From: Oleksandr Andrushchenko <[email protected]>
Xen-pciback driver was designed to be built for x86 only. But it
can also be used by other architectures, e.g. Arm.
Re-structure the driver in a way that it can be built for other
platforms as well.
Signed-off-by: Oleksandr Andrushchenko <[email protected]>
Signed-off-by: Anastasiia Lukianenko <[email protected]>
---
Tested on Arm and x86.
---
arch/x86/include/asm/xen/pci.h | 24 ----------
arch/x86/pci/xen.c | 74 +----------------------------
drivers/xen/Kconfig | 2 +-
drivers/xen/events/events_base.c | 1 +
drivers/xen/pci.c | 75 ++++++++++++++++++++++++++++++
drivers/xen/xen-pciback/pci_stub.c | 3 +-
drivers/xen/xen-pciback/xenbus.c | 2 +-
include/xen/pci.h | 34 ++++++++++++++
8 files changed, 115 insertions(+), 100 deletions(-)
create mode 100644 include/xen/pci.h
diff --git a/arch/x86/include/asm/xen/pci.h b/arch/x86/include/asm/xen/pci.h
index 3506d8c598c1..9ff7b49bca08 100644
--- a/arch/x86/include/asm/xen/pci.h
+++ b/arch/x86/include/asm/xen/pci.h
@@ -14,30 +14,6 @@ static inline int pci_xen_hvm_init(void)
return -1;
}
#endif
-#if defined(CONFIG_XEN_DOM0)
-int __init pci_xen_initial_domain(void);
-int xen_find_device_domain_owner(struct pci_dev *dev);
-int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
-int xen_unregister_device_domain_owner(struct pci_dev *dev);
-#else
-static inline int __init pci_xen_initial_domain(void)
-{
- return -1;
-}
-static inline int xen_find_device_domain_owner(struct pci_dev *dev)
-{
- return -1;
-}
-static inline int xen_register_device_domain_owner(struct pci_dev *dev,
- uint16_t domain)
-{
- return -1;
-}
-static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
-{
- return -1;
-}
-#endif
#if defined(CONFIG_PCI_MSI)
#if defined(CONFIG_PCI_XEN)
diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
index 3d41a09c2c14..4a45b0bf9ae4 100644
--- a/arch/x86/pci/xen.c
+++ b/arch/x86/pci/xen.c
@@ -23,6 +23,7 @@
#include <xen/features.h>
#include <xen/events.h>
+#include <xen/pci.h>
#include <asm/xen/pci.h>
#include <asm/xen/cpuid.h>
#include <asm/apic.h>
@@ -583,77 +584,4 @@ int __init pci_xen_initial_domain(void)
}
return 0;
}
-
-struct xen_device_domain_owner {
- domid_t domain;
- struct pci_dev *dev;
- struct list_head list;
-};
-
-static DEFINE_SPINLOCK(dev_domain_list_spinlock);
-static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
-
-static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
-{
- struct xen_device_domain_owner *owner;
-
- list_for_each_entry(owner, &dev_domain_list, list) {
- if (owner->dev == dev)
- return owner;
- }
- return NULL;
-}
-
-int xen_find_device_domain_owner(struct pci_dev *dev)
-{
- struct xen_device_domain_owner *owner;
- int domain = -ENODEV;
-
- spin_lock(&dev_domain_list_spinlock);
- owner = find_device(dev);
- if (owner)
- domain = owner->domain;
- spin_unlock(&dev_domain_list_spinlock);
- return domain;
-}
-EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
-
-int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
-{
- struct xen_device_domain_owner *owner;
-
- owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
- if (!owner)
- return -ENODEV;
-
- spin_lock(&dev_domain_list_spinlock);
- if (find_device(dev)) {
- spin_unlock(&dev_domain_list_spinlock);
- kfree(owner);
- return -EEXIST;
- }
- owner->domain = domain;
- owner->dev = dev;
- list_add_tail(&owner->list, &dev_domain_list);
- spin_unlock(&dev_domain_list_spinlock);
- return 0;
-}
-EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
-
-int xen_unregister_device_domain_owner(struct pci_dev *dev)
-{
- struct xen_device_domain_owner *owner;
-
- spin_lock(&dev_domain_list_spinlock);
- owner = find_device(dev);
- if (!owner) {
- spin_unlock(&dev_domain_list_spinlock);
- return -ENODEV;
- }
- list_del(&owner->list);
- spin_unlock(&dev_domain_list_spinlock);
- kfree(owner);
- return 0;
-}
-EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
#endif
diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
index a37eb52fb401..057ddf61ef61 100644
--- a/drivers/xen/Kconfig
+++ b/drivers/xen/Kconfig
@@ -182,7 +182,7 @@ config SWIOTLB_XEN
config XEN_PCIDEV_BACKEND
tristate "Xen PCI-device backend driver"
- depends on PCI && X86 && XEN
+ depends on PCI && XEN
depends on XEN_BACKEND
default m
help
diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c
index a78704ae3618..35493ff0d146 100644
--- a/drivers/xen/events/events_base.c
+++ b/drivers/xen/events/events_base.c
@@ -65,6 +65,7 @@
#include <xen/interface/vcpu.h>
#include <xen/xenbus.h>
#include <asm/hw_irq.h>
+#include <xen/pci.h>
#include "events_internal.h"
diff --git a/drivers/xen/pci.c b/drivers/xen/pci.c
index 224df03ce42e..fc8c1249d49f 100644
--- a/drivers/xen/pci.c
+++ b/drivers/xen/pci.c
@@ -254,3 +254,78 @@ static int xen_mcfg_late(void)
return 0;
}
#endif
+
+#ifdef CONFIG_XEN_DOM0
+struct xen_device_domain_owner {
+ domid_t domain;
+ struct pci_dev *dev;
+ struct list_head list;
+};
+
+static DEFINE_SPINLOCK(dev_domain_list_spinlock);
+static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
+
+static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
+{
+ struct xen_device_domain_owner *owner;
+
+ list_for_each_entry(owner, &dev_domain_list, list) {
+ if (owner->dev == dev)
+ return owner;
+ }
+ return NULL;
+}
+
+int xen_find_device_domain_owner(struct pci_dev *dev)
+{
+ struct xen_device_domain_owner *owner;
+ int domain = -ENODEV;
+
+ spin_lock(&dev_domain_list_spinlock);
+ owner = find_device(dev);
+ if (owner)
+ domain = owner->domain;
+ spin_unlock(&dev_domain_list_spinlock);
+ return domain;
+}
+EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
+
+int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
+{
+ struct xen_device_domain_owner *owner;
+
+ owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
+ if (!owner)
+ return -ENODEV;
+
+ spin_lock(&dev_domain_list_spinlock);
+ if (find_device(dev)) {
+ spin_unlock(&dev_domain_list_spinlock);
+ kfree(owner);
+ return -EEXIST;
+ }
+ owner->domain = domain;
+ owner->dev = dev;
+ list_add_tail(&owner->list, &dev_domain_list);
+ spin_unlock(&dev_domain_list_spinlock);
+ return 0;
+}
+EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
+
+int xen_unregister_device_domain_owner(struct pci_dev *dev)
+{
+ struct xen_device_domain_owner *owner;
+
+ spin_lock(&dev_domain_list_spinlock);
+ owner = find_device(dev);
+ if (!owner) {
+ spin_unlock(&dev_domain_list_spinlock);
+ return -ENODEV;
+ }
+ list_del(&owner->list);
+ spin_unlock(&dev_domain_list_spinlock);
+ kfree(owner);
+ return 0;
+}
+EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
+#endif
diff --git a/drivers/xen/xen-pciback/pci_stub.c b/drivers/xen/xen-pciback/pci_stub.c
index f8e4faa96ad6..bba527620507 100644
--- a/drivers/xen/xen-pciback/pci_stub.c
+++ b/drivers/xen/xen-pciback/pci_stub.c
@@ -19,7 +19,8 @@
#include <linux/sched.h>
#include <linux/atomic.h>
#include <xen/events.h>
-#include <asm/xen/pci.h>
+#include <xen/pci.h>
+#include <xen/xen.h>
#include <asm/xen/hypervisor.h>
#include <xen/interface/physdev.h>
#include "pciback.h"
diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
index c09c7ebd6968..da34ce85dc88 100644
--- a/drivers/xen/xen-pciback/xenbus.c
+++ b/drivers/xen/xen-pciback/xenbus.c
@@ -14,7 +14,7 @@
#include <linux/workqueue.h>
#include <xen/xenbus.h>
#include <xen/events.h>
-#include <asm/xen/pci.h>
+#include <xen/pci.h>
#include "pciback.h"
#define INVALID_EVTCHN_IRQ (-1)
diff --git a/include/xen/pci.h b/include/xen/pci.h
new file mode 100644
index 000000000000..5c988d5ff38f
--- /dev/null
+++ b/include/xen/pci.h
@@ -0,0 +1,34 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef __XEN_PCI_H__
+#define __XEN_PCI_H__
+
+#if defined(CONFIG_XEN_DOM0)
+int __init pci_xen_initial_domain(void);
+int xen_find_device_domain_owner(struct pci_dev *dev);
+int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
+int xen_unregister_device_domain_owner(struct pci_dev *dev);
+#else
+static inline int __init pci_xen_initial_domain(void)
+{
+ return -1;
+}
+
+static inline int xen_find_device_domain_owner(struct pci_dev *dev)
+{
+ return -1;
+}
+
+static inline int xen_register_device_domain_owner(struct pci_dev *dev,
+ uint16_t domain)
+{
+ return -1;
+}
+
+static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
+{
+ return -1;
+}
+#endif
+
+#endif
--
2.25.1
Hi Oleksandr,
Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
device in Dom0 so that it can be safely assigned to a DomU?
