Skip to contentSkip to conversation

xen/virtio: Handle PCI devices which Host controller is described in DT

PATCH Superseded Oleksandr Tyshchenko 4 participants 16 messages Oct 15, 2022 – Oct 21, 2022 Xen hypervisor interface

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 xen/virtio: Handle PCI devices which Host controller is described in DT

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://lore.kernel.org/xen-devel/[email protected]/

Based on:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1

---
 drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
 1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +76 -11
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
 #include <linux/module.h>
 #include <linux/dma-map-ops.h>
 #include <linux/of.h>
+#include <linux/pci.h>
 #include <linux/pfn.h>
 #include <linux/xarray.h>
 #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
 	.dma_supported = xen_grant_dma_supported,
 };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}
+
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+
+	return ret;
+}
+
 static bool xen_is_dt_grant_dma_device(struct device *dev)
 {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
 	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
 	has_iommu = iommu_np &&
 		    of_device_is_compatible(iommu_np, "xen,grant-dma");
 	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
 bool xen_is_grant_dma_device(struct device *dev)
 {
+	struct device_node *np;
+
 	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}
 
 	return false;
 }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
 static int xen_dt_grant_init_backend_domid(struct device *dev,
 					   struct xen_grant_dma_data *data)
 {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
 	}
 
 	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
 void xen_grant_setup_dma_ops(struct device *dev)
 {
 	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
 	data = find_xen_grant_dma_data(dev);
 	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
 	if (!data)
 		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
 			goto err;
 	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
 		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
#1 lore raw
Quoted 9 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Hi Oleksandr,

From another email I understood that you successfully managed to describe in device tree all the individual virtio pci devices so that you can have iommu-map/iommu-map-mask properties under each virtio device node. Is that right?

If that is the case, then I would rather jump straight to that approach because I think it is far better than this one.

Cheers,

Stefano

Quoted 167 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://lore.kernel.org/xen-devel/[email protected]/

Based on:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1

---
 drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
 1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +76 -11
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
 #include <linux/module.h>
 #include <linux/dma-map-ops.h>
 #include <linux/of.h>
+#include <linux/pci.h>
 #include <linux/pfn.h>
 #include <linux/xarray.h>
 #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
 	.dma_supported = xen_grant_dma_supported,
 };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}
+
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+
+	return ret;
+}
+
 static bool xen_is_dt_grant_dma_device(struct device *dev)
 {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
 	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
 	has_iommu = iommu_np &&
 		    of_device_is_compatible(iommu_np, "xen,grant-dma");
 	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
 bool xen_is_grant_dma_device(struct device *dev)
 {
+	struct device_node *np;
+
 	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}
 
 	return false;
 }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
 static int xen_dt_grant_init_backend_domid(struct device *dev,
 					   struct xen_grant_dma_data *data)
 {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
 	}
 
 	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
 void xen_grant_setup_dma_ops(struct device *dev)
 {
 	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
 	data = find_xen_grant_dma_data(dev);
 	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
 	if (!data)
 		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
 			goto err;
 	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
 		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
#2 lore raw

On 18.10.22 03:33, Stefano Stabellini wrote:

Hello Stefano

Quoted 10 lines · Oleksandr Tyshchenko
Quoted 9 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 6 lines

Hi Oleksandr,

From another email I understood that you successfully managed to
describe in device tree all the individual virtio pci devices so that you can have iommu-map/iommu-map-mask properties under each virtio device node. Is that right?

No. Here [1] I mentioned that I had experimented with PCI-IOMMU bindings (iommu-map/iommu-map-mask properties) as IOMMU bindings (iommu property) is insufficient for us and got it worked.
Also I provided a link to the current patch. Sorry, if I was unclear.

Just to be clear:

We do not describe in device-tree all the individual virtio-pci devices (and we do not have to), we only describe generic PCI host bridge node. So we have only a *single* iommu-map property under that PCI host bridge node.
The iommu-map property in turn describes the IOMMU connections for the endpoints within that PCI Host bridge according to:
https://www.kernel.org/doc/Documentation/devicetree/bindings/pci/pci-iommu.txt

For the instance, the following iommu-map property under that PCI host bridge node describes the relationship between IOMMU and two PCI devices (0000:00:01.0 and 0000:00:02.0):
iommu-map = <0x08 0xfde9 0x01 0x08 0x10 0xfde9 0x02 0x08>; For 0000:00:01.0 we pass the endpoint ID 1 (backend domid 1) For 0000:00:02.0 we pass the endpoint ID 2 (backend domid 2) Other PCI devices (i.e 0000:00:03.0) are untranslated (are not required to use grants for the virtio).

Quoted 3 lines

If that is the case, then I would rather jump straight to that approach because I think it is far better than this one.

Please see above, I don't have any other approach except the one implemented in current patch.

[1]
https://lore.kernel.org/xen-devel/[email protected]/

Quoted 174 lines

Cheers,

Stefano

Quoted 167 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyV6yWUy3w$  [lore[.]kernel[.]org]

Based on:

https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyWa-6yyug$  [git[.]kernel[.]org]
---
  drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +76 -11
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
  	.dma_supported = xen_grant_dma_supported,
  };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}
+
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+
+	return ret;
+}
+
  static bool xen_is_dt_grant_dma_device(struct device *dev)
  {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
  	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
  	has_iommu = iommu_np &&
  		    of_device_is_compatible(iommu_np, "xen,grant-dma");
  	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
  bool xen_is_grant_dma_device(struct device *dev)
  {
+	struct device_node *np;
+
  	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}
 
  	return false;
  }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
  static int xen_dt_grant_init_backend_domid(struct device *dev,
  					   struct xen_grant_dma_data *data)
  {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
  	}
 
  	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
  void xen_grant_setup_dma_ops(struct device *dev)
  {
  	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
  	data = find_xen_grant_dma_data(dev);
  	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
  	if (!data)
  		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
  			goto err;
  	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
  		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#3 lore raw
Quoted 33 lines · Oleksandr Tyshchenko
Quoted 10 lines · Stefano Stabellini
Quoted 9 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 6 lines

Hi Oleksandr,

From another email I understood that you successfully managed to
describe in device tree all the individual virtio pci devices so that you can have iommu-map/iommu-map-mask properties under each virtio device node. Is that right?

No. Here [1] I mentioned that I had experimented with PCI-IOMMU bindings (iommu-map/iommu-map-mask properties) as IOMMU bindings (iommu property) is insufficient for us and got it worked.
Also I provided a link to the current patch. Sorry, if I was unclear.

Just to be clear:

We do not describe in device-tree all the individual virtio-pci devices (and we do not have to), we only describe generic PCI host bridge node. So we have only a *single* iommu-map property under that PCI host bridge node.
The iommu-map property in turn describes the IOMMU connections for the endpoints within that PCI Host bridge according to:
https://www.kernel.org/doc/Documentation/devicetree/bindings/pci/pci-iommu.txt

Quoted 8 lines

For the instance, the following iommu-map property under that PCI host bridge node describes the relationship between IOMMU and two PCI devices (0000:00:01.0 and 0000:00:02.0):
iommu-map = <0x08 0xfde9 0x01 0x08 0x10 0xfde9 0x02 0x08>; For 0000:00:01.0 we pass the endpoint ID 1 (backend domid 1) For 0000:00:02.0 we pass the endpoint ID 2 (backend domid 2) Other PCI devices (i.e 0000:00:03.0) are untranslated (are not required to use grants for the virtio).

That's great! I misunderstood. Actually I wonder if iommu-map might be suitable also for hotplug devices (as long as the backend domid is known beforehand). I think that should work? It should be possible to specify PCI device IDs even if those device IDs are not present yet?

If this work, it could be the best solution actually.

Quoted 189 lines
Quoted 2 lines

If that is the case, then I would rather jump straight to that approach because I think it is far better than this one.

Please see above, I don't have any other approach except the one implemented in current patch.

[1]
https://lore.kernel.org/xen-devel/[email protected]/

Quoted 174 lines

Cheers,

Stefano

Quoted 167 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyV6yWUy3w$  [lore[.]kernel[.]org]

Based on:

https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyWa-6yyug$  [git[.]kernel[.]org]
---
  drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +76 -11
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
  	.dma_supported = xen_grant_dma_supported,
  };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}
+
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+
+	return ret;
+}
+
  static bool xen_is_dt_grant_dma_device(struct device *dev)
  {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
  	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
  	has_iommu = iommu_np &&
  		    of_device_is_compatible(iommu_np, "xen,grant-dma");
  	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
  bool xen_is_grant_dma_device(struct device *dev)
  {
+	struct device_node *np;
+
  	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}
 
  	return false;
  }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
  static int xen_dt_grant_init_backend_domid(struct device *dev,
  					   struct xen_grant_dma_data *data)
  {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
  	}
 
  	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
  void xen_grant_setup_dma_ops(struct device *dev)
  {
  	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
  	data = find_xen_grant_dma_data(dev);
  	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
  	if (!data)
  		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
  			goto err;
  	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
  		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#4 lore raw

On 19.10.22 03:34, Stefano Stabellini wrote:

Hello Stefano

Quoted 41 lines · Oleksandr Tyshchenko
Quoted 32 lines · Stefano Stabellini
Quoted 10 lines · Oleksandr Tyshchenko
Quoted 9 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 6 lines

Hi Oleksandr,

From another email I understood that you successfully managed to
describe in device tree all the individual virtio pci devices so that you can have iommu-map/iommu-map-mask properties under each virtio device node. Is that right?

No. Here [1] I mentioned that I had experimented with PCI-IOMMU bindings (iommu-map/iommu-map-mask properties) as IOMMU bindings (iommu property) is insufficient for us and got it worked.
Also I provided a link to the current patch. Sorry, if I was unclear.

