like it.

I also like this much better than hardcoding specific busses. The only real downside is that it makes it too easy to add exceptions :)

#5 lore raw

PATCH 2/2 of: configure DMA for host1x devices

In reply to Robin Murphy
Quoted 39 lines · Robin Murphy
Quoted 33 lines · Thierry Reding
Quoted 17 lines · Mikko Perttunen

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
  drivers/of/device.c | 3 +++
  1 file changed, 3 insertions(+)
drivers/of/device.c +0 -0
index 64b710265d39..12368418cd33 100644

drivers/of/device.c int of_dma_configure(struct device *dev, struct device_node *np) · line 112

Quoted 9 lines
drivers/of/device.c +3 -0
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
  		if (!dev_is_pci(dev) &&
  #ifdef CONFIG_ARM_AMBA
  		    dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+		    !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
  #endif
  		    dev->bus != &platform_bus_type)
  			return ret == -ENODEV ? 0 : ret;

Maybe a slightly better solution would be to add a dev_is_host1x() function that returns dev->bus == &host1x_bus_type if TEGRA_HOST1X is enabled and returns false otherwise. That way we can have a "proper" check for the bus type and avoid the #if in this file.

Quoted 4 lines

It's slightly more complicated because of the dependencies between the DT and Tegra DRM trees, but I'm sure we can get Rob to either give his Acked-by on the drivers/of/device.c change or carry the Tegra DRM patch in the DT tree.

Quoted 6 lines

Alternatively (and modulo Greg's opinion, obviously) it would arguably be neatest to handle this at a slightly lower level as below - I only really considered this idea after writing the existing fix, but now that I have actually written it up I do rather like it.

Robin.

Yep, this solution seems the cleanest to me as well.

Mikko

Quoted 132 lines

----->8-----
From: Robin Murphy <[email protected]>
Date: Mon, 25 Sep 2017 12:45:58 +0100
Subject: [PATCH] drivers: Flag buses which demand DMA configuration

We do not want the common dma_configure() pathway to apply indiscriminately to all devices, since there are plenty of buses which do not have DMA capability, and if their child devices were used for DMA API calls it would only be indicative of a driver bug. However, there are a number of buses for which DMA is implicitly expected even when not described by firmware, so we currently whitelist those to automatically opt in to dma_configure(), assuming a 1:1 mapping between the DMA address space and the physical address space.

Commit 723288836628 ("of: restrict DMA configuration") introduced a short-term fix by comparing explicit bus types, but this approach is far from pretty, doesn't scale well, and fails to cope at all with bus drivers which may be built as modules. Let's refine things by making that opt-in a property of the bus type, which neatly addresses those problems and lets the decision of whether firmware description of DMA capability should be optional or mandatory stay internal to the bus drivers themselves.

Signed-off-by Robin Murphy <[email protected]>
---
  drivers/amba/bus.c       | 1 +
  drivers/base/platform.c  | 1 +
  drivers/gpu/host1x/bus.c | 1 +
  drivers/of/device.c      | 8 +-------
  drivers/pci/pci-driver.c | 1 +
  include/linux/device.h   | 4 ++++
  6 files changed, 9 insertions(+), 7 deletions(-)
drivers/amba/bus.c +1 -0
index e0f74ddc22b7..594c228d2f02 100644
@@ -195,6 +195,7 @@ struct bus_type amba_bustype = {
  	.match		= amba_match,
  	.uevent		= amba_uevent,
  	.pm		= &amba_pm,
+	.force_dma	= true,
  };
 
  static int __init amba_init(void)
drivers/base/platform.c +1 -0
index d1bd99271066..9c95c8db00f8 100644
@@ -1142,6 +1142,7 @@ struct bus_type platform_bus_type = {
  	.match		= platform_match,
  	.uevent		= platform_uevent,
  	.pm		= &platform_dev_pm_ops,
+	.force_dma	= true,
  };
  EXPORT_SYMBOL_GPL(platform_bus_type);
drivers/gpu/host1x/bus.c +1 -0
index f9cde03030fd..ed03b3243195 100644
@@ -320,6 +320,7 @@ struct bus_type host1x_bus_type = {
  	.name = "host1x",
  	.match = host1x_device_match,
  	.pm = &host1x_device_pm_ops,
+	.force_dma = true,
  };
 