I am asking because actually I don't think we want to enable the PV PCI
backend feature of pciback on ARM, right? That would clash with the PCI
assignment work you have been doing in Xen. They couldn't both work at
the same time.
If we only need pciback to "park" a device in Dom0, wouldn't it be
possible and better to use pci-stub instead?
On Fri, 17 Sep 2021, Oleksandr Andrushchenko wrote:
> From: Oleksandr Andrushchenko <[email protected]>
>
> Xen-pciback driver was designed to be built for x86 only. But it
> can also be used by other architectures, e.g. Arm.
> Re-structure the driver in a way that it can be built for other
> platforms as well.
>
> Signed-off-by: Oleksandr Andrushchenko <[email protected]>
> Signed-off-by: Anastasiia Lukianenko <[email protected]>
>
> ---
> Tested on Arm and x86.
> ---
> arch/x86/include/asm/xen/pci.h | 24 ----------
> arch/x86/pci/xen.c | 74 +----------------------------
> drivers/xen/Kconfig | 2 +-
> drivers/xen/events/events_base.c | 1 +
> drivers/xen/pci.c | 75 ++++++++++++++++++++++++++++++
> drivers/xen/xen-pciback/pci_stub.c | 3 +-
> drivers/xen/xen-pciback/xenbus.c | 2 +-
> include/xen/pci.h | 34 ++++++++++++++
> 8 files changed, 115 insertions(+), 100 deletions(-)
> create mode 100644 include/xen/pci.h
>
> diff --git a/arch/x86/include/asm/xen/pci.h b/arch/x86/include/asm/xen/pci.h
> index 3506d8c598c1..9ff7b49bca08 100644
> --- a/arch/x86/include/asm/xen/pci.h
> +++ b/arch/x86/include/asm/xen/pci.h
> @@ -14,30 +14,6 @@ static inline int pci_xen_hvm_init(void)
> return -1;
> }
> #endif
> -#if defined(CONFIG_XEN_DOM0)
> -int __init pci_xen_initial_domain(void);
> -int xen_find_device_domain_owner(struct pci_dev *dev);
> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
> -int xen_unregister_device_domain_owner(struct pci_dev *dev);
> -#else
> -static inline int __init pci_xen_initial_domain(void)
> -{
> - return -1;
> -}
> -static inline int xen_find_device_domain_owner(struct pci_dev *dev)
> -{
> - return -1;
> -}
> -static inline int xen_register_device_domain_owner(struct pci_dev *dev,
> - uint16_t domain)
> -{
> - return -1;
> -}
> -static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
> -{
> - return -1;
> -}
> -#endif
>
> #if defined(CONFIG_PCI_MSI)
> #if defined(CONFIG_PCI_XEN)
> diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
> index 3d41a09c2c14..4a45b0bf9ae4 100644
> --- a/arch/x86/pci/xen.c
> +++ b/arch/x86/pci/xen.c
> @@ -23,6 +23,7 @@
>
> #include <xen/features.h>
> #include <xen/events.h>
> +#include <xen/pci.h>
> #include <asm/xen/pci.h>
> #include <asm/xen/cpuid.h>
> #include <asm/apic.h>
> @@ -583,77 +584,4 @@ int __init pci_xen_initial_domain(void)
> }
> return 0;
> }
> -
> -struct xen_device_domain_owner {
> - domid_t domain;
> - struct pci_dev *dev;
> - struct list_head list;
> -};
> -
> -static DEFINE_SPINLOCK(dev_domain_list_spinlock);
> -static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
> -
> -static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - list_for_each_entry(owner, &dev_domain_list, list) {
> - if (owner->dev == dev)
> - return owner;
> - }
> - return NULL;
> -}
> -
> -int xen_find_device_domain_owner(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> - int domain = -ENODEV;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - owner = find_device(dev);
> - if (owner)
> - domain = owner->domain;
> - spin_unlock(&dev_domain_list_spinlock);
> - return domain;
> -}
> -EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
> -
> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
> - if (!owner)
> - return -ENODEV;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - if (find_device(dev)) {
> - spin_unlock(&dev_domain_list_spinlock);
> - kfree(owner);
> - return -EEXIST;
> - }
> - owner->domain = domain;
> - owner->dev = dev;
> - list_add_tail(&owner->list, &dev_domain_list);
> - spin_unlock(&dev_domain_list_spinlock);
> - return 0;
> -}
> -EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
> -
> -int xen_unregister_device_domain_owner(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - owner = find_device(dev);
> - if (!owner) {
> - spin_unlock(&dev_domain_list_spinlock);
> - return -ENODEV;
> - }
> - list_del(&owner->list);
> - spin_unlock(&dev_domain_list_spinlock);
> - kfree(owner);
> - return 0;
> -}
> -EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
> #endif
> diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
> index a37eb52fb401..057ddf61ef61 100644
> --- a/drivers/xen/Kconfig
> +++ b/drivers/xen/Kconfig
> @@ -182,7 +182,7 @@ config SWIOTLB_XEN
>
> config XEN_PCIDEV_BACKEND
> tristate "Xen PCI-device backend driver"
> - depends on PCI && X86 && XEN
> + depends on PCI && XEN
> depends on XEN_BACKEND
> default m
> help
> diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c
> index a78704ae3618..35493ff0d146 100644
> --- a/drivers/xen/events/events_base.c
> +++ b/drivers/xen/events/events_base.c
> @@ -65,6 +65,7 @@
> #include <xen/interface/vcpu.h>
> #include <xen/xenbus.h>
> #include <asm/hw_irq.h>
> +#include <xen/pci.h>
>
> #include "events_internal.h"
>
> diff --git a/drivers/xen/pci.c b/drivers/xen/pci.c
> index 224df03ce42e..fc8c1249d49f 100644
> --- a/drivers/xen/pci.c
> +++ b/drivers/xen/pci.c
> @@ -254,3 +254,78 @@ static int xen_mcfg_late(void)
> return 0;
> }
> #endif
> +
> +#ifdef CONFIG_XEN_DOM0
> +struct xen_device_domain_owner {
> + domid_t domain;
> + struct pci_dev *dev;
> + struct list_head list;
> +};
> +
> +static DEFINE_SPINLOCK(dev_domain_list_spinlock);
> +static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
> +
> +static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
> +{
> + struct xen_device_domain_owner *owner;
> +
> + list_for_each_entry(owner, &dev_domain_list, list) {
> + if (owner->dev == dev)
> + return owner;
> + }
> + return NULL;
> +}
> +
> +int xen_find_device_domain_owner(struct pci_dev *dev)
> +{
> + struct xen_device_domain_owner *owner;
> + int domain = -ENODEV;
> +
> + spin_lock(&dev_domain_list_spinlock);
> + owner = find_device(dev);
> + if (owner)
> + domain = owner->domain;
> + spin_unlock(&dev_domain_list_spinlock);
> + return domain;
> +}
> +EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
> +
> +int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
> +{
> + struct xen_device_domain_owner *owner;
> +
> + owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
> + if (!owner)
> + return -ENODEV;
> +
> + spin_lock(&dev_domain_list_spinlock);
> + if (find_device(dev)) {
> + spin_unlock(&dev_domain_list_spinlock);
> + kfree(owner);
> + return -EEXIST;
> + }
> + owner->domain = domain;
> + owner->dev = dev;
> + list_add_tail(&owner->list, &dev_domain_list);
> + spin_unlock(&dev_domain_list_spinlock);
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
> +
> +int xen_unregister_device_domain_owner(struct pci_dev *dev)
> +{
> + struct xen_device_domain_owner *owner;
> +
> + spin_lock(&dev_domain_list_spinlock);
> + owner = find_device(dev);
> + if (!owner) {
> + spin_unlock(&dev_domain_list_spinlock);
> + return -ENODEV;
> + }
> + list_del(&owner->list);
> + spin_unlock(&dev_domain_list_spinlock);
> + kfree(owner);
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
> +#endif
> diff --git a/drivers/xen/xen-pciback/pci_stub.c b/drivers/xen/xen-pciback/pci_stub.c
> index f8e4faa96ad6..bba527620507 100644
> --- a/drivers/xen/xen-pciback/pci_stub.c
> +++ b/drivers/xen/xen-pciback/pci_stub.c
> @@ -19,7 +19,8 @@
> #include <linux/sched.h>
> #include <linux/atomic.h>
> #include <xen/events.h>
> -#include <asm/xen/pci.h>
> +#include <xen/pci.h>
> +#include <xen/xen.h>
> #include <asm/xen/hypervisor.h>
> #include <xen/interface/physdev.h>
> #include "pciback.h"
> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
> index c09c7ebd6968..da34ce85dc88 100644
> --- a/drivers/xen/xen-pciback/xenbus.c
> +++ b/drivers/xen/xen-pciback/xenbus.c
> @@ -14,7 +14,7 @@
> #include <linux/workqueue.h>
> #include <xen/xenbus.h>
> #include <xen/events.h>
> -#include <asm/xen/pci.h>
> +#include <xen/pci.h>
> #include "pciback.h"
>
> #define INVALID_EVTCHN_IRQ (-1)
> diff --git a/include/xen/pci.h b/include/xen/pci.h
> new file mode 100644
> index 000000000000..5c988d5ff38f
> --- /dev/null
> +++ b/include/xen/pci.h
> @@ -0,0 +1,34 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +
> +#ifndef __XEN_PCI_H__
> +#define __XEN_PCI_H__
> +
> +#if defined(CONFIG_XEN_DOM0)
> +int __init pci_xen_initial_domain(void);
> +int xen_find_device_domain_owner(struct pci_dev *dev);
> +int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
> +int xen_unregister_device_domain_owner(struct pci_dev *dev);
> +#else
> +static inline int __init pci_xen_initial_domain(void)
> +{
> + return -1;
> +}
> +
> +static inline int xen_find_device_domain_owner(struct pci_dev *dev)
> +{
> + return -1;
> +}
> +
> +static inline int xen_register_device_domain_owner(struct pci_dev *dev,
> + uint16_t domain)
> +{
> + return -1;
> +}
> +
> +static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
> +{
> + return -1;
> +}
> +#endif
> +
> +#endif
> --
> 2.25.1
>
Hello, Stefano!