Just to be clear:

We do not describe in device-tree all the individual virtio-pci devices (and we do not have to), we only describe generic PCI host bridge node. So we have only a *single* iommu-map property under that PCI host bridge node.
The iommu-map property in turn describes the IOMMU connections for the endpoints within that PCI Host bridge according to:

https://urldefense.com/v3/__https://www.kernel.org/doc/Documentation/devicetree/bindings/pci/pci-iommu.txt__;!!GF_29dbcQIUBPA!078KT_6M5f7P5_m6O2EotvWED-yuSZKHzzqykDoW5DPtQOWJQeoZB4QWsJCqCkP-wFtLry5TiLAz3uhNnB2ccNY9CsN57Q$  [kernel[.]org]
Quoted 8 lines

For the instance, the following iommu-map property under that PCI host bridge node describes the relationship between IOMMU and two PCI devices (0000:00:01.0 and 0000:00:02.0):
iommu-map = <0x08 0xfde9 0x01 0x08 0x10 0xfde9 0x02 0x08>; For 0000:00:01.0 we pass the endpoint ID 1 (backend domid 1) For 0000:00:02.0 we pass the endpoint ID 2 (backend domid 2) Other PCI devices (i.e 0000:00:03.0) are untranslated (are not required to use grants for the virtio).

Quoted 3 lines

That's great! I misunderstood. Actually I wonder if iommu-map might be suitable also for hotplug devices (as long as the backend domid is known beforehand). I think that should work?

I don't see at the moment any reasons why actually not. I assume, it would also work.
For hotplug devices the arch_setup_dma_ops() -> ... -> xen_grant_setup_dma_ops()
will also be called as it is called for *boot* devices, isn't it?

Quoted 2 lines

It should be possible to specify
PCI device IDs even if those device IDs are not present yet?

I may mistake, but I think, yes. I think, nothing prevents us from doing so when creating PCI Host bridge node in the toolstack (or Xen if it is a dom0less).
I think, we do not violate anything. We just describe the IOMMU mapping scheme for PCI devices. If PCI device with specified RID appears at some point in future,
the corresponding SID (backend domid) will be assigned to it, if not appears - nothing bad will happen.

I think, this is similar to interrupt-map property where we describe the interrupt mapping scheme for PCI devices.

Quoted 2 lines

If this work, it could be the best solution actually.

thanks, I cannot say for sure whether it will 100% work as we don't have a working hotplug at the moment, so it is not possible to re-check, but I don't see why that solution won't work for us.

Quoted 189 lines
Quoted 189 lines
Quoted 4 lines

If that is the case, then I would rather jump straight to that approach because I think it is far better than this one.

Please see above, I don't have any other approach except the one implemented in current patch.

[1]

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!078KT_6M5f7P5_m6O2EotvWED-yuSZKHzzqykDoW5DPtQOWJQeoZB4QWsJCqCkP-wFtLry5TiLAz3uhNnB2ccNaoN0AUOw$  [lore[.]kernel[.]org]
Quoted 173 lines

Cheers,

Stefano

Quoted 167 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyV6yWUy3w$  [lore[.]kernel[.]org]

Based on:

https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!1mSETxg8CRohlL5OpYo0VaLBXtbWRLZlam9QABMP_YUzsYcrn8no1FxBPvhQnNRCSzp3pkC1dXIgmhdaZmJ3oyWa-6yyug$  [git[.]kernel[.]org]
---
   drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
   1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +76 -11
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
   #include <linux/module.h>
   #include <linux/dma-map-ops.h>
   #include <linux/of.h>
+#include <linux/pci.h>
   #include <linux/pfn.h>
   #include <linux/xarray.h>
   #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
   	.dma_supported = xen_grant_dma_supported,
   };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}
+
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+
+	return ret;
+}
+
   static bool xen_is_dt_grant_dma_device(struct device *dev)
   {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
   	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
   	has_iommu = iommu_np &&
   		    of_device_is_compatible(iommu_np, "xen,grant-dma");
   	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
   bool xen_is_grant_dma_device(struct device *dev)
   {
+	struct device_node *np;
+
   	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}
 
   	return false;
   }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
   static int xen_dt_grant_init_backend_domid(struct device *dev,
   					   struct xen_grant_dma_data *data)
   {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
   	}
 
   	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
   void xen_grant_setup_dma_ops(struct device *dev)
   {
   	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
   	data = find_xen_grant_dma_data(dev);
   	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
   	if (!data)
   		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
   			goto err;
   	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
   		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#5 lore raw
In reply to Oleksandr Tyshchenko
Quoted 10 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

Please add an example of the bindings in the commit message.

Quoted 37 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://lore.kernel.org/xen-devel/[email protected]/

Based on:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1

---
 drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
 1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +2 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
 #include <linux/module.h>
 #include <linux/dma-map-ops.h>
 #include <linux/of.h>
+#include <linux/pci.h>
 #include <linux/pfn.h>
 #include <linux/xarray.h>
 #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
 	.dma_supported = xen_grant_dma_supported,
 };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 48 lines
drivers/xen/grant-dma-ops.c +34 -2
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+	return ret;
+}
+
 static bool xen_is_dt_grant_dma_device(struct device *dev)
 {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
 	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
 	has_iommu = iommu_np &&
 		    of_device_is_compatible(iommu_np, "xen,grant-dma");
 	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
 bool xen_is_grant_dma_device(struct device *dev)
 {
+	struct device_node *np;
+
 	/* XXX Handle only DT devices for now */

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 10 lines
drivers/xen/grant-dma-ops.c +8 -2
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 54 lines
drivers/xen/grant-dma-ops.c +20 -7
 	return false;
 }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
 static int xen_dt_grant_init_backend_domid(struct device *dev,
 					   struct xen_grant_dma_data *data)
 {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
 	}
 
 	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
 void xen_grant_setup_dma_ops(struct device *dev)
 {
 	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
 	data = find_xen_grant_dma_data(dev);
 	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
 	if (!data)
 		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
 			goto err;
 	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
 		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
#6 lore raw
Quoted 128 lines · Stefano Stabellini
Quoted 10 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

Please add an example of the bindings in the commit message.

Quoted 37 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://lore.kernel.org/xen-devel/[email protected]/

Based on:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1

---
  drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +2 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
  	.dma_supported = xen_grant_dma_supported,
  };
 
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 53 lines
drivers/xen/grant-dma-ops.c +37 -4
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+	return ret;
+}
+
  static bool xen_is_dt_grant_dma_device(struct device *dev)
  {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
  	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
  	has_iommu = iommu_np &&
  		    of_device_is_compatible(iommu_np, "xen,grant-dma");
  	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
  bool xen_is_grant_dma_device(struct device *dev)
  {
+	struct device_node *np;
+
  	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
+		bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)). So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves.
This will simplify the code.

Quoted 56 lines
Quoted 55 lines
drivers/xen/grant-dma-ops.c +20 -7
  	return false;
  }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
  static int xen_dt_grant_init_backend_domid(struct device *dev,
  					   struct xen_grant_dma_data *data)
  {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
  	}
 
  	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
  void xen_grant_setup_dma_ops(struct device *dev)
  {
  	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
  	data = find_xen_grant_dma_data(dev);
  	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
  	if (!data)
  		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
  			goto err;
  	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
  		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 1 line
Xenia
#7 lore raw

On 19.10.22 11:47, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 149 lines · Stefano Stabellini
Quoted 138 lines · Oleksandr Tyshchenko
Quoted 11 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

Please add an example of the bindings in the commit message.

Quoted 45 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid to the guest
than for platform devices.

Changes V1 -> V2:
    - update commit description
    - rebase
    - rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$
[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$
[git[.]kernel[.]org]
---
  drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----

  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
   #include <linux/module.h>
   #include <linux/dma-map-ops.h>
   #include <linux/of.h>
+#include <linux/pci.h>
   #include <linux/pfn.h>
   #include <linux/xarray.h>
   #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops 
 xen_grant_dma_ops = {
       .dma_supported = xen_grant_dma_supported,
   };

  +static struct device_node *xen_dt_get_pci_host_node(struct device *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +9 -0
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node 

**iommu_np,
+             u32 *sid)

drivers/xen/grant-dma-ops.c +11 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", 

iommu_np, sid);
+    of_node_put(host_np);
+    return ret;

drivers/xen/grant-dma-ops.c +2 -0
+}
+

  static bool xen_is_dt_grant_dma_device(struct device *dev) {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
      bool has_iommu;
  -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +
      has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

drivers/xen/grant-dma-ops.c +5 -2
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct 
 device *dev)
     bool xen_is_grant_dma_device(struct device *dev)
   {
+    struct device_node *np;
+
       /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 7 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves. This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev) {
    struct pci_host_bridge *bridge =
pci_find_host_bridge(to_pci_dev(dev)->bus);

    return of_node_get(bridge->dev.parent->of_node);
}

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

Quoted 58 lines
Quoted 57 lines
Quoted 56 lines
drivers/xen/grant-dma-ops.c +12 -4
       return false;
   }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device 
 *dev)
   static int xen_dt_grant_init_backend_domid(struct device *dev,
                          struct xen_grant_dma_data *data)
   {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
   -    if (of_parse_phandle_with_args(dev->of_node, "iommus", 
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus", 
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }

      }
        if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int 
 xen_dt_grant_init_backend_domid(struct device *dev,
   void xen_grant_setup_dma_ops(struct device *dev)
   {
       struct xen_grant_dma_data *data;
+    struct device_node *np;
         data = find_xen_grant_dma_data(dev);
       if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
       if (!data)
           goto err;
   -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
               goto err;
       } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
           dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#8 lore raw

On 10/19/22 22:41, Oleksandr Tyshchenko wrote:

Hi Oleksandr

Quoted 170 lines · Xenia Ragiadakou

Hello Xenia

Quoted 149 lines · Stefano Stabellini
Quoted 138 lines · Oleksandr Tyshchenko
Quoted 11 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

Please add an example of the bindings in the commit message.

Quoted 45 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid to the guest
than for platform devices.

Changes V1 -> V2:
    - update commit description
    - rebase
    - rework to use generic PCI-IOMMU bindings instead of generic
IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$
[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$
[git[.]kernel[.]org]

---
   drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----
   1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
   #include <linux/module.h>
   #include <linux/dma-map-ops.h>
   #include <linux/of.h>
+#include <linux/pci.h>
   #include <linux/pfn.h>
   #include <linux/xarray.h>
   #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
       .dma_supported = xen_grant_dma_supported,
   };
   +static struct device_node *xen_dt_get_pci_host_node(struct device
 *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops · line 307

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +37 -3
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node
 **iommu_np,
+             u32 *sid)
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",
 iommu_np, sid);
+    of_node_put(host_np);
+    return ret;
+}
+
   static bool xen_is_dt_grant_dma_device(struct device *dev)
   {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
       bool has_iommu;
   -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
       has_iommu = iommu_np &&
               of_device_is_compatible(iommu_np, "xen,grant-dma");
       of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
     bool xen_is_grant_dma_device(struct device *dev)
   {
+    struct device_node *np;
+
       /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 7 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves. This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

Quoted 8 lines

static struct device_node *xen_dt_get_pci_host_node(struct device *dev) {
    struct pci_host_bridge *bridge =
pci_find_host_bridge(to_pci_dev(dev)->bus);

    return of_node_get(bridge->dev.parent->of_node);
}

You are right. I prefer your version instead of the above.