  static void __host1x_device_del(struct host1x_device *device)
drivers/of/device.c +1 -7
index 64b710265d39..25bddf9c9fe1 100644
@@ -9,9 +9,7 @@
  #include <linux/module.h>
  #include <linux/mod_devicetable.h>
  #include <linux/slab.h>
-#include <linux/pci.h>
  #include <linux/platform_device.h>
-#include <linux/amba/bus.h>
 
  #include <asm/errno.h>
  #include "of_private.h"

@@ -101,11 +99,7 @@ int of_dma_configure(struct device *dev, struct device_node *np)
  		 * DMA configuration regardless of whether "dma-ranges" is
  		 * correctly specified or not.
  		 */
-		if (!dev_is_pci(dev) &&
-#ifdef CONFIG_ARM_AMBA
-		    dev->bus != &amba_bustype &&
-#endif
-		    dev->bus != &platform_bus_type)
+		if (!dev->bus->force_dma)
  			return ret == -ENODEV ? 0 : ret;
 
  		dma_addr = offset = 0;
drivers/pci/pci-driver.c +1 -0
index 11bd267fc137..38bdb97b6dc6 100644
@@ -1466,6 +1466,7 @@ struct bus_type pci_bus_type = {
  	.drv_groups	= pci_drv_groups,
  	.pm		= PCI_PM_OPS_PTR,
  	.num_vf		= pci_bus_num_vf,
+	.force_dma	= true,
  };
  EXPORT_SYMBOL(pci_bus_type);
include/linux/device.h +4 -0
index c6f27207dbe8..3a1efa56e7c4 100644
@@ -97,6 +97,8 @@ extern void bus_remove_file(struct bus_type *, struct bus_attribute *);
   * @p:		The private data of the driver core, only the driver core can
   *		touch this.
   * @lock_key:	Lock class key for use by the lock validator
+ * @force_dma:	Assume devices on this bus should be set up by dma_configure()
+ * 		even if DMA capability is not explicitly described by firmware.
   *
   * A bus is a channel between the processor and one or more devices. For the
   * purposes of the device model, all devices are connected via a bus, even if

@@ -135,6 +137,8 @@ struct bus_type {
 
  	struct subsys_private *p;
  	struct lock_class_key lock_key;
+
+	bool force_dma;
  };
 
  extern int __must_check bus_register(struct bus_type *bus);
#6 lore raw

PATCH 2/2 of: configure DMA for host1x devices

In reply to Robin Murphy
Quoted 168 lines · Robin Murphy
Quoted 33 lines · Thierry Reding
Quoted 17 lines · Mikko Perttunen

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/of/device.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/of/device.c +0 -0
index 64b710265d39..12368418cd33 100644

drivers/of/device.c int of_dma_configure(struct device *dev, struct device_node *np) · line 112

Quoted 9 lines
drivers/of/device.c +3 -0
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		if (!dev_is_pci(dev) &&
 #ifdef CONFIG_ARM_AMBA
 		    dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+		    !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
 #endif
 		    dev->bus != &platform_bus_type)
 			return ret == -ENODEV ? 0 : ret;

Maybe a slightly better solution would be to add a dev_is_host1x() function that returns dev->bus == &host1x_bus_type if TEGRA_HOST1X is enabled and returns false otherwise. That way we can have a "proper" check for the bus type and avoid the #if in this file.

Quoted 4 lines

It's slightly more complicated because of the dependencies between the DT and Tegra DRM trees, but I'm sure we can get Rob to either give his Acked-by on the drivers/of/device.c change or carry the Tegra DRM patch in the DT tree.

Alternatively (and modulo Greg's opinion, obviously) it would arguably be neatest to handle this at a slightly lower level as below - I only really considered this idea after writing the existing fix, but now that I have actually written it up I do rather like it.

Robin.