On 18.09.21 00:45, Stefano Stabellini wrote:
> Hi Oleksandr,
>
> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
> device in Dom0 so that it can be safely assigned to a DomU?
Not only that
>
> I am asking because actually I don't think we want to enable the PV PCI
> backend feature of pciback on ARM, right? That would clash with the PCI
> assignment work you have been doing in Xen. They couldn't both work at
> the same time.
Correct, it is not used
>
> If we only need pciback to "park" a device in Dom0, wouldn't it be
> possible and better to use pci-stub instead?
Not only that, so pci-stub is not enough
The functionality which is implemented by the pciback and the toolstack
and which is relevant/missing/needed for ARM:
1. pciback is used as a database for assignable PCI devices, e.g. xl
pci-assignable-{add|remove|list} manipulates that list. So, whenever the
toolstack needs to know which PCI devices can be passed through it reads
that from the relevant sysfs entries of the pciback.
2. pciback is used to hold the unbound PCI devices, e.g. when passing through
a PCI device it needs to be unbound from the relevant device driver and bound
to pciback (strictly speaking it is not required that the device is bound to
pciback, but pciback is again used as a database of the passed through PCI
devices, so we can re-bind the devices back to their original drivers when
guest domain shuts down)
3. Device reset
We have previously discussed on xen-devel ML possible solutions to that as from the
above we see that pciback functionality is going to be only partially used on Arm.
Please see [1] and [2]:
1. It is not acceptable to manage the assignable list in Xen itself
2. pciback can be split into two parts: PCI assignable/bind/reset handling and
the rest like vPCI etc.
3. pcifront is not used on Arm
So, limited use of the pciback is one of the bricks used to enable PCI passthrough
on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
all the goals above.
If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
parts then it can be done, yet the patch is going to be the very first brick in that building.
So, I think this patch is still going to be needed besides which direction we take.
Hope this sheds some light,
Oleksandr
>
>
> On Fri, 17 Sep 2021, Oleksandr Andrushchenko wrote:
>> From: Oleksandr Andrushchenko <[email protected]>
>>
>> Xen-pciback driver was designed to be built for x86 only. But it
>> can also be used by other architectures, e.g. Arm.
>> Re-structure the driver in a way that it can be built for other
>> platforms as well.
>>
>> Signed-off-by: Oleksandr Andrushchenko <[email protected]>
>> Signed-off-by: Anastasiia Lukianenko <[email protected]>
>>
>> ---
>> Tested on Arm and x86.
>> ---
>> arch/x86/include/asm/xen/pci.h | 24 ----------
>> arch/x86/pci/xen.c | 74 +----------------------------
>> drivers/xen/Kconfig | 2 +-
>> drivers/xen/events/events_base.c | 1 +
>> drivers/xen/pci.c | 75 ++++++++++++++++++++++++++++++
>> drivers/xen/xen-pciback/pci_stub.c | 3 +-
>> drivers/xen/xen-pciback/xenbus.c | 2 +-
>> include/xen/pci.h | 34 ++++++++++++++
>> 8 files changed, 115 insertions(+), 100 deletions(-)
>> create mode 100644 include/xen/pci.h
>>
>> diff --git a/arch/x86/include/asm/xen/pci.h b/arch/x86/include/asm/xen/pci.h
>> index 3506d8c598c1..9ff7b49bca08 100644
>> --- a/arch/x86/include/asm/xen/pci.h
>> +++ b/arch/x86/include/asm/xen/pci.h
>> @@ -14,30 +14,6 @@ static inline int pci_xen_hvm_init(void)
>> return -1;
>> }
>> #endif
>> -#if defined(CONFIG_XEN_DOM0)
>> -int __init pci_xen_initial_domain(void);
>> -int xen_find_device_domain_owner(struct pci_dev *dev);
>> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
>> -int xen_unregister_device_domain_owner(struct pci_dev *dev);
>> -#else
>> -static inline int __init pci_xen_initial_domain(void)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_find_device_domain_owner(struct pci_dev *dev)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_register_device_domain_owner(struct pci_dev *dev,
>> - uint16_t domain)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> -{
>> - return -1;
>> -}
>> -#endif
>>
>> #if defined(CONFIG_PCI_MSI)
>> #if defined(CONFIG_PCI_XEN)
>> diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
>> index 3d41a09c2c14..4a45b0bf9ae4 100644
>> --- a/arch/x86/pci/xen.c
>> +++ b/arch/x86/pci/xen.c
>> @@ -23,6 +23,7 @@
>>
>> #include <xen/features.h>
>> #include <xen/events.h>
>> +#include <xen/pci.h>
>> #include <asm/xen/pci.h>
>> #include <asm/xen/cpuid.h>
>> #include <asm/apic.h>
>> @@ -583,77 +584,4 @@ int __init pci_xen_initial_domain(void)
>> }
>> return 0;
>> }
>> -
>> -struct xen_device_domain_owner {
>> - domid_t domain;
>> - struct pci_dev *dev;
>> - struct list_head list;
>> -};
>> -
>> -static DEFINE_SPINLOCK(dev_domain_list_spinlock);
>> -static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
>> -
>> -static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - list_for_each_entry(owner, &dev_domain_list, list) {
>> - if (owner->dev == dev)
>> - return owner;
>> - }
>> - return NULL;
>> -}
>> -
>> -int xen_find_device_domain_owner(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> - int domain = -ENODEV;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - owner = find_device(dev);
>> - if (owner)
>> - domain = owner->domain;
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return domain;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
>> -
>> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
>> - if (!owner)
>> - return -ENODEV;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - if (find_device(dev)) {
>> - spin_unlock(&dev_domain_list_spinlock);
>> - kfree(owner);
>> - return -EEXIST;
>> - }
>> - owner->domain = domain;
>> - owner->dev = dev;
>> - list_add_tail(&owner->list, &dev_domain_list);
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return 0;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
>> -
>> -int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - owner = find_device(dev);
>> - if (!owner) {
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return -ENODEV;
>> - }
>> - list_del(&owner->list);
>> - spin_unlock(&dev_domain_list_spinlock);
>> - kfree(owner);
>> - return 0;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
>> #endif
>> diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
>> index a37eb52fb401..057ddf61ef61 100644
>> --- a/drivers/xen/Kconfig
>> +++ b/drivers/xen/Kconfig
>> @@ -182,7 +182,7 @@ config SWIOTLB_XEN
>>
>> config XEN_PCIDEV_BACKEND
>> tristate "Xen PCI-device backend driver"
>> - depends on PCI && X86 && XEN
>> + depends on PCI && XEN
>> depends on XEN_BACKEND
>> default m
>> help
>> diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c
>> index a78704ae3618..35493ff0d146 100644
>> --- a/drivers/xen/events/events_base.c
>> +++ b/drivers/xen/events/events_base.c
>> @@ -65,6 +65,7 @@
>> #include <xen/interface/vcpu.h>
>> #include <xen/xenbus.h>
>> #include <asm/hw_irq.h>
>> +#include <xen/pci.h>
>>
>> #include "events_internal.h"
>>
>> diff --git a/drivers/xen/pci.c b/drivers/xen/pci.c
>> index 224df03ce42e..fc8c1249d49f 100644
>> --- a/drivers/xen/pci.c
>> +++ b/drivers/xen/pci.c
>> @@ -254,3 +254,78 @@ static int xen_mcfg_late(void)
>> return 0;
>> }
>> #endif
>> +
>> +#ifdef CONFIG_XEN_DOM0
>> +struct xen_device_domain_owner {
>> + domid_t domain;
>> + struct pci_dev *dev;
>> + struct list_head list;
>> +};
>> +
>> +static DEFINE_SPINLOCK(dev_domain_list_spinlock);
>> +static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
>> +
>> +static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
>> +{
>> + struct xen_device_domain_owner *owner;
>> +
>> + list_for_each_entry(owner, &dev_domain_list, list) {
>> + if (owner->dev == dev)
>> + return owner;
>> + }
>> + return NULL;
>> +}
>> +
>> +int xen_find_device_domain_owner(struct pci_dev *dev)
>> +{
>> + struct xen_device_domain_owner *owner;
>> + int domain = -ENODEV;
>> +
>> + spin_lock(&dev_domain_list_spinlock);
>> + owner = find_device(dev);
>> + if (owner)
>> + domain = owner->domain;