Quoted 10 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just an idea):

bool xen_is_grant_dma_device(struct device *dev)

{
    struct device_node *np;
    bool has_iommu = false;

    /* XXX Handle only DT devices for now */
    np = xen_dt_get_node(dev);
    if (np)
        has_iommu = xen_is_dt_grant_dma_device(dev, np);
    of_node_put(np);
    return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev,
struct device_node *np)

{
    struct device_node *iommu_np = NULL;
    bool has_iommu;

    if (dev_is_pci(dev)) {
        struct pci_dev *pdev = to_pci_dev(dev);
	u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn);
        of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np, NULL);
    } else {
        iommu_np = of_parse_phandle(np, "iommus", 0);
    }
	
    has_iommu = iommu_np && of_device_is_compatible(iommu_np, 
"xen,grant-dma");
    of_node_put(iommu_np);

    return has_iommu;
}

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Quoted 58 lines
Quoted 58 lines
Quoted 57 lines
Quoted 56 lines
drivers/xen/grant-dma-ops.c +20 -5
       return false;
   }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device
 *dev)
   static int xen_dt_grant_init_backend_domid(struct device *dev,
                          struct xen_grant_dma_data *data)
   {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
   -    if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }
       }
         if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
   void xen_grant_setup_dma_ops(struct device *dev)
   {
       struct xen_grant_dma_data *data;
+    struct device_node *np;
         data = find_xen_grant_dma_data(dev);
       if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
       if (!data)
           goto err;
   -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
               goto err;
       } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
           dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 1 line
Xenia
#9 lore raw
Quoted 3 lines · Xenia Ragiadakou

On 10/19/22 22:41, Oleksandr Tyshchenko wrote:

Hi Oleksandr

Hello Xenia

Quoted 180 lines
Quoted 180 lines · Xenia Ragiadakou

Hello Xenia

Quoted 153 lines · Stefano Stabellini
Quoted 142 lines · Oleksandr Tyshchenko
Quoted 11 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review
the patch :-)

Please add an example of the bindings in the commit message.

Quoted 48 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid to the guest
than for platform devices.

Changes V1 -> V2:
     - update commit description
     - rebase
     - rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$

[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$

[git[.]kernel[.]org]
---
   drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----

   1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
    #include <linux/module.h>
    #include <linux/dma-map-ops.h>
    #include <linux/of.h>
+#include <linux/pci.h>
    #include <linux/pfn.h>
    #include <linux/xarray.h>
    #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
        .dma_supported = xen_grant_dma_supported,
    };

   +static struct device_node *xen_dt_get_pci_host_node(struct device *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +9 -0
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node

**iommu_np,
+             u32 *sid)

drivers/xen/grant-dma-ops.c +11 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",

iommu_np, sid);
+    of_node_put(host_np);
+    return ret;

drivers/xen/grant-dma-ops.c +2 -0
+}
+

   static bool xen_is_dt_grant_dma_device(struct device *dev) {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
       bool has_iommu;
   -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +
       has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

drivers/xen/grant-dma-ops.c +5 -2
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
      bool xen_is_grant_dma_device(struct device *dev)
    {
+    struct device_node *np;
+
        /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 7 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves. This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev) {
      struct pci_host_bridge *bridge = pci_find_host_bridge(to_pci_dev(dev)->bus);

Quoted 5 lines
Quoted 3 lines

      return of_node_get(bridge->dev.parent->of_node); }

You are right. I prefer your version instead of the above.

ok, thanks

Quoted 45 lines
Quoted 12 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just an idea):

bool xen_is_grant_dma_device(struct device *dev)
{
    struct device_node *np;
    bool has_iommu = false;

    /* XXX Handle only DT devices for now */
    np = xen_dt_get_node(dev);
    if (np)
        has_iommu = xen_is_dt_grant_dma_device(dev, np); of_node_put(np);
    return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev, struct device_node *np) {
    struct device_node *iommu_np = NULL;
    bool has_iommu;

    if (dev_is_pci(dev)) {
        struct pci_dev *pdev = to_pci_dev(dev);
    u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn); of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np, NULL);
    } else {
        iommu_np = of_parse_phandle(np, "iommus", 0);
    }

Quoted 6 lines

    has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma");
    of_node_put(iommu_np);

    return has_iommu;
}

I got it.

xen_is_grant_dma_device() for V3 won't call xen_dt_get_node(), but call xen_is_dt_grant_dma_device() directly.

static bool xen_is_dt_grant_dma_device(struct device *dev) {
    struct device_node *iommu_np = NULL;
    bool has_iommu;

    if (dev_is_pci(dev)) {
        if (xen_dt_map_id(dev, &iommu_np, NULL)) return false;
    } else if (dev->of_node)
        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); else
        return false;

    has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

    return has_iommu;
}

bool xen_is_grant_dma_device(struct device *dev)
{
    /* XXX Handle only DT devices for now */
    return xen_is_dt_grant_dma_device(dev);
}

Quoted 4 lines

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Good question. I think, it is possible... This is technically what V1 is doing.

Are you asking because to support "iommus" for PCI devices as well to describe that use-case with all PCI devices having the same endpoint ID (backend ID)?
If yes, I think, this could be still described by "iommu-map" property, something like that (if we don't want to describe mapping for each PCI device one-by-one).

iommu-map = <0x0 &iommu X 0x1>;

iommu-map-mask = <0x0>;

where the X is backend ID.

It feels to me that it should be written down somewhere that for platform devices we expect "iommus" and for PCI devices we expect "iommu-map/iommu-map-mask" to be present.

Quoted 62 lines
Quoted 61 lines
Quoted 61 lines
Quoted 60 lines
Quoted 59 lines
drivers/xen/grant-dma-ops.c +12 -4
        return false;
    }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device
 *dev)
    static int xen_dt_grant_init_backend_domid(struct device *dev,
                           struct xen_grant_dma_data *data)
    {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
    -    if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }

       }
         if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
    void xen_grant_setup_dma_ops(struct device *dev)
    {
        struct xen_grant_dma_data *data;
+    struct device_node *np;
          data = find_xen_grant_dma_data(dev);
        if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
        if (!data)
            goto err;
    -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
                goto err;
        } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
            dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#10 lore raw
Quoted 275 lines · Oleksandr Tyshchenko
Quoted 3 lines · Xenia Ragiadakou

On 10/19/22 22:41, Oleksandr Tyshchenko wrote:

Hi Oleksandr

Hello Xenia

Quoted 180 lines
Quoted 180 lines · Xenia Ragiadakou

Hello Xenia

Quoted 153 lines · Stefano Stabellini
Quoted 142 lines · Oleksandr Tyshchenko
Quoted 11 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review
the patch :-)

Please add an example of the bindings in the commit message.

Quoted 48 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid to the guest
than for platform devices.

Changes V1 -> V2:
     - update commit description
     - rebase
     - rework to use generic PCI-IOMMU bindings instead of generic
IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$

[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$

[git[.]kernel[.]org]

---
    drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----
    1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
    #include <linux/module.h>
    #include <linux/dma-map-ops.h>
    #include <linux/of.h>
+#include <linux/pci.h>
    #include <linux/pfn.h>
    #include <linux/xarray.h>
    #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
        .dma_supported = xen_grant_dma_supported,
    };
    +static struct device_node *xen_dt_get_pci_host_node(struct device
 *dev)

drivers/xen/grant-dma-ops.c static const struct dma_map_ops · line 307

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +37 -3
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node
 **iommu_np,
+             u32 *sid)
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",
 iommu_np, sid);
+    of_node_put(host_np);
+    return ret;
+}
+
    static bool xen_is_dt_grant_dma_device(struct device *dev)
    {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
        bool has_iommu;
    -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
        has_iommu = iommu_np &&
                of_device_is_compatible(iommu_np, "xen,grant-dma");
        of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
      bool xen_is_grant_dma_device(struct device *dev)
    {
+    struct device_node *np;
+
        /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 7 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves. This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev) {
      struct pci_host_bridge *bridge =
pci_find_host_bridge(to_pci_dev(dev)->bus);

Quoted 5 lines
Quoted 3 lines

      return of_node_get(bridge->dev.parent->of_node);
}

You are right. I prefer your version instead of the above.

ok, thanks

Quoted 45 lines
Quoted 12 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just an idea):

bool xen_is_grant_dma_device(struct device *dev)
{
    struct device_node *np;
    bool has_iommu = false;

    /* XXX Handle only DT devices for now */
    np = xen_dt_get_node(dev);
    if (np)
        has_iommu = xen_is_dt_grant_dma_device(dev, np); of_node_put(np);
    return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev,
                                       struct device_node *np)
{
    struct device_node *iommu_np = NULL;
    bool has_iommu;

    if (dev_is_pci(dev)) {
        struct pci_dev *pdev = to_pci_dev(dev);
    u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn); of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np,
NULL);
    } else {
        iommu_np = of_parse_phandle(np, "iommus", 0);
    }

Quoted 6 lines

    has_iommu = iommu_np && of_device_is_compatible(iommu_np,
"xen,grant-dma");
    of_node_put(iommu_np);

    return has_iommu;
}

I got it.

xen_is_grant_dma_device() for V3 won't call xen_dt_get_node(), but call xen_is_dt_grant_dma_device() directly.

static bool xen_is_dt_grant_dma_device(struct device *dev) {
    struct device_node *iommu_np = NULL;
    bool has_iommu;

      if (dev_is_pci(dev)) {
          if (xen_dt_map_id(dev, &iommu_np, NULL))
              return false;
      } else if (dev->of_node)
          iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
      else
          return false;

      has_iommu = iommu_np &&
              of_device_is_compatible(iommu_np, "xen,grant-dma");
      of_node_put(iommu_np);
Quoted 10 lines

    return has_iommu;
}

bool xen_is_grant_dma_device(struct device *dev)
{
    /* XXX Handle only DT devices for now */
    return xen_is_dt_grant_dma_device(dev);
}

Ok. One difference, that I see from the previous, is that here you don't use the dynamic interface when you access the dev->of_node (of_node_get/of_node_put). Before, this was guarded through the external xen_dt_get_node().