----->8-----
From: Robin Murphy <[email protected]>
Date: Mon, 25 Sep 2017 12:45:58 +0100
Subject: [PATCH] drivers: Flag buses which demand DMA configuration

We do not want the common dma_configure() pathway to apply indiscriminately to all devices, since there are plenty of buses which do not have DMA capability, and if their child devices were used for DMA API calls it would only be indicative of a driver bug. However, there are a number of buses for which DMA is implicitly expected even when not described by firmware, so we currently whitelist those to automatically opt in to dma_configure(), assuming a 1:1 mapping between the DMA address space and the physical address space.

Commit 723288836628 ("of: restrict DMA configuration") introduced a short-term fix by comparing explicit bus types, but this approach is far from pretty, doesn't scale well, and fails to cope at all with bus drivers which may be built as modules. Let's refine things by making that opt-in a property of the bus type, which neatly addresses those problems and lets the decision of whether firmware description of DMA capability should be optional or mandatory stay internal to the bus drivers themselves.

Signed-off-by Robin Murphy <[email protected]>
---
 drivers/amba/bus.c       | 1 +
 drivers/base/platform.c  | 1 +
 drivers/gpu/host1x/bus.c | 1 +
 drivers/of/device.c      | 8 +-------
 drivers/pci/pci-driver.c | 1 +
 include/linux/device.h   | 4 ++++
 6 files changed, 9 insertions(+), 7 deletions(-)
drivers/amba/bus.c +1 -0
index e0f74ddc22b7..594c228d2f02 100644
@@ -195,6 +195,7 @@ struct bus_type amba_bustype = {
 	.match		= amba_match,
 	.uevent		= amba_uevent,
 	.pm		= &amba_pm,
+	.force_dma	= true,
 };
 
 static int __init amba_init(void)
drivers/base/platform.c +1 -0
index d1bd99271066..9c95c8db00f8 100644
@@ -1142,6 +1142,7 @@ struct bus_type platform_bus_type = {
 	.match		= platform_match,
 	.uevent		= platform_uevent,
 	.pm		= &platform_dev_pm_ops,
+	.force_dma	= true,
 };
 EXPORT_SYMBOL_GPL(platform_bus_type);
drivers/gpu/host1x/bus.c +1 -0
index f9cde03030fd..ed03b3243195 100644
@@ -320,6 +320,7 @@ struct bus_type host1x_bus_type = {
 	.name = "host1x",
 	.match = host1x_device_match,
 	.pm = &host1x_device_pm_ops,
+	.force_dma = true,
 };
 
 static void __host1x_device_del(struct host1x_device *device)
drivers/of/device.c +1 -7
index 64b710265d39..25bddf9c9fe1 100644
@@ -9,9 +9,7 @@
 #include <linux/module.h>
 #include <linux/mod_devicetable.h>
 #include <linux/slab.h>
-#include <linux/pci.h>
 #include <linux/platform_device.h>
-#include <linux/amba/bus.h>
 
 #include <asm/errno.h>
 #include "of_private.h"

@@ -101,11 +99,7 @@ int of_dma_configure(struct device *dev, struct device_node *np)
 		 * DMA configuration regardless of whether "dma-ranges" is
 		 * correctly specified or not.
 		 */
-		if (!dev_is_pci(dev) &&
-#ifdef CONFIG_ARM_AMBA
-		    dev->bus != &amba_bustype &&
-#endif
-		    dev->bus != &platform_bus_type)
+		if (!dev->bus->force_dma)
 			return ret == -ENODEV ? 0 : ret;
 
 		dma_addr = offset = 0;
drivers/pci/pci-driver.c +1 -0
index 11bd267fc137..38bdb97b6dc6 100644
@@ -1466,6 +1466,7 @@ struct bus_type pci_bus_type = {
 	.drv_groups	= pci_drv_groups,
 	.pm		= PCI_PM_OPS_PTR,
 	.num_vf		= pci_bus_num_vf,
+	.force_dma	= true,
 };
 EXPORT_SYMBOL(pci_bus_type);
include/linux/device.h +2 -0
index c6f27207dbe8..3a1efa56e7c4 100644
@@ -97,6 +97,8 @@ extern void bus_remove_file(struct bus_type *, struct bus_attribute *);
  * @p:		The private data of the driver core, only the driver core can
  *		touch this.
  * @lock_key:	Lock class key for use by the lock validator
+ * @force_dma:	Assume devices on this bus should be set up by dma_configure()
+ * 		even if DMA capability is not explicitly described by firmware.
  *
  * A bus is a channel between the processor and one or more devices. For the
  * purposes of the device model, all devices are connected via a bus, even if