>> + spin_unlock(&dev_domain_list_spinlock);
>> + return domain;
>> +}
>> +EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
>> +
>> +int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
>> +{
>> + struct xen_device_domain_owner *owner;
>> +
>> + owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
>> + if (!owner)
>> + return -ENODEV;
>> +
>> + spin_lock(&dev_domain_list_spinlock);
>> + if (find_device(dev)) {
>> + spin_unlock(&dev_domain_list_spinlock);
>> + kfree(owner);
>> + return -EEXIST;
>> + }
>> + owner->domain = domain;
>> + owner->dev = dev;
>> + list_add_tail(&owner->list, &dev_domain_list);
>> + spin_unlock(&dev_domain_list_spinlock);
>> + return 0;
>> +}
>> +EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
>> +
>> +int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> +{
>> + struct xen_device_domain_owner *owner;
>> +
>> + spin_lock(&dev_domain_list_spinlock);
>> + owner = find_device(dev);
>> + if (!owner) {
>> + spin_unlock(&dev_domain_list_spinlock);
>> + return -ENODEV;
>> + }
>> + list_del(&owner->list);
>> + spin_unlock(&dev_domain_list_spinlock);
>> + kfree(owner);
>> + return 0;
>> +}
>> +EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
>> +#endif
>> diff --git a/drivers/xen/xen-pciback/pci_stub.c b/drivers/xen/xen-pciback/pci_stub.c
>> index f8e4faa96ad6..bba527620507 100644
>> --- a/drivers/xen/xen-pciback/pci_stub.c
>> +++ b/drivers/xen/xen-pciback/pci_stub.c
>> @@ -19,7 +19,8 @@
>> #include <linux/sched.h>
>> #include <linux/atomic.h>
>> #include <xen/events.h>
>> -#include <asm/xen/pci.h>
>> +#include <xen/pci.h>
>> +#include <xen/xen.h>
>> #include <asm/xen/hypervisor.h>
>> #include <xen/interface/physdev.h>
>> #include "pciback.h"
>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>> index c09c7ebd6968..da34ce85dc88 100644
>> --- a/drivers/xen/xen-pciback/xenbus.c
>> +++ b/drivers/xen/xen-pciback/xenbus.c
>> @@ -14,7 +14,7 @@
>> #include <linux/workqueue.h>
>> #include <xen/xenbus.h>
>> #include <xen/events.h>
>> -#include <asm/xen/pci.h>
>> +#include <xen/pci.h>
>> #include "pciback.h"
>>
>> #define INVALID_EVTCHN_IRQ (-1)
>> diff --git a/include/xen/pci.h b/include/xen/pci.h
>> new file mode 100644
>> index 000000000000..5c988d5ff38f
>> --- /dev/null
>> +++ b/include/xen/pci.h
>> @@ -0,0 +1,34 @@
>> +/* SPDX-License-Identifier: GPL-2.0 */
>> +
>> +#ifndef __XEN_PCI_H__
>> +#define __XEN_PCI_H__
>> +
>> +#if defined(CONFIG_XEN_DOM0)
>> +int __init pci_xen_initial_domain(void);
>> +int xen_find_device_domain_owner(struct pci_dev *dev);
>> +int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
>> +int xen_unregister_device_domain_owner(struct pci_dev *dev);
>> +#else
>> +static inline int __init pci_xen_initial_domain(void)
>> +{
>> + return -1;
>> +}
>> +
>> +static inline int xen_find_device_domain_owner(struct pci_dev *dev)
>> +{
>> + return -1;
>> +}
>> +
>> +static inline int xen_register_device_domain_owner(struct pci_dev *dev,
>> + uint16_t domain)
>> +{
>> + return -1;
>> +}
>> +
>> +static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> +{
>> + return -1;
>> +}
>> +#endif
>> +
>> +#endif
>> --
>> 2.25.1
>>
[1] https://www.mail-archive.com/[email protected]/msg84452.html
[2] https://patchwork.kernel.org/project/xen-devel/patch/[email protected]/
On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
> Hello, Stefano!
>
> On 18.09.21 00:45, Stefano Stabellini wrote:
>> Hi Oleksandr,
>>
>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>> device in Dom0 so that it can be safely assigned to a DomU?
> Not only that
>>
>> I am asking because actually I don't think we want to enable the PV PCI
>> backend feature of pciback on ARM, right? That would clash with the PCI
>> assignment work you have been doing in Xen. They couldn't both work at
>> the same time.
> Correct, it is not used
>>
>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>> possible and better to use pci-stub instead?
>
> Not only that, so pci-stub is not enough
>
> The functionality which is implemented by the pciback and the toolstack
> and which is relevant/missing/needed for ARM:
>
> 1. pciback is used as a database for assignable PCI devices, e.g. xl
> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
> toolstack needs to know which PCI devices can be passed through it reads
> that from the relevant sysfs entries of the pciback.
>
> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
> a PCI device it needs to be unbound from the relevant device driver and bound
> to pciback (strictly speaking it is not required that the device is bound to
> pciback, but pciback is again used as a database of the passed through PCI
> devices, so we can re-bind the devices back to their original drivers when
> guest domain shuts down)
>
> 3. Device reset
>
> We have previously discussed on xen-devel ML possible solutions to that as from the
> above we see that pciback functionality is going to be only partially used on Arm.
>
> Please see [1] and [2]:
>
> 1. It is not acceptable to manage the assignable list in Xen itself
>
> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
> the rest like vPCI etc.
>
> 3. pcifront is not used on Arm
It is neither in x86 PVH/HVM guests.
> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
> all the goals above.
>
> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
> parts then it can be done, yet the patch is going to be the very first brick in that building.
Doing this split should be done, as the pcifront specific part could be
omitted on x86, too, in case no PV guests using PCI passthrough have to
be supported.
> So, I think this patch is still going to be needed besides which direction we take.
Some kind of this patch, yes. It might look different in case the split
is done first.
I don't mind doing it in either sequence.
Juergen
Hello, Juergen!
On 20.09.21 14:30, Juergen Gross wrote:
> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>> Hello, Stefano!
>>
>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>> Hi Oleksandr,
>>>
>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>> device in Dom0 so that it can be safely assigned to a DomU?
>> Not only that
>>>
>>> I am asking because actually I don't think we want to enable the PV PCI
>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>> assignment work you have been doing in Xen. They couldn't both work at
>>> the same time.
>> Correct, it is not used
>>>
>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>> possible and better to use pci-stub instead?
>>
>> Not only that, so pci-stub is not enough
>>
>> The functionality which is implemented by the pciback and the toolstack
>> and which is relevant/missing/needed for ARM:
>>
>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>> toolstack needs to know which PCI devices can be passed through it reads
>> that from the relevant sysfs entries of the pciback.
>>
>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>> a PCI device it needs to be unbound from the relevant device driver and bound
>> to pciback (strictly speaking it is not required that the device is bound to
>> pciback, but pciback is again used as a database of the passed through PCI
>> devices, so we can re-bind the devices back to their original drivers when
>> guest domain shuts down)
>>
>> 3. Device reset
>>
>> We have previously discussed on xen-devel ML possible solutions to that as from the
>> above we see that pciback functionality is going to be only partially used on Arm.
>>
>> Please see [1] and [2]:
>>
>> 1. It is not acceptable to manage the assignable list in Xen itself
>>
>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>> the rest like vPCI etc.
>>
>> 3. pcifront is not used on Arm
>
> It is neither in x86 PVH/HVM guests.
Didn't know that, thank you for pointing
>
>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>> all the goals above.
>>
>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>
> Doing this split should be done, as the pcifront specific part could be
> omitted on x86, too, in case no PV guests using PCI passthrough have to
> be supported.