I suspect that the same needs to be done for the function xen_grant_setup_dma_ops(). There, also, the code walks up to the root bus twice.

Quoted 18 lines
Quoted 4 lines

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Good question. I think, it is possible... This is technically what V1 is doing.

Are you asking because to support "iommus" for PCI devices as well to describe that use-case with all PCI devices having the same endpoint ID (backend ID)?
If yes, I think, this could be still described by "iommu-map" property, something like that (if we don't want to describe mapping for each PCI device one-by-one).

iommu-map = <0x0 &iommu X 0x1>;

Quoted 8 lines

iommu-map-mask = <0x0>;

where the X is backend ID.

It feels to me that it should be written down somewhere that for platform devices we expect "iommus" and for PCI devices we expect "iommu-map/iommu-map-mask" to be present.

Thanks for the clarification, now I got it. Yes I agree.

Quoted 62 lines
Quoted 62 lines
Quoted 61 lines
Quoted 61 lines
Quoted 60 lines
Quoted 59 lines
drivers/xen/grant-dma-ops.c +12 -4
        return false;
    }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device
 *dev)
    static int xen_dt_grant_init_backend_domid(struct device *dev,
                           struct xen_grant_dma_data *data)
    {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
    -    if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }
        }
          if (!of_device_is_compatible(iommu_spec.np,

"xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
    void xen_grant_setup_dma_ops(struct device *dev)
    {
        struct xen_grant_dma_data *data;
+    struct device_node *np;
          data = find_xen_grant_dma_data(dev);
        if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
        if (!data)
            goto err;
    -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
                goto err;
        } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
            dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 1 line
Xenia
#11 lore raw

On 20.10.22 21:11, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 297 lines · Oleksandr Tyshchenko
Quoted 285 lines · Xenia Ragiadakou
Quoted 3 lines · Oleksandr Tyshchenko

Hi Oleksandr

Hello Xenia

Quoted 190 lines
Quoted 190 lines · Xenia Ragiadakou

Hello Xenia

Quoted 160 lines · Stefano Stabellini
Quoted 149 lines · Oleksandr Tyshchenko
Quoted 14 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration.
The problem is that PCI devices under the single PCI Host controller
may have the backends running in different Xen domains and thus have
different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review
the patch :-)

Please add an example of the bindings in the commit message.

Quoted 52 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid
to the guest
than for platform devices.

Changes V1 -> V2:
      - update commit description
      - rebase
      - rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$

[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$

[git[.]kernel[.]org]
---
    drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----

    1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
     #include <linux/module.h>
     #include <linux/dma-map-ops.h>
     #include <linux/of.h>
+#include <linux/pci.h>
     #include <linux/pfn.h>
     #include <linux/xarray.h>
     #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
         .dma_supported = xen_grant_dma_supported,
     };

    +static struct device_node *xen_dt_get_pci_host_node(struct device
*dev)

drivers/xen/grant-dma-ops.c

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +9 -0
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node

**iommu_np,
+             u32 *sid)

drivers/xen/grant-dma-ops.c +11 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",

iommu_np, sid);
+    of_node_put(host_np);
+    return ret;

drivers/xen/grant-dma-ops.c +2 -0
+}
+

    static bool xen_is_dt_grant_dma_device(struct device *dev) {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
        bool has_iommu;
    -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +
        has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

drivers/xen/grant-dma-ops.c +5 -2
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
       bool xen_is_grant_dma_device(struct device *dev)
     {
+    struct device_node *np;
+
         /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 8 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves.
This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have
to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
{
       struct pci_host_bridge *bridge = pci_find_host_bridge(to_pci_dev(dev)->bus);

Quoted 5 lines
Quoted 3 lines

       return of_node_get(bridge->dev.parent->of_node); }

You are right. I prefer your version instead of the above.

ok, thanks

Quoted 45 lines
Quoted 12 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just an idea):

bool xen_is_grant_dma_device(struct device *dev)
{
     struct device_node *np;
     bool has_iommu = false;

     /* XXX Handle only DT devices for now */
     np = xen_dt_get_node(dev);
     if (np)
         has_iommu = xen_is_dt_grant_dma_device(dev, np); of_node_put(np);
     return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev, struct device_node *np) {
     struct device_node *iommu_np = NULL;
     bool has_iommu;

     if (dev_is_pci(dev)) {
         struct pci_dev *pdev = to_pci_dev(dev);
     u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn); of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np, NULL);
     } else {
         iommu_np = of_parse_phandle(np, "iommus", 0);
     }

Quoted 6 lines

     has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma");
     of_node_put(iommu_np);

     return has_iommu;
}

I got it.

xen_is_grant_dma_device() for V3 won't call xen_dt_get_node(), but call xen_is_dt_grant_dma_device() directly.

static bool xen_is_dt_grant_dma_device(struct device *dev) {
      struct device_node *iommu_np = NULL;
      bool has_iommu;

      if (dev_is_pci(dev)) {
          if (xen_dt_map_id(dev, &iommu_np, NULL)) return false;
      } else if (dev->of_node)
          iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); else
          return false;

      has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

Quoted 10 lines

      return has_iommu;
}

bool xen_is_grant_dma_device(struct device *dev)
{
      /* XXX Handle only DT devices for now */
      return xen_is_dt_grant_dma_device(dev);
}

Quoted 8 lines

Ok. One difference, that I see from the previous, is that here you don't use the dynamic interface when you access the dev->of_node (of_node_get/of_node_put). Before, this was guarded through the external xen_dt_get_node().

I suspect that the same needs to be done for the function xen_grant_setup_dma_ops(). There, also, the code walks up to the root bus twice.

Hmm, xen_dt_grant_init_backend_domid() should only be called if we deal with device-tree based device.

I think you are completely right, thanks!

In order to address both your comments, I think I need to rework the code (taking into the account your example with xen_is_dt_grant_dma_device()

provided a few letters ago and extrapolate this example to xen_dt_grant_init_backend_domid()). Below the patch (not tested) which seems to address both your comments (also I dropped

xen_dt_map_id() and squashed xen_dt_get_pci_host_node() with xen_dt_get_node()).

drivers/xen/grant-dma-ops.c +5 -2
index daa525df7bdc..dae24dbd2ef7 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,33 @@ static const struct dma_map_ops xen_grant_dma_ops = {
         .dma_supported = xen_grant_dma_supported,
  };
 
-static bool xen_is_dt_grant_dma_device(struct device *dev)
+static struct device_node *xen_dt_get_node(struct device *dev)
  {
-       struct device_node *iommu_np;
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               struct pci_host_bridge *bridge = 

pci_find_host_bridge(pdev->bus);
+
+               return of_node_get(bridge->dev.parent->of_node); +       }
+
+       return of_node_get(dev->of_node);
+}
+
+static bool xen_is_dt_grant_dma_device(struct device *dev, +                                       struct device_node *np) +{
+       struct device_node *iommu_np = NULL;
        bool has_iommu;

-       iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn); +
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask", &iommu_np, NULL))
+                       return false;
+       } else
+               iommu_np = of_parse_phandle(np, "iommus", 0); +
        has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);
@@ -307,9 +329,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)

 bool xen_is_grant_dma_device(struct device *dev) {
+       struct device_node *np;
+
        /* XXX Handle only DT devices for now */
-       if (dev->of_node)
-               return xen_is_dt_grant_dma_device(dev); +       np = xen_dt_get_node(dev);
+       if (np) {
+               bool ret;
+
+               ret = xen_is_dt_grant_dma_device(dev, np); +               of_node_put(np);
+               return ret;
+       }

        return false;
 }
@@ -323,14 +353,26 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
 }

 static int xen_dt_grant_init_backend_domid(struct device *dev,
+                                          struct device_node *np,
                                           struct xen_grant_dma_data *data) {
-       struct of_phandle_args iommu_spec;
+       struct of_phandle_args iommu_spec = { .args_count = 1 };

-       if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-                       0, &iommu_spec)) {
-               dev_err(dev, "Cannot parse iommus property\n"); -               return -ESRCH;
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn); +
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask", &iommu_spec.np,
+                               iommu_spec.args)) {
+                       dev_err(dev, "Cannot translate ID\n");
+                       return -ESRCH;
+               }
+       } else {
+               if (of_parse_phandle_with_args(np, "iommus", "#iommu-cells", +                               0, &iommu_spec)) {
+                       dev_err(dev, "Cannot parse iommus property\n"); +                       return -ESRCH;
+               }
        }

        if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +396,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
 void xen_grant_setup_dma_ops(struct device *dev)
 {
        struct xen_grant_dma_data *data;
+       struct device_node *np;

        data = find_xen_grant_dma_data(dev);
        if (data) {
@@ -365,8 +408,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
        if (!data)
                goto err;

-       if (dev->of_node) {
-               if (xen_dt_grant_init_backend_domid(dev, data)) +       np = xen_dt_get_node(dev);
+       if (np) {
+               int ret;
+
+               ret = xen_dt_grant_init_backend_domid(dev, np, data); +               of_node_put(np);
+               if (ret)
                        goto err;
        } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) { dev_info(dev, "Using dom0 as backend\n");

Does it look ok now?

Quoted 22 lines
Quoted 22 lines
Quoted 6 lines

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Good question. I think, it is possible... This is technically what V1 is doing.

Are you asking because to support "iommus" for PCI devices as well to describe that use-case with all PCI devices having the same endpoint ID (backend ID)?
If yes, I think, this could be still described by "iommu-map" property, something like that (if we don't want to describe mapping for each PCI device one-by-one).

iommu-map = <0x0 &iommu X 0x1>;

iommu-map-mask = <0x0>;

Quoted 8 lines
Quoted 6 lines

where the X is backend ID.

It feels to me that it should be written down somewhere that for platform devices we expect "iommus" and for PCI devices we expect "iommu-map/iommu-map-mask" to be present.