include/linux/device.h struct bus_type { · line 141

Quoted 5 lines
include/linux/device.h +2 -0
@@ -135,6 +137,8 @@ struct bus_type {
 
 	struct subsys_private *p;
 	struct lock_class_key lock_key;
+
+	bool force_dma;

Maybe force_dma_setup to clarify that we're not trying to force DMA usage, but rather setup of DMA properties.

Either way:

Acked-by Thierry Reding <[email protected]>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170926/ac787ff8/attachment.sig>

maintainer
#7 lore raw

PATCH 2/2 of: configure DMA for host1x devices

In reply to Robin Murphy
Quoted 69 lines · Robin Murphy
Quoted 33 lines · Thierry Reding
Quoted 17 lines · Mikko Perttunen

Devices in the host1x bus rely on the old behavior of of_dma_configure to set up DMA ops. Add a check for them into of_dma_configure.

We must do the check using a string comparison instead of using pointers since the host1x bus can be compiled into a module.

Fixes 723288836628 ("of: restrict DMA configuration")
Signed-off-by Mikko Perttunen <[email protected]>
---
 drivers/of/device.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/of/device.c +0 -0
index 64b710265d39..12368418cd33 100644

drivers/of/device.c int of_dma_configure(struct device *dev, struct device_node *np) · line 112

Quoted 9 lines
drivers/of/device.c +3 -0
@@ -104,6 +104,9 @@ int of_dma_configure(struct device *dev, struct device_node *np)
             if (!dev_is_pci(dev) &&
 #ifdef CONFIG_ARM_AMBA
                 dev->bus != &amba_bustype &&
+#endif
+#if IS_ENABLED(CONFIG_TEGRA_HOST1X)
+                !(dev->bus && !strcmp("host1x", dev->bus->name)) &&
 #endif
                 dev->bus != &platform_bus_type)
                     return ret == -ENODEV ? 0 : ret;

Maybe a slightly better solution would be to add a dev_is_host1x() function that returns dev->bus == &host1x_bus_type if TEGRA_HOST1X is enabled and returns false otherwise. That way we can have a "proper" check for the bus type and avoid the #if in this file.

Quoted 4 lines

It's slightly more complicated because of the dependencies between the DT and Tegra DRM trees, but I'm sure we can get Rob to either give his Acked-by on the drivers/of/device.c change or carry the Tegra DRM patch in the DT tree.

Alternatively (and modulo Greg's opinion, obviously) it would arguably be neatest to handle this at a slightly lower level as below - I only really considered this idea after writing the existing fix, but now that I have actually written it up I do rather like it.

Robin.

----->8-----
From: Robin Murphy <[email protected]>
Date: Mon, 25 Sep 2017 12:45:58 +0100
Subject: [PATCH] drivers: Flag buses which demand DMA configuration

We do not want the common dma_configure() pathway to apply indiscriminately to all devices, since there are plenty of buses which do not have DMA capability, and if their child devices were used for DMA API calls it would only be indicative of a driver bug. However, there are a number of buses for which DMA is implicitly expected even when not described by firmware, so we currently whitelist those to automatically opt in to dma_configure(), assuming a 1:1 mapping between the DMA address space and the physical address space.

Commit 723288836628 ("of: restrict DMA configuration") introduced a short-term fix by comparing explicit bus types, but this approach is far from pretty, doesn't scale well, and fails to cope at all with bus drivers which may be built as modules. Let's refine things by making that opt-in a property of the bus type, which neatly addresses those problems and lets the decision of whether firmware description of DMA capability should be optional or mandatory stay internal to the bus drivers themselves.