Agree, that the final solution should have the driver split
>
>> So, I think this patch is still going to be needed besides which direction we take.
>
> Some kind of this patch, yes. It might look different in case the split
> is done first.
>
> I don't mind doing it in either sequence.
>
With this patch we have Arm on the same page as the above mentioned x86 guests,
e.g. the driver has unused code, but yet allows Arm to function now.
At this stage of PCI passthrough on Arm it is yet enough. Long term, when
the driver gets split, Arm will benefit from that split too, but unfortunately I do not
have enough bandwidth for that piece of work at the moment.
>
> Juergen
Thank you,
Oleksandr
On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
> On 20.09.21 14:30, Juergen Gross wrote:
> > On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
> >> Hello, Stefano!
> >>
> >> On 18.09.21 00:45, Stefano Stabellini wrote:
> >>> Hi Oleksandr,
> >>>
> >>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
> >>> device in Dom0 so that it can be safely assigned to a DomU?
> >> Not only that
> >>>
> >>> I am asking because actually I don't think we want to enable the PV PCI
> >>> backend feature of pciback on ARM, right? That would clash with the PCI
> >>> assignment work you have been doing in Xen. They couldn't both work at
> >>> the same time.
> >> Correct, it is not used
> >>>
> >>> If we only need pciback to "park" a device in Dom0, wouldn't it be
> >>> possible and better to use pci-stub instead?
> >>
> >> Not only that, so pci-stub is not enough
> >>
> >> The functionality which is implemented by the pciback and the toolstack
> >> and which is relevant/missing/needed for ARM:
> >>
> >> 1. pciback is used as a database for assignable PCI devices, e.g. xl
> >> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
> >> toolstack needs to know which PCI devices can be passed through it reads
> >> that from the relevant sysfs entries of the pciback.
> >>
> >> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
> >> a PCI device it needs to be unbound from the relevant device driver and bound
> >> to pciback (strictly speaking it is not required that the device is bound to
> >> pciback, but pciback is again used as a database of the passed through PCI
> >> devices, so we can re-bind the devices back to their original drivers when
> >> guest domain shuts down)
> >>
> >> 3. Device reset
> >>
> >> We have previously discussed on xen-devel ML possible solutions to that as from the
> >> above we see that pciback functionality is going to be only partially used on Arm.
> >>
> >> Please see [1] and [2]:
> >>
> >> 1. It is not acceptable to manage the assignable list in Xen itself
> >>
> >> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
> >> the rest like vPCI etc.
> >>
> >> 3. pcifront is not used on Arm
> >
> > It is neither in x86 PVH/HVM guests.
> Didn't know that, thank you for pointing
> >
> >> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
> >> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
> >> all the goals above.
> >>
> >> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
> >> parts then it can be done, yet the patch is going to be the very first brick in that building.
> >
> > Doing this split should be done, as the pcifront specific part could be
> > omitted on x86, too, in case no PV guests using PCI passthrough have to
> > be supported.
> Agree, that the final solution should have the driver split
> >
> >> So, I think this patch is still going to be needed besides which direction we take.
> >
> > Some kind of this patch, yes. It might look different in case the split
> > is done first.
> >
> > I don't mind doing it in either sequence.
> >
> With this patch we have Arm on the same page as the above mentioned x86 guests,
>
> e.g. the driver has unused code, but yet allows Arm to function now.
>
> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>
> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>
> have enough bandwidth for that piece of work at the moment.
That's fair and I don't want to scope-creep this simple patch asking for
an enormous rework. At the same time I don't think we should enable the
whole of pciback on ARM because it would be erroneous and confusing.
I am wonder if there is a simple:
if (!xen_pv_domain())
return;
That we could add in a couple of places in pciback to stop it from
initializing the parts we don't care about. Something along these lines
(untested and probably incomplete).
What do you guys think?
diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
index da34ce85dc88..991ba0a9b359 100644
--- a/drivers/xen/xen-pciback/xenbus.c
+++ b/drivers/xen/xen-pciback/xenbus.c
@@ -15,6 +15,7 @@
#include <xen/xenbus.h>
#include <xen/events.h>
#include <xen/pci.h>
+#include <xen/xen.h>
#include "pciback.h"
#define INVALID_EVTCHN_IRQ (-1)
@@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
const struct xenbus_device_id *id)
{
int err = 0;
- struct xen_pcibk_device *pdev = alloc_pdev(dev);
+ struct xen_pcibk_device *pdev;
+
+ if (!xen_pv_domain())
+ return 0;
+ pdev = alloc_pdev(dev);
if (pdev == NULL) {
err = -ENOMEM;
xenbus_dev_fatal(dev, err,
@@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
int __init xen_pcibk_xenbus_register(void)
{
+ if (!xen_pv_domain())
+ return 0;
+
xen_pcibk_backend = &xen_pcibk_vpci_backend;
if (passthrough)
xen_pcibk_backend = &xen_pcibk_passthrough_backend;
@@ -752,5 +760,7 @@ int __init xen_pcibk_xenbus_register(void)
void __exit xen_pcibk_xenbus_unregister(void)
{
+ if (!xen_pv_domain())
+ return;
xenbus_unregister_driver(&xen_pcibk_driver);
}
On 21.09.21 01:16, Stefano Stabellini wrote:
> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>> On 20.09.21 14:30, Juergen Gross wrote:
>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>> Hello, Stefano!
>>>>
>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>> Hi Oleksandr,
>>>>>
>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>> Not only that
>>>>>
>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>> the same time.
>>>> Correct, it is not used
>>>>>
>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>> possible and better to use pci-stub instead?
>>>>
>>>> Not only that, so pci-stub is not enough
>>>>
>>>> The functionality which is implemented by the pciback and the toolstack
>>>> and which is relevant/missing/needed for ARM:
>>>>
>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>> that from the relevant sysfs entries of the pciback.
>>>>
>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>> devices, so we can re-bind the devices back to their original drivers when
>>>> guest domain shuts down)
>>>>
>>>> 3. Device reset
>>>>
>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>
>>>> Please see [1] and [2]:
>>>>
>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>
>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>> the rest like vPCI etc.
>>>>
>>>> 3. pcifront is not used on Arm
>>>
>>> It is neither in x86 PVH/HVM guests.
>> Didn't know that, thank you for pointing
>>>
>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>> all the goals above.
>>>>
>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>
>>> Doing this split should be done, as the pcifront specific part could be
>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>> be supported.
>> Agree, that the final solution should have the driver split
>>>
>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>
>>> Some kind of this patch, yes. It might look different in case the split
>>> is done first.
>>>
>>> I don't mind doing it in either sequence.
>>>
>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>
>> e.g. the driver has unused code, but yet allows Arm to function now.
>>
>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>
>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>
>> have enough bandwidth for that piece of work at the moment.
>
> That's fair and I don't want to scope-creep this simple patch asking for
> an enormous rework. At the same time I don't think we should enable the
> whole of pciback on ARM because it would be erroneous and confusing.
>
> I am wonder if there is a simple:
>
> if (!xen_pv_domain())
> return;
>
> That we could add in a couple of places in pciback to stop it from
> initializing the parts we don't care about. Something along these lines
> (untested and probably incomplete).
>
> What do you guys think?
Uh no, not in this way, please. This will kill pci passthrough on x86
with dom0 running as PVH. I don't think this is working right now, but
adding more code making it even harder to work should be avoided.
> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
> index da34ce85dc88..991ba0a9b359 100644
> --- a/drivers/xen/xen-pciback/xenbus.c
> +++ b/drivers/xen/xen-pciback/xenbus.c
> @@ -15,6 +15,7 @@
> #include <xen/xenbus.h>
> #include <xen/events.h>
> #include <xen/pci.h>
> +#include <xen/xen.h>
> #include "pciback.h"
>
> #define INVALID_EVTCHN_IRQ (-1)
> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
> const struct xenbus_device_id *id)
> {
> int err = 0;
> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
> + struct xen_pcibk_device *pdev;
> +
> + if (!xen_pv_domain())
> + return 0;
>
> + pdev = alloc_pdev(dev);
This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
early will result in xen_pcibk_xenbus_probe never being called.
> if (pdev == NULL) {
> err = -ENOMEM;
> xenbus_dev_fatal(dev, err,
> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>
> int __init xen_pcibk_xenbus_register(void)
> {
> + if (!xen_pv_domain())
> + return 0;
> +
Use #ifdef CONFIG_X86 instead.
> xen_pcibk_backend = &xen_pcibk_vpci_backend;
> if (passthrough)
> xen_pcibk_backend = &xen_pcibk_passthrough_backend;
> @@ -752,5 +760,7 @@ int __init xen_pcibk_xenbus_register(void)
>
> void __exit xen_pcibk_xenbus_unregister(void)
> {
> + if (!xen_pv_domain())
> + return;
#ifdef again.
> xenbus_unregister_driver(&xen_pcibk_driver);
> }
>
Juergen
On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>
> On 21.09.21 08:20, Juergen Gross wrote:
>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>> Hello, Stefano!