Thanks for the clarification, now I got it. Yes I agree.

ok, good

Quoted 65 lines
Quoted 64 lines
Quoted 64 lines
Quoted 63 lines
Quoted 63 lines
Quoted 62 lines
Quoted 61 lines
drivers/xen/grant-dma-ops.c +12 -4
         return false;
     }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device
 *dev)
     static int xen_dt_grant_init_backend_domid(struct device *dev,
                            struct xen_grant_dma_data *data)
     {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
     -    if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }

        }
          if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
     void xen_grant_setup_dma_ops(struct device *dev)
     {
         struct xen_grant_dma_data *data;
+    struct device_node *np;
           data = find_xen_grant_dma_data(dev);
         if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device 
 *dev)
         if (!data)
             goto err;
     -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
                 goto err;
         } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {

            dev_info(dev, "Using dom0 as backend\n");

Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#12 lore raw

On 10/20/22 23:07, Oleksandr Tyshchenko wrote:
Hi Oleksandr

Quoted 350 lines · Xenia Ragiadakou

Hello Xenia

Quoted 297 lines · Oleksandr Tyshchenko
Quoted 285 lines · Xenia Ragiadakou
Quoted 3 lines · Oleksandr Tyshchenko

Hi Oleksandr

Hello Xenia

Quoted 190 lines
Quoted 190 lines · Xenia Ragiadakou

Hello Xenia

Quoted 160 lines · Stefano Stabellini
Quoted 149 lines · Oleksandr Tyshchenko
Quoted 14 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration.
The problem is that PCI devices under the single PCI Host controller
may have the backends running in different Xen domains and thus have
different endpoints ID (backend domains ID).

Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review
the patch :-)

Please add an example of the bindings in the commit message.

Quoted 52 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid
to the guest
than for platform devices.

Changes V1 -> V2:
      - update commit description
      - rebase
      - rework to use generic PCI-IOMMU bindings instead of generic
IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$

[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$

[git[.]kernel[.]org]

---
     drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----
     1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
     #include <linux/module.h>
     #include <linux/dma-map-ops.h>
     #include <linux/of.h>
+#include <linux/pci.h>
     #include <linux/pfn.h>
     #include <linux/xarray.h>
     #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
         .dma_supported = xen_grant_dma_supported,
     };
     +static struct device_node *xen_dt_get_pci_host_node(struct

device
*dev)

drivers/xen/grant-dma-ops.c

Quoted 10 lines
drivers/xen/grant-dma-ops.c +10 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 54 lines
drivers/xen/grant-dma-ops.c +9 -0
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node

**iommu_np,
+             u32 *sid)

drivers/xen/grant-dma-ops.c +11 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",

iommu_np, sid);
+    of_node_put(host_np);
+    return ret;

drivers/xen/grant-dma-ops.c +14 -3
+}
+
     static bool xen_is_dt_grant_dma_device(struct device *dev)
     {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
         bool has_iommu;
     -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
         has_iommu = iommu_np &&
                 of_device_is_compatible(iommu_np, "xen,grant-dma");
         of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
       bool xen_is_grant_dma_device(struct device *dev)
     {
+    struct device_node *np;
+
         /* XXX Handle only DT devices for now */
-    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev);
+    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;
Quoted 9 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct · line 364

Quoted 5 lines
drivers/xen/grant-dma-ops.c +5 -0
+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 8 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves.
This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have
to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
{
       struct pci_host_bridge *bridge =
pci_find_host_bridge(to_pci_dev(dev)->bus);

Quoted 5 lines
Quoted 3 lines

       return of_node_get(bridge->dev.parent->of_node);
}

You are right. I prefer your version instead of the above.

ok, thanks

Quoted 45 lines
Quoted 12 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when executing xen_is_grant_dma_device() for PCI device:

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just an idea):

bool xen_is_grant_dma_device(struct device *dev)
{
     struct device_node *np;
     bool has_iommu = false;

     /* XXX Handle only DT devices for now */
     np = xen_dt_get_node(dev);
     if (np)
         has_iommu = xen_is_dt_grant_dma_device(dev, np); of_node_put(np);
     return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev,
                                        struct device_node *np)
{
     struct device_node *iommu_np = NULL;
     bool has_iommu;

     if (dev_is_pci(dev)) {
         struct pci_dev *pdev = to_pci_dev(dev);
     u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn); of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np,
NULL);
     } else {
         iommu_np = of_parse_phandle(np, "iommus", 0);
     }

Quoted 6 lines

     has_iommu = iommu_np && of_device_is_compatible(iommu_np,
"xen,grant-dma");
     of_node_put(iommu_np);

     return has_iommu;
}

I got it.

xen_is_grant_dma_device() for V3 won't call xen_dt_get_node(), but call xen_is_dt_grant_dma_device() directly.

static bool xen_is_dt_grant_dma_device(struct device *dev) {
      struct device_node *iommu_np = NULL;
      bool has_iommu;

      if (dev_is_pci(dev)) {
          if (xen_dt_map_id(dev, &iommu_np, NULL)) return false;
      } else if (dev->of_node)
          iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); else
          return false;

      has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

Quoted 10 lines

      return has_iommu;
}

bool xen_is_grant_dma_device(struct device *dev)
{
      /* XXX Handle only DT devices for now */
      return xen_is_dt_grant_dma_device(dev);
}

Quoted 8 lines

Ok. One difference, that I see from the previous, is that here you don't use the dynamic interface when you access the dev->of_node (of_node_get/of_node_put). Before, this was guarded through the external xen_dt_get_node().

I suspect that the same needs to be done for the function xen_grant_setup_dma_ops(). There, also, the code walks up to the root bus twice.

Hmm, xen_dt_grant_init_backend_domid() should only be called if we deal with device-tree based device.

I think you are completely right, thanks!

In order to address both your comments, I think I need to rework the code (taking into the account your example with xen_is_dt_grant_dma_device()

provided a few letters ago and extrapolate this example to xen_dt_grant_init_backend_domid()). Below the patch (not tested) which seems to address both your comments (also I dropped

xen_dt_map_id() and squashed xen_dt_get_pci_host_node() with xen_dt_get_node()).

drivers/xen/grant-dma-ops.c +3 -2
index daa525df7bdc..dae24dbd2ef7 100644
@@ -10,6 +10,7 @@
   #include <linux/module.h>
   #include <linux/dma-map-ops.h>
   #include <linux/of.h>
+#include <linux/pci.h>
   #include <linux/pfn.h>
   #include <linux/xarray.h>
   #include <linux/virtio_anchor.h>
@@ -292,12 +293,33 @@ static const struct dma_map_ops xen_grant_dma_ops = {
          .dma_supported = xen_grant_dma_supported,
   };
 
-static bool xen_is_dt_grant_dma_device(struct device *dev)
+static struct device_node *xen_dt_get_node(struct device *dev)
   {
-       struct device_node *iommu_np;
+       if (dev_is_pci(dev)) {

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 308

Quoted 10 lines
drivers/xen/grant-dma-ops.c +9 -0
+               struct pci_dev *pdev = to_pci_dev(dev);
+               struct pci_host_bridge *bridge =
 pci_find_host_bridge(pdev->bus);
+
+               return of_node_get(bridge->dev.parent->of_node);
+       }
+
+       return of_node_get(dev->of_node);
+}
+

It does not seem right to me to expose the struct pci_host_bridge (which we would need to check if it is null by the way). I would prefer your version for the above i.e static struct device_node *xen_dt_get_node(struct device *dev)

{
	if (dev_is_pci(dev)) {
		struct pci_bus *bus = to_pci_dev(dev)->bus;

		/* Walk up to the root bus to look for PCI Host controller */
		while (!pci_is_root_bus(bus))
			bus = bus->parent;
		return of_node_get(bus->bridge->parent->of_node);
	}

	return of_node_get(dev->of_node);
}
Quoted 65 lines
drivers/xen/grant-dma-ops.c +13 -1
+static bool xen_is_dt_grant_dma_device(struct device *dev,
+                                       struct device_node *np)
+{
+       struct device_node *iommu_np = NULL;
          bool has_iommu;
 
-       iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask",
 &iommu_np, NULL))
+                       return false;
+       } else
+               iommu_np = of_parse_phandle(np, "iommus", 0);
+
          has_iommu = iommu_np &&
                      of_device_is_compatible(iommu_np, "xen,grant-dma");
          of_node_put(iommu_np);
drivers/xen/grant-dma-ops.c +10 -2
@@ -307,9 +329,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
 
   bool xen_is_grant_dma_device(struct device *dev)
   {
+       struct device_node *np;
+
          /* XXX Handle only DT devices for now */
-       if (dev->of_node)
-               return xen_is_dt_grant_dma_device(dev);
+       np = xen_dt_get_node(dev);
+       if (np) {
+               bool ret;
+
+               ret = xen_is_dt_grant_dma_device(dev, np);
+               of_node_put(np);
+               return ret;
+       }
 
          return false;
   }
drivers/xen/grant-dma-ops.c +8 -5
@@ -323,14 +353,26 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
   }
 
   static int xen_dt_grant_init_backend_domid(struct device *dev,
+                                          struct device_node *np,
                                             struct xen_grant_dma_data *data)
   {
-       struct of_phandle_args iommu_spec;
+       struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-       if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-                       0, &iommu_spec)) {
-               dev_err(dev, "Cannot parse iommus property\n");
-               return -ESRCH;
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask",
 &iommu_spec.np,
+                               iommu_spec.args)) {

drivers/xen/grant-dma-ops.c bool xen_virtio_mem_acc(struct virtio_device *dev) · line 378

Quoted 10 lines
drivers/xen/grant-dma-ops.c +9 -0
+                       dev_err(dev, "Cannot translate ID\n");
+                       return -ESRCH;
+               }
+       } else {
+               if (of_parse_phandle_with_args(np, "iommus", "#iommu-cells",
+                               0, &iommu_spec)) {
+                       dev_err(dev, "Cannot parse iommus property\n");
+                       return -ESRCH;
+               }
          }

IMO, instead of passing struct xen_grant_dma_data *data to xen_dt_grant_init_backend_domid(), you could pass domid_t *backend_domid (e.g xen_dt_grant_init_backend_domid(dev, np, &data->backend_domid)). I think this way the internal struct xen_grant_dma_datain is manipulated in a single place and xen_dt_grant_init_backend_domid() does not depend on it.