Quoted 1 line
Signed-off-by Robin Murphy <[email protected]>
Acked-by Rob Herring <[email protected]>

Are you going to send out as a proper patch? We've now got another case with pci_epf_bus_type...

Rob

Quoted 108 lines
---
 drivers/amba/bus.c       | 1 +
 drivers/base/platform.c  | 1 +
 drivers/gpu/host1x/bus.c | 1 +
 drivers/of/device.c      | 8 +-------
 drivers/pci/pci-driver.c | 1 +
 include/linux/device.h   | 4 ++++
 6 files changed, 9 insertions(+), 7 deletions(-)
drivers/amba/bus.c +1 -0
index e0f74ddc22b7..594c228d2f02 100644
@@ -195,6 +195,7 @@ struct bus_type amba_bustype = {
        .match          = amba_match,
        .uevent         = amba_uevent,
        .pm             = &amba_pm,
+       .force_dma      = true,
 };
 
 static int __init amba_init(void)
drivers/base/platform.c +1 -0
index d1bd99271066..9c95c8db00f8 100644
@@ -1142,6 +1142,7 @@ struct bus_type platform_bus_type = {
        .match          = platform_match,
        .uevent         = platform_uevent,
        .pm             = &platform_dev_pm_ops,
+       .force_dma      = true,
 };
 EXPORT_SYMBOL_GPL(platform_bus_type);
drivers/gpu/host1x/bus.c +1 -0
index f9cde03030fd..ed03b3243195 100644
@@ -320,6 +320,7 @@ struct bus_type host1x_bus_type = {
        .name = "host1x",
        .match = host1x_device_match,
        .pm = &host1x_device_pm_ops,
+       .force_dma = true,
 };
 
 static void __host1x_device_del(struct host1x_device *device)
drivers/of/device.c +1 -7
index 64b710265d39..25bddf9c9fe1 100644
@@ -9,9 +9,7 @@
 #include <linux/module.h>
 #include <linux/mod_devicetable.h>
 #include <linux/slab.h>
-#include <linux/pci.h>
 #include <linux/platform_device.h>
-#include <linux/amba/bus.h>
 
 #include <asm/errno.h>
 #include "of_private.h"

@@ -101,11 +99,7 @@ int of_dma_configure(struct device *dev, struct device_node *np)
                 * DMA configuration regardless of whether "dma-ranges" is
                 * correctly specified or not.
                 */
-               if (!dev_is_pci(dev) &&
-#ifdef CONFIG_ARM_AMBA
-                   dev->bus != &amba_bustype &&
-#endif
-                   dev->bus != &platform_bus_type)
+               if (!dev->bus->force_dma)
                        return ret == -ENODEV ? 0 : ret;
 
                dma_addr = offset = 0;
drivers/pci/pci-driver.c +1 -0
index 11bd267fc137..38bdb97b6dc6 100644
@@ -1466,6 +1466,7 @@ struct bus_type pci_bus_type = {
        .drv_groups     = pci_drv_groups,
        .pm             = PCI_PM_OPS_PTR,
        .num_vf         = pci_bus_num_vf,
+       .force_dma      = true,
 };
 EXPORT_SYMBOL(pci_bus_type);
include/linux/device.h +4 -1
index c6f27207dbe8..3a1efa56e7c4 100644
@@ -97,6 +97,8 @@ extern void bus_remove_file(struct bus_type *, struct bus_attribute *);
  * @p:         The private data of the driver core, only the driver core can
  *             touch this.
  * @lock_key:  Lock class key for use by the lock validator
+ * @force_dma: Assume devices on this bus should be set up by dma_configure()
+ *             even if DMA capability is not explicitly described by firmware.
  *
  * A bus is a channel between the processor and one or more devices. For the
  * purposes of the device model, all devices are connected via a bus, even if

@@ -135,6 +137,8 @@ struct bus_type {
 
        struct subsys_private *p;
        struct lock_class_key lock_key;
+
+       bool force_dma;
 };
 
 extern int __must_check bus_register(struct bus_type *bus);
--

2.13.4.dirty