>>>>>>
>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>> Hi Oleksandr,
>>>>>>>
>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>> Not only that
>>>>>>>
>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>> the same time.
>>>>>> Correct, it is not used
>>>>>>>
>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>> possible and better to use pci-stub instead?
>>>>>>
>>>>>> Not only that, so pci-stub is not enough
>>>>>>
>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>
>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>
>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>> guest domain shuts down)
>>>>>>
>>>>>> 3. Device reset
>>>>>>
>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>
>>>>>> Please see [1] and [2]:
>>>>>>
>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>
>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>> the rest like vPCI etc.
>>>>>>
>>>>>> 3. pcifront is not used on Arm
>>>>>
>>>>> It is neither in x86 PVH/HVM guests.
>>>> Didn't know that, thank you for pointing
>>>>>
>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>> all the goals above.
>>>>>>
>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>
>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>> be supported.
>>>> Agree, that the final solution should have the driver split
>>>>>
>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>
>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>> is done first.
>>>>>
>>>>> I don't mind doing it in either sequence.
>>>>>
>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>
>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>
>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>
>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>
>>>> have enough bandwidth for that piece of work at the moment.
>>>
>>> That's fair and I don't want to scope-creep this simple patch asking for
>>> an enormous rework. At the same time I don't think we should enable the
>>> whole of pciback on ARM because it would be erroneous and confusing.
>
> As the first stage before the driver is split or ifdef's used - can we take the patch
> as is now? In either way we chose this needs to be done, e.g. enable compiling
> for other architectures and common code move.
Fine with me in principle. I need to take a more thorough look
at the patch, though.
>
>>>
>>> I am wonder if there is a simple:
>>>
>>> if (!xen_pv_domain())
>>> return;
>>>
>>> That we could add in a couple of places in pciback to stop it from
>>> initializing the parts we don't care about. Something along these lines
>>> (untested and probably incomplete).
>>>
>>> What do you guys think?
>>
>> Uh no, not in this way, please. This will kill pci passthrough on x86
>> with dom0 running as PVH. I don't think this is working right now, but
>> adding more code making it even harder to work should be avoided.
>>
>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>> index da34ce85dc88..991ba0a9b359 100644
>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>> @@ -15,6 +15,7 @@
>>> #include <xen/xenbus.h>
>>> #include <xen/events.h>
>>> #include <xen/pci.h>
>>> +#include <xen/xen.h>
>>> #include "pciback.h"
>>> #define INVALID_EVTCHN_IRQ (-1)
>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>> const struct xenbus_device_id *id)
>>> {
>>> int err = 0;
>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>> + struct xen_pcibk_device *pdev;
>>> +
>>> + if (!xen_pv_domain())
>>> + return 0;
>>> + pdev = alloc_pdev(dev);
>>
>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>> early will result in xen_pcibk_xenbus_probe never being called.
>>
>>> if (pdev == NULL) {
>>> err = -ENOMEM;
>>> xenbus_dev_fatal(dev, err,
>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>> int __init xen_pcibk_xenbus_register(void)
>>> {
>>> + if (!xen_pv_domain())
>>> + return 0;
>>> +
>>
>> Use #ifdef CONFIG_X86 instead.
>
> The title of this patch says that we want to allow this driver for other archs
> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
> right with that respect. Instead, we may want having something like a
> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
> or something which is architecture agnostic.
Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
wanted. The splitting of the driver can still be done later.
> Gating also means that we are not thinking about splitting the backend driver into
> two different ones, e.g. one for "common" code and one for PV stuff.
> Otherwise this ifdefery won't be needed.
I just wanted to avoid the xen_pv_domain() tests creeping in, as
they are wrong IMO.
Juergen
On 21.09.21 09:07, Juergen Gross wrote:
> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>
>> On 21.09.21 08:20, Juergen Gross wrote:
>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>> Hello, Stefano!
>>>>>>>
>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>> Hi Oleksandr,
>>>>>>>>
>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>> Not only that
>>>>>>>>
>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>> the same time.
>>>>>>> Correct, it is not used
>>>>>>>>
>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>
>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>
>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>
>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>
>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>> guest domain shuts down)
>>>>>>>
>>>>>>> 3. Device reset
>>>>>>>
>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>
>>>>>>> Please see [1] and [2]:
>>>>>>>
>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>
>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>> the rest like vPCI etc.
>>>>>>>
>>>>>>> 3. pcifront is not used on Arm
>>>>>>
>>>>>> It is neither in x86 PVH/HVM guests.
>>>>> Didn't know that, thank you for pointing
>>>>>>
>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>> all the goals above.
>>>>>>>
>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>
>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>> be supported.
>>>>> Agree, that the final solution should have the driver split
>>>>>>
>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>
>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>> is done first.
>>>>>>
>>>>>> I don't mind doing it in either sequence.
>>>>>>
>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>
>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>
>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>
>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>
>>>>> have enough bandwidth for that piece of work at the moment.
>>>>
>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>> an enormous rework. At the same time I don't think we should enable the
>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>
>> As the first stage before the driver is split or ifdef's used - can we take the patch
>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>> for other architectures and common code move.
>
> Fine with me in principle. I need to take a more thorough look
> at the patch, though.
Of course
>
>>
>>>>
>>>> I am wonder if there is a simple:
>>>>
>>>> if (!xen_pv_domain())
>>>> return;
>>>>
>>>> That we could add in a couple of places in pciback to stop it from
>>>> initializing the parts we don't care about. Something along these lines
>>>> (untested and probably incomplete).
>>>>
>>>> What do you guys think?
>>>
>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>> with dom0 running as PVH. I don't think this is working right now, but
>>> adding more code making it even harder to work should be avoided.
>>>
>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>> index da34ce85dc88..991ba0a9b359 100644
>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>> @@ -15,6 +15,7 @@
>>>> #include <xen/xenbus.h>
>>>> #include <xen/events.h>
>>>> #include <xen/pci.h>
>>>> +#include <xen/xen.h>
>>>> #include "pciback.h"
>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>> const struct xenbus_device_id *id)
>>>> {
>>>> int err = 0;
>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>> + struct xen_pcibk_device *pdev;
>>>> +
>>>> + if (!xen_pv_domain())
>>>> + return 0;
>>>> + pdev = alloc_pdev(dev);
>>>
>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>
>>>> if (pdev == NULL) {
>>>> err = -ENOMEM;
>>>> xenbus_dev_fatal(dev, err,
>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>> int __init xen_pcibk_xenbus_register(void)
>>>> {
>>>> + if (!xen_pv_domain())
>>>> + return 0;
>>>> +
>>>
>>> Use #ifdef CONFIG_X86 instead.
>>
>> The title of this patch says that we want to allow this driver for other archs
>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>> right with that respect. Instead, we may want having something like a
>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>> or something which is architecture agnostic.
>
> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
> wanted. The splitting of the driver can still be done later.
Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>
>> Gating also means that we are not thinking about splitting the backend driver into
>> two different ones, e.g. one for "common" code and one for PV stuff.
>> Otherwise this ifdefery won't be needed.
>
> I just wanted to avoid the xen_pv_domain() tests creeping in, as
> they are wrong IMO.
>
I understand that
>
> Juergen
On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>
> On 21.09.21 09:07, Juergen Gross wrote:
>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>
>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>> Hello, Stefano!
>>>>>>>>
>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>> Hi Oleksandr,
>>>>>>>>>
>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>> Not only that
>>>>>>>>>
>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>> the same time.
>>>>>>>> Correct, it is not used
>>>>>>>>>
>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>
>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>
>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>
>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>>
>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>>> guest domain shuts down)
>>>>>>>>
>>>>>>>> 3. Device reset
>>>>>>>>
>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>
>>>>>>>> Please see [1] and [2]:
>>>>>>>>
>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>
>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>> the rest like vPCI etc.
>>>>>>>>
>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>
>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>> Didn't know that, thank you for pointing
>>>>>>>
>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>> all the goals above.
>>>>>>>>
>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>
>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>> be supported.
>>>>>> Agree, that the final solution should have the driver split
>>>>>>>
>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>
>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>> is done first.
>>>>>>>
>>>>>>> I don't mind doing it in either sequence.
>>>>>>>
>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>
>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>
>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>
>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>
>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>
>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>
>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>> for other architectures and common code move.
>>
>> Fine with me in principle. I need to take a more thorough look
>> at the patch, though.
> Of course
>>
>>>
>>>>>
>>>>> I am wonder if there is a simple:
>>>>>
>>>>> if (!xen_pv_domain())
>>>>> return;
>>>>>
>>>>> That we could add in a couple of places in pciback to stop it from
>>>>> initializing the parts we don't care about. Something along these lines
>>>>> (untested and probably incomplete).
>>>>>
>>>>> What do you guys think?
>>>>
>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>> adding more code making it even harder to work should be avoided.