Quoted 28 lines

        if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +1 -0
@@ -354,6 +396,7 @@ static int xen_dt_grant_init_backend_domid(struct
 device *dev,
   void xen_grant_setup_dma_ops(struct device *dev)
   {
          struct xen_grant_dma_data *data;
+       struct device_node *np;
 
          data = find_xen_grant_dma_data(dev);
          if (data) {
drivers/xen/grant-dma-ops.c +7 -2
@@ -365,8 +408,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
          if (!data)
                  goto err;
 
-       if (dev->of_node) {
-               if (xen_dt_grant_init_backend_domid(dev, data))
+       np = xen_dt_get_node(dev);
+       if (np) {
+               int ret;
+
+               ret = xen_dt_grant_init_backend_domid(dev, np, data);
+               of_node_put(np);
+               if (ret)
                          goto err;
          } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
                  dev_info(dev, "Using dom0 as backend\n");
Quoted 1 line

Does it look ok now?

That is what I had in mind. I do not know if Stefano agrees with this approach.

Quoted 100 lines
Quoted 22 lines
Quoted 22 lines
Quoted 6 lines

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Good question. I think, it is possible... This is technically what V1 is doing.

Are you asking because to support "iommus" for PCI devices as well to describe that use-case with all PCI devices having the same endpoint ID (backend ID)?
If yes, I think, this could be still described by "iommu-map" property, something like that (if we don't want to describe mapping for each PCI device one-by-one).

iommu-map = <0x0 &iommu X 0x1>;

iommu-map-mask = <0x0>;

Quoted 8 lines
Quoted 6 lines

where the X is backend ID.

It feels to me that it should be written down somewhere that for platform devices we expect "iommus" and for PCI devices we expect "iommu-map/iommu-map-mask" to be present.

Thanks for the clarification, now I got it. Yes I agree.

ok, good

Quoted 65 lines
Quoted 64 lines
Quoted 64 lines
Quoted 63 lines
Quoted 63 lines
Quoted 62 lines
Quoted 61 lines
drivers/xen/grant-dma-ops.c +12 -4
         return false;
     }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device
 *dev)
     static int xen_dt_grant_init_backend_domid(struct device *dev,
                            struct xen_grant_dma_data *data)
     {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 };
     -    if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n");
-        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus",
 "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n");
+            return -ESRCH;
+        }
         }
           if (!of_device_is_compatible(iommu_spec.np,

"xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
     void xen_grant_setup_dma_ops(struct device *dev)
     {
         struct xen_grant_dma_data *data;
+    struct device_node *np;
           data = find_xen_grant_dma_data(dev);
         if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device
 *dev)
         if (!data)
             goto err;
     -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
                 goto err;
         } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
             dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 1 line
Xenia
#13 lore raw

On 21.10.22 09:08, Xenia Ragiadakou wrote:

Hello Xenia

Quoted 384 lines · Oleksandr Tyshchenko

Hi Oleksandr

Quoted 363 lines · Xenia Ragiadakou

Hello Xenia

Quoted 308 lines · Oleksandr Tyshchenko
Quoted 296 lines · Xenia Ragiadakou
Quoted 3 lines · Oleksandr Tyshchenko

Hi Oleksandr

Hello Xenia

Quoted 198 lines
Quoted 198 lines · Xenia Ragiadakou

Hello Xenia

Quoted 168 lines · Stefano Stabellini
Quoted 157 lines · Oleksandr Tyshchenko
Quoted 14 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification.
Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration.
The problem is that PCI devices under the single PCI Host controller
may have the backends running in different Xen domains and thus have
different endpoints ID (backend domains ID).

Quoted 7 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask
properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Signed-off-by: Oleksandr Tyshchenko
<[email protected]>

Now that I understood the approach and the reasons for it, I can review
the patch :-)

Please add an example of the bindings in the commit message.

Quoted 63 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices
on Arm at some point in the future. The Xen toolstack side is not completely ready yet.
Here, for PCI devices we use more flexible way to pass backend domid
to the guest
than for platform devices.

Changes V1 -> V2:
       - update commit description
       - rebase
       - rework to use generic PCI-IOMMU bindings instead of generic
IOMMU bindings

Previous discussion is at:
https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauACie_ZAy$

[lore[.]kernel[.]org]

Based on:
https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!3-vq7Edm3XfKtD5cnNjnOzDQvuo_XrhJ73yH-nPfqOkGGU0IjLG7R7MR_nAJCAPeOutHRLT44wKYwQwz3SauAEnMDHAq$

[git[.]kernel[.]org]
---
     drivers/xen/grant-dma-ops.c | 87
++++++++++++++++++++++++++++++++-----

     1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
      #include <linux/module.h>
      #include <linux/dma-map-ops.h>
      #include <linux/of.h>
+#include <linux/pci.h>
      #include <linux/pfn.h>
      #include <linux/xarray.h>
      #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops
 xen_grant_dma_ops = {
          .dma_supported = xen_grant_dma_supported,
      };

     +static struct device_node *xen_dt_get_pci_host_node(struct device
*dev)

drivers/xen/grant-dma-ops.c +5 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    struct pci_bus *bus = pdev->bus;
+
+    /* Walk up to the root bus to look for PCI Host 

controller */
+    while (!pci_is_root_bus(bus))
+        bus = bus->parent;

drivers/xen/grant-dma-ops.c

Quoted 3 lines
drivers/xen/grant-dma-ops.c +3 -0
+
+    return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

Quoted 56 lines
drivers/xen/grant-dma-ops.c +9 -0
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+    if (dev_is_pci(dev))
+        return xen_dt_get_pci_host_node(dev);
+
+    return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node

**iommu_np,
+             u32 *sid)

drivers/xen/grant-dma-ops.c +11 -0
+{
+    struct pci_dev *pdev = to_pci_dev(dev);
+    u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+    struct device_node *host_np;
+    int ret;
+
+    host_np = xen_dt_get_pci_host_node(dev);
+    if (!host_np)
+        return -ENODEV;
+
+    ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask",

iommu_np, sid);
+    of_node_put(host_np);
+    return ret;

drivers/xen/grant-dma-ops.c +2 -0
+}
+

     static bool xen_is_dt_grant_dma_device(struct device *dev) {
-    struct device_node *iommu_np;
+    struct device_node *iommu_np = NULL;
         bool has_iommu;
     -    iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_np, NULL))
+            return false;
+    } else
+        iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); +
         has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma");
         of_node_put(iommu_np);

drivers/xen/grant-dma-ops.c +0 -0
@@ -307,9 +351,17 @@ static bool 

xen_is_dt_grant_dma_device(struct
device *dev)
       bool xen_is_grant_dma_device(struct device *dev) {
+    struct device_node *np;
+
         /* XXX Handle only DT devices for now */ -    if (dev->of_node)
-        return xen_is_dt_grant_dma_device(dev); +    np = xen_dt_get_node(dev);
+    if (np) {
+        bool ret;

Quoted 9 lines
Quoted 5 lines

+
+        ret = xen_is_dt_grant_dma_device(dev);
+        of_node_put(np);
+        return ret;
+    }

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Quoted 8 lines

I think in general we could pass directly the host bridge device if dev_is_pci(dev) (which can be retrieved with pci_get_host_bridge_device(to_pci_dev(dev), and after done with it pci_put_host_bridge_device(phb)).
So that, xen_is_dt_grant_dma_device() and xen_dt_grant_init_backend_domid() won't need to discover it themselves.
This will simplify the code.

Good point. I have some remark. Can we use pci_find_host_bridge() instead? This way we don't have to add #include "../pci/pci.h", and have
to drop reference afterwards.

With that xen_dt_get_pci_host_node() will became the following:

static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
{
        struct pci_host_bridge *bridge = pci_find_host_bridge(to_pci_dev(dev)->bus);

Quoted 5 lines
Quoted 3 lines

        return of_node_get(bridge->dev.parent->of_node); }

You are right. I prefer your version instead of the above.

ok, thanks

Quoted 47 lines
Quoted 7 lines

With Stefano's suggestion, we won't walk the PCI hierarchy twice when
executing xen_is_grant_dma_device() for PCI device:

Quoted 6 lines

xen_is_grant_dma_device() -> xen_is_dt_grant_dma_device() -> xen_dt_map_id() -> xen_dt_get_pci_host_node()

What do you think?

I was thinking passing the device_node along with the device in the function arguments. More specifically, of doing this (not tested, just
an idea):

bool xen_is_grant_dma_device(struct device *dev)
{
      struct device_node *np;
      bool has_iommu = false;

      /* XXX Handle only DT devices for now */
      np = xen_dt_get_node(dev);
      if (np)
          has_iommu = xen_is_dt_grant_dma_device(dev, np); of_node_put(np);
      return has_iommu;
}

static bool xen_is_dt_grant_dma_device(struct device *dev, struct device_node *np) {
      struct device_node *iommu_np = NULL;
      bool has_iommu;

      if (dev_is_pci(dev)) {
          struct pci_dev *pdev = to_pci_dev(dev);
      u32 id = PCI_DEVID(pdev->bus->number, pdev->devfn); of_map_id(np, id, "iommu-map", "iommu-map-mask", &iommu_np, NULL);
      } else {
          iommu_np = of_parse_phandle(np, "iommus", 0);
      }

Quoted 6 lines

      has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma");
      of_node_put(iommu_np);

      return has_iommu;
}

I got it.

xen_is_grant_dma_device() for V3 won't call xen_dt_get_node(), but call
xen_is_dt_grant_dma_device() directly.

static bool xen_is_dt_grant_dma_device(struct device *dev) {
       struct device_node *iommu_np = NULL;
       bool has_iommu;

       if (dev_is_pci(dev)) {
           if (xen_dt_map_id(dev, &iommu_np, NULL)) return false;
       } else if (dev->of_node)
           iommu_np = of_parse_phandle(dev->of_node, "iommus", 0); else
           return false;

       has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma"); of_node_put(iommu_np);

Quoted 10 lines

       return has_iommu;
}

bool xen_is_grant_dma_device(struct device *dev)
{
       /* XXX Handle only DT devices for now */
       return xen_is_dt_grant_dma_device(dev);
}

Quoted 8 lines

Ok. One difference, that I see from the previous, is that here you don't use the dynamic interface when you access the dev->of_node (of_node_get/of_node_put). Before, this was guarded through the external xen_dt_get_node().

I suspect that the same needs to be done for the function xen_grant_setup_dma_ops(). There, also, the code walks up to the root bus twice.

Hmm, xen_dt_grant_init_backend_domid() should only be called if we deal with device-tree based device.