>>>>
>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>> @@ -15,6 +15,7 @@
>>>>> #include <xen/xenbus.h>
>>>>> #include <xen/events.h>
>>>>> #include <xen/pci.h>
>>>>> +#include <xen/xen.h>
>>>>> #include "pciback.h"
>>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>> const struct xenbus_device_id *id)
>>>>> {
>>>>> int err = 0;
>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>> + struct xen_pcibk_device *pdev;
>>>>> +
>>>>> + if (!xen_pv_domain())
>>>>> + return 0;
>>>>> + pdev = alloc_pdev(dev);
>>>>
>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>
>>>>> if (pdev == NULL) {
>>>>> err = -ENOMEM;
>>>>> xenbus_dev_fatal(dev, err,
>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>> int __init xen_pcibk_xenbus_register(void)
>>>>> {
>>>>> + if (!xen_pv_domain())
>>>>> + return 0;
>>>>> +
>>>>
>>>> Use #ifdef CONFIG_X86 instead.
>>>
>>> The title of this patch says that we want to allow this driver for other archs
>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>> right with that respect. Instead, we may want having something like a
>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>> or something which is architecture agnostic.
>>
>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>> wanted. The splitting of the driver can still be done later.
>
> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
not set (this will be the case on Arm).
This is another step in the right direction preparing the split.
But as said before, this is not a requirement by me to take your patch.
Juergen
Hi, Stefano!
On 21.09.21 02:16, Stefano Stabellini wrote:
> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>> On 20.09.21 14:30, Juergen Gross wrote:
>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>> Hello, Stefano!
>>>>
>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>> Hi Oleksandr,
>>>>>
>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>> Not only that
>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>> the same time.
>>>> Correct, it is not used
>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>> possible and better to use pci-stub instead?
>>>> Not only that, so pci-stub is not enough
>>>>
>>>> The functionality which is implemented by the pciback and the toolstack
>>>> and which is relevant/missing/needed for ARM:
>>>>
>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>> that from the relevant sysfs entries of the pciback.
>>>>
>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>> devices, so we can re-bind the devices back to their original drivers when
>>>> guest domain shuts down)
>>>>
>>>> 3. Device reset
>>>>
>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>
>>>> Please see [1] and [2]:
>>>>
>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>
>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>> the rest like vPCI etc.
>>>>
>>>> 3. pcifront is not used on Arm
>>> It is neither in x86 PVH/HVM guests.
>> Didn't know that, thank you for pointing
>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>> all the goals above.
>>>>
>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>> Doing this split should be done, as the pcifront specific part could be
>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>> be supported.
>> Agree, that the final solution should have the driver split
>>>> So, I think this patch is still going to be needed besides which direction we take.
>>> Some kind of this patch, yes. It might look different in case the split
>>> is done first.
>>>
>>> I don't mind doing it in either sequence.
>>>
>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>
>> e.g. the driver has unused code, but yet allows Arm to function now.
>>
>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>
>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>
>> have enough bandwidth for that piece of work at the moment.
> That's fair and I don't want to scope-creep this simple patch asking for
> an enormous rework. At the same time I don't think we should enable the
> whole of pciback on ARM because it would be erroneous and confusing.
>
> I am wonder if there is a simple:
>
> if (!xen_pv_domain())
> return;
>
> That we could add in a couple of places in pciback to stop it from
> initializing the parts we don't care about. Something along these lines
> (untested and probably incomplete).
>
> What do you guys think?
I think that it needs to be an additional patch and the PV check seems
reasonable to me. We need to check if gating only part of the initialization
with xen_pv_domain is just enough, e.g. if the rest of the code is ok that
something was not initialized and won't be touched at run-time.
Let's see what other think about the approach
>
>
> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
> index da34ce85dc88..991ba0a9b359 100644
> --- a/drivers/xen/xen-pciback/xenbus.c
> +++ b/drivers/xen/xen-pciback/xenbus.c
> @@ -15,6 +15,7 @@
> #include <xen/xenbus.h>
> #include <xen/events.h>
> #include <xen/pci.h>
> +#include <xen/xen.h>
> #include "pciback.h"
>
> #define INVALID_EVTCHN_IRQ (-1)
> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
> const struct xenbus_device_id *id)
> {
> int err = 0;
> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
> + struct xen_pcibk_device *pdev;
> +
> + if (!xen_pv_domain())
> + return 0;
>
> + pdev = alloc_pdev(dev);
> if (pdev == NULL) {
> err = -ENOMEM;
> xenbus_dev_fatal(dev, err,
> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>
> int __init xen_pcibk_xenbus_register(void)
> {
> + if (!xen_pv_domain())
> + return 0;
> +
> xen_pcibk_backend = &xen_pcibk_vpci_backend;
> if (passthrough)
> xen_pcibk_backend = &xen_pcibk_passthrough_backend;
> @@ -752,5 +760,7 @@ int __init xen_pcibk_xenbus_register(void)
>
> void __exit xen_pcibk_xenbus_unregister(void)
> {
> + if (!xen_pv_domain())
> + return;
> xenbus_unregister_driver(&xen_pcibk_driver);
> }
On 21.09.21 09:49, Juergen Gross wrote:
> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>>
>> On 21.09.21 09:07, Juergen Gross wrote:
>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>>
>>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>>> Hello, Stefano!
>>>>>>>>>
>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>>> Hi Oleksandr,
>>>>>>>>>>
>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>>> Not only that
>>>>>>>>>>
>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>>> the same time.
>>>>>>>>> Correct, it is not used
>>>>>>>>>>
>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>>
>>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>>
>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>>
>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>>>
>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>>>> guest domain shuts down)
>>>>>>>>>
>>>>>>>>> 3. Device reset
>>>>>>>>>
>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>>
>>>>>>>>> Please see [1] and [2]:
>>>>>>>>>
>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>>
>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>>> the rest like vPCI etc.
>>>>>>>>>
>>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>>
>>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>>> Didn't know that, thank you for pointing
>>>>>>>>
>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>>> all the goals above.
>>>>>>>>>
>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>>
>>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>>> be supported.
>>>>>>> Agree, that the final solution should have the driver split
>>>>>>>>
>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>>
>>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>>> is done first.
>>>>>>>>
>>>>>>>> I don't mind doing it in either sequence.
>>>>>>>>
>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>>
>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>>
>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>>
>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>>
>>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>>
>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>>
>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>>> for other architectures and common code move.
>>>
>>> Fine with me in principle. I need to take a more thorough look
>>> at the patch, though.
>> Of course
>>>
>>>>
>>>>>>
>>>>>> I am wonder if there is a simple:
>>>>>>
>>>>>> if (!xen_pv_domain())
>>>>>> return;
>>>>>>
>>>>>> That we could add in a couple of places in pciback to stop it from
>>>>>> initializing the parts we don't care about. Something along these lines
>>>>>> (untested and probably incomplete).
>>>>>>
>>>>>> What do you guys think?
>>>>>
>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>>> adding more code making it even harder to work should be avoided.
>>>>>
>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>>> @@ -15,6 +15,7 @@
>>>>>> #include <xen/xenbus.h>
>>>>>> #include <xen/events.h>
>>>>>> #include <xen/pci.h>
>>>>>> +#include <xen/xen.h>
>>>>>> #include "pciback.h"
>>>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>>> const struct xenbus_device_id *id)
>>>>>> {
>>>>>> int err = 0;
>>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>>> + struct xen_pcibk_device *pdev;
>>>>>> +
>>>>>> + if (!xen_pv_domain())
>>>>>> + return 0;
>>>>>> + pdev = alloc_pdev(dev);
>>>>>
>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>>
>>>>>> if (pdev == NULL) {
>>>>>> err = -ENOMEM;
>>>>>> xenbus_dev_fatal(dev, err,
>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>>> int __init xen_pcibk_xenbus_register(void)
>>>>>> {
>>>>>> + if (!xen_pv_domain())
>>>>>> + return 0;
>>>>>> +
>>>>>
>>>>> Use #ifdef CONFIG_X86 instead.
>>>>
>>>> The title of this patch says that we want to allow this driver for other archs
>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>>> right with that respect. Instead, we may want having something like a
>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>>> or something which is architecture agnostic.
>>>
>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>>> wanted. The splitting of the driver can still be done later.
>>
>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>
> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
> not set (this will be the case on Arm).
But this will require that the existing kernel configurations out there have to additionally
define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling
CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
the things while doing the split/re-work.
I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
I am not quite sure about this though.
>
> This is another step in the right direction preparing the split.
>
> But as said before, this is not a requirement by me to take your patch.
Thank you
>
>
> Juergen
On 21.09.21 09:00, Oleksandr Andrushchenko wrote:
>
> On 21.09.21 09:49, Juergen Gross wrote:
>> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>>>
>>> On 21.09.21 09:07, Juergen Gross wrote:
>>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>>>
>>>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>>>> Hello, Stefano!
>>>>>>>>>>
>>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>>>> Hi Oleksandr,
>>>>>>>>>>>
>>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>>>> Not only that
>>>>>>>>>>>
>>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>>>> the same time.