I think you are completely right, thanks!

In order to address both your comments, I think I need to rework the code (taking into the account your example with xen_is_dt_grant_dma_device()

provided a few letters ago and extrapolate this example to xen_dt_grant_init_backend_domid()). Below the patch (not tested) which seems to address both your comments (also I dropped

xen_dt_map_id() and squashed xen_dt_get_pci_host_node() with xen_dt_get_node()).

drivers/xen/grant-dma-ops.c +3 -2
index daa525df7bdc..dae24dbd2ef7 100644
@@ -10,6 +10,7 @@
    #include <linux/module.h>
    #include <linux/dma-map-ops.h>
    #include <linux/of.h>
+#include <linux/pci.h>
    #include <linux/pfn.h>
    #include <linux/xarray.h>
    #include <linux/virtio_anchor.h>
@@ -292,12 +293,33 @@ static const struct dma_map_ops 
 xen_grant_dma_ops = {
           .dma_supported = xen_grant_dma_supported,
    };
 
-static bool xen_is_dt_grant_dma_device(struct device *dev)
+static struct device_node *xen_dt_get_node(struct device *dev)
    {
-       struct device_node *iommu_np;
+       if (dev_is_pci(dev)) {

drivers/xen/grant-dma-ops.c static const struct dma_map_ops · line 309

Quoted 10 lines
drivers/xen/grant-dma-ops.c +9 -0
+               struct pci_dev *pdev = to_pci_dev(dev);
+               struct pci_host_bridge *bridge =
 pci_find_host_bridge(pdev->bus);
+
+               return of_node_get(bridge->dev.parent->of_node);
+       }
+
+       return of_node_get(dev->of_node);
+}
+

It does not seem right to me to expose the struct pci_host_bridge (which we would need to check if it is null by the way). I would prefer your version for the above i.e
static struct device_node *xen_dt_get_node(struct device *dev) {
    if (dev_is_pci(dev)) {
        struct pci_bus *bus = to_pci_dev(dev)->bus;

Quoted 8 lines

        /* Walk up to the root bus to look for PCI Host controller */ while (!pci_is_root_bus(bus))
            bus = bus->parent;
        return of_node_get(bus->bridge->parent->of_node); }

    return of_node_get(dev->of_node);
}

ok, will return it back

Quoted 80 lines
Quoted 70 lines
drivers/xen/grant-dma-ops.c +13 -1
+static bool xen_is_dt_grant_dma_device(struct device *dev,
+                                       struct device_node *np)
+{
+       struct device_node *iommu_np = NULL;
           bool has_iommu;
 
-       iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask",
 &iommu_np, NULL))
+                       return false;
+       } else
+               iommu_np = of_parse_phandle(np, "iommus", 0);
+

          has_iommu = iommu_np && of_device_is_compatible(iommu_np, "xen,grant-dma");
          of_node_put(iommu_np);

drivers/xen/grant-dma-ops.c +10 -2
@@ -307,9 +329,17 @@ static bool xen_is_dt_grant_dma_device(struct
 device *dev)
 
    bool xen_is_grant_dma_device(struct device *dev)
    {
+       struct device_node *np;
+
           /* XXX Handle only DT devices for now */
-       if (dev->of_node)
-               return xen_is_dt_grant_dma_device(dev);
+       np = xen_dt_get_node(dev);
+       if (np) {
+               bool ret;
+
+               ret = xen_is_dt_grant_dma_device(dev, np);
+               of_node_put(np);
+               return ret;
+       }
 
           return false;

   }

drivers/xen/grant-dma-ops.c +1 -0
@@ -323,14 +353,26 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
    }
 
    static int xen_dt_grant_init_backend_domid(struct device *dev,
+                                          struct device_node *np,

                                             struct xen_grant_dma_data *data)
   {
-       struct of_phandle_args iommu_spec;
+       struct of_phandle_args iommu_spec = { .args_count = 1 };

-       if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-                       0, &iommu_spec)) {
-               dev_err(dev, "Cannot parse iommus property\n"); -               return -ESRCH;
+       if (dev_is_pci(dev)) {
+               struct pci_dev *pdev = to_pci_dev(dev);
+               u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn); +
+               if (of_map_id(np, rid, "iommu-map", "iommu-map-mask", &iommu_spec.np,
+                               iommu_spec.args)) {
+                       dev_err(dev, "Cannot translate ID\n"); +                       return -ESRCH;
+               }

Quoted 9 lines

+       } else {
+               if (of_parse_phandle_with_args(np, "iommus", "#iommu-cells",
+                               0, &iommu_spec)) {
+                       dev_err(dev, "Cannot parse iommus property\n"); +                       return -ESRCH;
+               }
          }

Quoted 7 lines

IMO, instead of passing struct xen_grant_dma_data *data to xen_dt_grant_init_backend_domid(), you could pass domid_t *backend_domid (e.g xen_dt_grant_init_backend_domid(dev, np, &data->backend_domid)).
I think this way the internal struct xen_grant_dma_datain is manipulated in a single place and xen_dt_grant_init_backend_domid() does not depend on it.

Although I think it is not directly related to current patch, I agree we could make this change as we touch the list of arguments for xen_dt_grant_init_backend_domid() anyway.

Quoted 31 lines
Quoted 31 lines

          if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +1 -0
@@ -354,6 +396,7 @@ static int xen_dt_grant_init_backend_domid(struct
 device *dev,
    void xen_grant_setup_dma_ops(struct device *dev)
    {
           struct xen_grant_dma_data *data;
+       struct device_node *np;
 
           data = find_xen_grant_dma_data(dev);

          if (data) {

drivers/xen/grant-dma-ops.c +7 -2
@@ -365,8 +408,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
           if (!data)
                   goto err;
 
-       if (dev->of_node) {
-               if (xen_dt_grant_init_backend_domid(dev, data))
+       np = xen_dt_get_node(dev);
+       if (np) {
+               int ret;
+
+               ret = xen_dt_grant_init_backend_domid(dev, np, data);
+               of_node_put(np);
+               if (ret)
                           goto err;
           } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
                   dev_info(dev, "Using dom0 as backend\n");
Quoted 3 lines
Quoted 1 line

Does it look ok now?

That is what I had in mind.

Great, thanks!

Quoted 1 line

I do not know if Stefano agrees with this approach.

We will see

Quoted 108 lines
Quoted 107 lines
Quoted 26 lines
Quoted 26 lines
Quoted 7 lines

I 'm wondering ... is it possible for the host bridge device node to have the iommus property set? meaning that all of its pci devs will have the same backend?

Good question. I think, it is possible... This is technically what V1 is
doing.

Are you asking because to support "iommus" for PCI devices as well to describe that use-case with all PCI devices having the same endpoint ID
(backend ID)?
If yes, I think, this could be still described by "iommu-map" property,
something like that (if we don't want to describe mapping for each PCI device one-by-one).

iommu-map = <0x0 &iommu X 0x1>;

iommu-map-mask = <0x0>;

Quoted 8 lines
Quoted 6 lines

where the X is backend ID.

It feels to me that it should be written down somewhere that for platform devices we expect "iommus" and for PCI devices we expect "iommu-map/iommu-map-mask" to be present.

Thanks for the clarification, now I got it. Yes I agree.

ok, good

Quoted 68 lines
Quoted 67 lines
Quoted 67 lines
Quoted 66 lines
Quoted 66 lines
Quoted 65 lines
Quoted 64 lines

         return false;
     }

drivers/xen/grant-dma-ops.c +0 -0
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct 

virtio_device
*dev)
     static int xen_dt_grant_init_backend_domid(struct device *dev,
                            struct xen_grant_dma_data *data) {
-    struct of_phandle_args iommu_spec;
+    struct of_phandle_args iommu_spec = { .args_count = 1 }; -    if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-            0, &iommu_spec)) {
-        dev_err(dev, "Cannot parse iommus property\n"); -        return -ESRCH;
+    if (dev_is_pci(dev)) {
+        if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+            dev_err(dev, "Cannot translate ID\n");
+            return -ESRCH;
+        }
+    } else {
+        if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+                0, &iommu_spec)) {
+            dev_err(dev, "Cannot parse iommus property\n"); +            return -ESRCH;
+        }
         }
           if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||

drivers/xen/grant-dma-ops.c +8 -1
@@ -354,6 +413,7 @@ static int
 xen_dt_grant_init_backend_domid(struct device *dev,
      void xen_grant_setup_dma_ops(struct device *dev)
      {
          struct xen_grant_dma_data *data;
+    struct device_node *np;
            data = find_xen_grant_dma_data(dev);
          if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device
 *dev)
          if (!data)
              goto err;
      -    if (dev->of_node) {
-        if (xen_dt_grant_init_backend_domid(dev, data))
+    np = xen_dt_get_node(dev);
+    if (np) {
+        int ret;
+
+        ret = xen_dt_grant_init_backend_domid(dev, data);
+        of_node_put(np);
+        if (ret)
                  goto err;
          } else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {

             dev_info(dev, "Using dom0 as backend\n");

Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#14 lore raw
In reply to Stefano Stabellini

On 19.10.22 03:58, Stefano Stabellini wrote:

Hello Stefano

Quoted 11 lines · Oleksandr Tyshchenko
Quoted 10 lines

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 7 lines
Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

perfect, thanks.

Quoted 2 lines

Please add an example of the bindings in the commit message.

ok, will do

Quoted 38 lines
Quoted 38 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!xJPdZO3-3Wmgo_79HuDsD53RkH_eAN96NmwuwFE7dArt_xNYGdD6LeLsq4B_QPrrvX-x23tJl6jQlNqgyNjgzT2NE3Pqjg$  [lore[.]kernel[.]org]

Based on:

https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!xJPdZO3-3Wmgo_79HuDsD53RkH_eAN96NmwuwFE7dArt_xNYGdD6LeLsq4B_QPrrvX-x23tJl6jQlNqgyNjgzT2J40LOxg$  [git[.]kernel[.]org]
---
  drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
  	.dma_supported = xen_grant_dma_supported,
  };
 
Quoted 13 lines

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 11 lines
drivers/xen/grant-dma-ops.c +11 -0
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

I also couldn't, but is it a really problem? This code is only gets called during initialization.

Quoted 54 lines
Quoted 54 lines
drivers/xen/grant-dma-ops.c +36 -4
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+	return ret;
+}
+
  static bool xen_is_dt_grant_dma_device(struct device *dev)
  {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
  	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
  	has_iommu = iommu_np &&
  		    of_device_is_compatible(iommu_np, "xen,grant-dma");
  	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
  bool xen_is_grant_dma_device(struct device *dev)
  {
+	struct device_node *np;
+
  	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
Quoted 8 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 6 lines
drivers/xen/grant-dma-ops.c +6 -0
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Good point. I was thinking that we would likely need the following construct is the future:

if (np) /* DT device */
   return xen_is_dt_grant_dma_device(dev);
else /* ACPI device */
   return xen_is_acpi_grant_dma_device(dev);

So, if we use the check directly in xen_is_dt_grant_dma_device() and it returns false we won't be able to recognize a reason then (whether dev is not related to DT, or it is related to DT but it is not a "xen,grant-dma" device).
But, I am ok to eliminate one walk right now, then we will see.

xen_is_grant_dma_device() will became the following:

bool xen_is_grant_dma_device(struct device *dev)
{
    return xen_is_dt_grant_dma_device(dev);
}

xen_is_dt_grant_dma_device() will need to gain a check that dev->of_node is not a NULL.

Shall I?

Quoted 56 lines
Quoted 56 lines
drivers/xen/grant-dma-ops.c +20 -7
  	return false;
  }
@@ -325,12 +377,19 @@ bool xen_virtio_mem_acc(struct virtio_device *dev)
  static int xen_dt_grant_init_backend_domid(struct device *dev,
  					   struct xen_grant_dma_data *data)
  {
-	struct of_phandle_args iommu_spec;
+	struct of_phandle_args iommu_spec = { .args_count = 1 };
 
-	if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
-			0, &iommu_spec)) {
-		dev_err(dev, "Cannot parse iommus property\n");
-		return -ESRCH;
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_spec.np, iommu_spec.args)) {
+			dev_err(dev, "Cannot translate ID\n");
+			return -ESRCH;
+		}
+	} else {
+		if (of_parse_phandle_with_args(dev->of_node, "iommus", "#iommu-cells",
+				0, &iommu_spec)) {
+			dev_err(dev, "Cannot parse iommus property\n");
+			return -ESRCH;
+		}
  	}
 
  	if (!of_device_is_compatible(iommu_spec.np, "xen,grant-dma") ||
@@ -354,6 +413,7 @@ static int xen_dt_grant_init_backend_domid(struct device *dev,
  void xen_grant_setup_dma_ops(struct device *dev)
  {
  	struct xen_grant_dma_data *data;
+	struct device_node *np;
 
  	data = find_xen_grant_dma_data(dev);
  	if (data) {
@@ -365,8 +425,13 @@ void xen_grant_setup_dma_ops(struct device *dev)
  	if (!data)
  		goto err;
 
-	if (dev->of_node) {
-		if (xen_dt_grant_init_backend_domid(dev, data))
+	np = xen_dt_get_node(dev);
+	if (np) {
+		int ret;
+
+		ret = xen_dt_grant_init_backend_domid(dev, data);
+		of_node_put(np);
+		if (ret)
  			goto err;
  	} else if (IS_ENABLED(CONFIG_XEN_VIRTIO_FORCE_GRANT)) {
  		dev_info(dev, "Using dom0 as backend\n");
Signature · 1 line
2.25.1
Signature · 3 lines
Regards,

Oleksandr Tyshchenko
#15 lore raw
Quoted 159 lines · Oleksandr Tyshchenko
Quoted 11 lines · Stefano Stabellini
Quoted 10 lines · Oleksandr Tyshchenko

From: Oleksandr Tyshchenko <[email protected]>

Use the same "xen-grant-dma" device concept for the PCI devices behind device-tree based PCI Host controller, but with one modification. Unlike for platform devices, we cannot use generic IOMMU bindings (iommus property), as we need to support more flexible configuration. The problem is that PCI devices under the single PCI Host controller may have the backends running in different Xen domains and thus have different endpoints ID (backend domains ID).

Quoted 7 lines
Quoted 5 lines

So use generic PCI-IOMMU bindings instead (iommu-map/iommu-map-mask properties) which allows us to describe relationship between PCI devices and backend domains ID properly.

Now that I understood the approach and the reasons for it, I can review the patch :-)

perfect, thanks.

Quoted 2 lines

Please add an example of the bindings in the commit message.

ok, will do

Quoted 38 lines
Quoted 38 lines

---
Slightly RFC. This is needed to support Xen grant mappings for virtio-pci devices on Arm at some point in the future. The Xen toolstack side is not completely ready yet. Here, for PCI devices we use more flexible way to pass backend domid to the guest than for platform devices.

Changes V1 -> V2:
- update commit description
- rebase
- rework to use generic PCI-IOMMU bindings instead of generic IOMMU bindings

Previous discussion is at:

https://urldefense.com/v3/__https://lore.kernel.org/xen-devel/[email protected]/__;!!GF_29dbcQIUBPA!xJPdZO3-3Wmgo_79HuDsD53RkH_eAN96NmwuwFE7dArt_xNYGdD6LeLsq4B_QPrrvX-x23tJl6jQlNqgyNjgzT2NE3Pqjg$  [lore[.]kernel[.]org]

Based on:

https://urldefense.com/v3/__https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=for-linus-6.1__;!!GF_29dbcQIUBPA!xJPdZO3-3Wmgo_79HuDsD53RkH_eAN96NmwuwFE7dArt_xNYGdD6LeLsq4B_QPrrvX-x23tJl6jQlNqgyNjgzT2J40LOxg$  [git[.]kernel[.]org]
---
  drivers/xen/grant-dma-ops.c | 87 ++++++++++++++++++++++++++++++++-----
  1 file changed, 76 insertions(+), 11 deletions(-)
drivers/xen/grant-dma-ops.c +1 -0
index daa525df7bdc..b79d9d6ce154 100644
@@ -10,6 +10,7 @@
  #include <linux/module.h>
  #include <linux/dma-map-ops.h>
  #include <linux/of.h>
+#include <linux/pci.h>
  #include <linux/pfn.h>
  #include <linux/xarray.h>
  #include <linux/virtio_anchor.h>
@@ -292,12 +293,55 @@ static const struct dma_map_ops xen_grant_dma_ops = {
  	.dma_supported = xen_grant_dma_supported,
  };
 
Quoted 13 lines

drivers/xen/grant-dma-ops.c static const struct dma_map_ops xen_grant_dma_ops = { · line 306

Quoted 11 lines
drivers/xen/grant-dma-ops.c +11 -0
+static struct device_node *xen_dt_get_pci_host_node(struct device *dev)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	struct pci_bus *bus = pdev->bus;
+
+	/* Walk up to the root bus to look for PCI Host controller */
+	while (!pci_is_root_bus(bus))
+		bus = bus->parent;
+
+	return of_node_get(bus->bridge->parent->of_node);
+}

It seems silly that we need to walk the hierachy that way, but I couldn't find another way to do it

I also couldn't, but is it a really problem? This code is only gets called during initialization.

Quoted 54 lines
Quoted 54 lines
drivers/xen/grant-dma-ops.c +36 -4
+static struct device_node *xen_dt_get_node(struct device *dev)
+{
+	if (dev_is_pci(dev))
+		return xen_dt_get_pci_host_node(dev);
+
+	return of_node_get(dev->of_node);
+}
+
+static int xen_dt_map_id(struct device *dev, struct device_node **iommu_np,
+			 u32 *sid)
+{
+	struct pci_dev *pdev = to_pci_dev(dev);
+	u32 rid = PCI_DEVID(pdev->bus->number, pdev->devfn);
+	struct device_node *host_np;
+	int ret;
+
+	host_np = xen_dt_get_pci_host_node(dev);
+	if (!host_np)
+		return -ENODEV;
+
+	ret = of_map_id(host_np, rid, "iommu-map", "iommu-map-mask", iommu_np, sid);
+	of_node_put(host_np);
+	return ret;
+}
+
  static bool xen_is_dt_grant_dma_device(struct device *dev)
  {
-	struct device_node *iommu_np;
+	struct device_node *iommu_np = NULL;
  	bool has_iommu;
 
-	iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+	if (dev_is_pci(dev)) {
+		if (xen_dt_map_id(dev, &iommu_np, NULL))
+			return false;
+	} else
+		iommu_np = of_parse_phandle(dev->of_node, "iommus", 0);
+
  	has_iommu = iommu_np &&
  		    of_device_is_compatible(iommu_np, "xen,grant-dma");
  	of_node_put(iommu_np);
@@ -307,9 +351,17 @@ static bool xen_is_dt_grant_dma_device(struct device *dev)
 
  bool xen_is_grant_dma_device(struct device *dev)
  {
+	struct device_node *np;
+
  	/* XXX Handle only DT devices for now */
-	if (dev->of_node)
-		return xen_is_dt_grant_dma_device(dev);
+	np = xen_dt_get_node(dev);
+	if (np) {
Quoted 8 lines

drivers/xen/grant-dma-ops.c static bool xen_is_dt_grant_dma_device(struct device *dev) · line 364

Quoted 6 lines
drivers/xen/grant-dma-ops.c +6 -0
+		bool ret;
+
+		ret = xen_is_dt_grant_dma_device(dev);
+		of_node_put(np);
+		return ret;
+	}

We don't need to walk the PCI hierachy twice. Maybe we can add the of_node check directly to xen_is_dt_grant_dma_device?

Good point. I was thinking that we would likely need the following construct is the future:

if (np) /* DT device */
   return xen_is_dt_grant_dma_device(dev);
else /* ACPI device */
   return xen_is_acpi_grant_dma_device(dev);

Quoted 4 lines

So, if we use the check directly in xen_is_dt_grant_dma_device() and it returns false we won't be able to recognize a reason then (whether dev is not related to DT, or it is related to DT but it is not a "xen,grant-dma" device).

That problem can easily be solved by having xen_is_dt_grant_dma_device return more than 2 possible values. It could return an int for example, with 3 valid values.

Quoted 14 lines

But, I am ok to eliminate one walk right now, then we will see.

xen_is_grant_dma_device() will became the following:

bool xen_is_grant_dma_device(struct device *dev)
{
    return xen_is_dt_grant_dma_device(dev);
}

xen_is_dt_grant_dma_device() will need to gain a check that dev->of_node is not a NULL.

Shall I?