>>>>>>>>>> Correct, it is not used
>>>>>>>>>>>
>>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>>>
>>>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>>>
>>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>>>
>>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>>>>
>>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>>>>> guest domain shuts down)
>>>>>>>>>>
>>>>>>>>>> 3. Device reset
>>>>>>>>>>
>>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>>>
>>>>>>>>>> Please see [1] and [2]:
>>>>>>>>>>
>>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>>>
>>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>>>> the rest like vPCI etc.
>>>>>>>>>>
>>>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>>>
>>>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>>>> Didn't know that, thank you for pointing
>>>>>>>>>
>>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>>>> all the goals above.
>>>>>>>>>>
>>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>>>
>>>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>>>> be supported.
>>>>>>>> Agree, that the final solution should have the driver split
>>>>>>>>>
>>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>>>
>>>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>>>> is done first.
>>>>>>>>>
>>>>>>>>> I don't mind doing it in either sequence.
>>>>>>>>>
>>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>>>
>>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>>>
>>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>>>
>>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>>>
>>>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>>>
>>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>>>
>>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>>>> for other architectures and common code move.
>>>>
>>>> Fine with me in principle. I need to take a more thorough look
>>>> at the patch, though.
>>> Of course
>>>>
>>>>>
>>>>>>>
>>>>>>> I am wonder if there is a simple:
>>>>>>>
>>>>>>> if (!xen_pv_domain())
>>>>>>> return;
>>>>>>>
>>>>>>> That we could add in a couple of places in pciback to stop it from
>>>>>>> initializing the parts we don't care about. Something along these lines
>>>>>>> (untested and probably incomplete).
>>>>>>>
>>>>>>> What do you guys think?
>>>>>>
>>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>>>> adding more code making it even harder to work should be avoided.
>>>>>>
>>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>>>> @@ -15,6 +15,7 @@
>>>>>>> #include <xen/xenbus.h>
>>>>>>> #include <xen/events.h>
>>>>>>> #include <xen/pci.h>
>>>>>>> +#include <xen/xen.h>
>>>>>>> #include "pciback.h"
>>>>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>>>> const struct xenbus_device_id *id)
>>>>>>> {
>>>>>>> int err = 0;
>>>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>>>> + struct xen_pcibk_device *pdev;
>>>>>>> +
>>>>>>> + if (!xen_pv_domain())
>>>>>>> + return 0;
>>>>>>> + pdev = alloc_pdev(dev);
>>>>>>
>>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>>>
>>>>>>> if (pdev == NULL) {
>>>>>>> err = -ENOMEM;
>>>>>>> xenbus_dev_fatal(dev, err,
>>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>>>> int __init xen_pcibk_xenbus_register(void)
>>>>>>> {
>>>>>>> + if (!xen_pv_domain())
>>>>>>> + return 0;
>>>>>>> +
>>>>>>
>>>>>> Use #ifdef CONFIG_X86 instead.
>>>>>
>>>>> The title of this patch says that we want to allow this driver for other archs
>>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>>>> right with that respect. Instead, we may want having something like a
>>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>>>> or something which is architecture agnostic.
>>>>
>>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>>>> wanted. The splitting of the driver can still be done later.
>>>
>>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
>>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
>>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
>>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
>>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>>
>> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
>> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
>> not set (this will be the case on Arm).
>
> But this will require that the existing kernel configurations out there have to additionally
> define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling
> CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
> the things while doing the split/re-work.
By letting XEN_PCIDEV_BACKEND select XEN_PCI_STUB this won't happen.
> I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
> may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
> I am not quite sure about this though.
This would be a very weird semantics of XEN_PCI_STUB, as the stub part
is needed on X86 and on Arm.
Gating could even be done with Stefano's patch just by replacing his
"!xen_pv_domain()" tests with "!IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND)".
Juergen
On 21.09.21 10:09, Juergen Gross wrote:
> On 21.09.21 09:00, Oleksandr Andrushchenko wrote:
>>
>> On 21.09.21 09:49, Juergen Gross wrote:
>>> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>>>>
>>>> On 21.09.21 09:07, Juergen Gross wrote:
>>>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>>>>
>>>>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>>>>> Hello, Stefano!
>>>>>>>>>>>
>>>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>>>>> Hi Oleksandr,
>>>>>>>>>>>>
>>>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>>>>> Not only that
>>>>>>>>>>>>
>>>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>>>>> the same time.
>>>>>>>>>>> Correct, it is not used
>>>>>>>>>>>>
>>>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>>>>
>>>>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>>>>
>>>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>>>>
>>>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>>>>>
>>>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>>>>>> guest domain shuts down)
>>>>>>>>>>>
>>>>>>>>>>> 3. Device reset
>>>>>>>>>>>
>>>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>>>>
>>>>>>>>>>> Please see [1] and [2]:
>>>>>>>>>>>
>>>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>>>>
>>>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>>>>> the rest like vPCI etc.
>>>>>>>>>>>
>>>>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>>>>
>>>>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>>>>> Didn't know that, thank you for pointing
>>>>>>>>>>
>>>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>>>>> all the goals above.
>>>>>>>>>>>
>>>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>>>>
>>>>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>>>>> be supported.
>>>>>>>>> Agree, that the final solution should have the driver split
>>>>>>>>>>
>>>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>>>>
>>>>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>>>>> is done first.
>>>>>>>>>>
>>>>>>>>>> I don't mind doing it in either sequence.
>>>>>>>>>>
>>>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>>>>
>>>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>>>>
>>>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>>>>
>>>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>>>>
>>>>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>>>>
>>>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>>>>
>>>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>>>>> for other architectures and common code move.
>>>>>
>>>>> Fine with me in principle. I need to take a more thorough look
>>>>> at the patch, though.
>>>> Of course
>>>>>
>>>>>>
>>>>>>>>
>>>>>>>> I am wonder if there is a simple:
>>>>>>>>
>>>>>>>> if (!xen_pv_domain())
>>>>>>>> return;
>>>>>>>>
>>>>>>>> That we could add in a couple of places in pciback to stop it from
>>>>>>>> initializing the parts we don't care about. Something along these lines
>>>>>>>> (untested and probably incomplete).
>>>>>>>>
>>>>>>>> What do you guys think?
>>>>>>>
>>>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>>>>> adding more code making it even harder to work should be avoided.
>>>>>>>
>>>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>> @@ -15,6 +15,7 @@
>>>>>>>> #include <xen/xenbus.h>
>>>>>>>> #include <xen/events.h>
>>>>>>>> #include <xen/pci.h>
>>>>>>>> +#include <xen/xen.h>
>>>>>>>> #include "pciback.h"
>>>>>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>>>>> const struct xenbus_device_id *id)
>>>>>>>> {
>>>>>>>> int err = 0;
>>>>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>>>>> + struct xen_pcibk_device *pdev;
>>>>>>>> +
>>>>>>>> + if (!xen_pv_domain())
>>>>>>>> + return 0;
>>>>>>>> + pdev = alloc_pdev(dev);
>>>>>>>
>>>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>>>>
>>>>>>>> if (pdev == NULL) {
>>>>>>>> err = -ENOMEM;
>>>>>>>> xenbus_dev_fatal(dev, err,
>>>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>>>>> int __init xen_pcibk_xenbus_register(void)
>>>>>>>> {
>>>>>>>> + if (!xen_pv_domain())
>>>>>>>> + return 0;
>>>>>>>> +
>>>>>>>
>>>>>>> Use #ifdef CONFIG_X86 instead.
>>>>>>
>>>>>> The title of this patch says that we want to allow this driver for other archs
>>>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>>>>> right with that respect. Instead, we may want having something like a
>>>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>>>>> or something which is architecture agnostic.
>>>>>
>>>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>>>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>>>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>>>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>>>>> wanted. The splitting of the driver can still be done later.
>>>>
>>>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
>>>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
>>>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
>>>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
>>>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>>>
>>> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
>>> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
>>> not set (this will be the case on Arm).
>>
>> But this will require that the existing kernel configurations out there have to additionally define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
>> the things while doing the split/re-work.
>
> By letting XEN_PCIDEV_BACKEND select XEN_PCI_STUB this won't happen.
Indeed
>
>> I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
>> may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
>> I am not quite sure about this though.
>
> This would be a very weird semantics of XEN_PCI_STUB, as the stub part
> is needed on X86 and on Arm.
>
> Gating could even be done with Stefano's patch just by replacing his
> "!xen_pv_domain()" tests with "!IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND)".
Makes sense.
Another question if we do not want the code to be compiled or not executed?
If the later then we can define something like:
bool need_pv_part(void)
{
return IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND);
}
and then just use need_pv_part() for the checks where it is needed.
This allows avoiding multiple ifdef's through the code
>
>
> Juergen
On 21.09.21 08:20, Juergen Gross wrote:
> On 21.09.21 01:16, Stefano Stabellini wrote:
>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>> Hello, Stefano!
>>>>>
>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>> Hi Oleksandr,
>>>>>>
>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>> Not only that
>>>>>>
>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>> the same time.
>>>>> Correct, it is not used
>>>>>>
>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>> possible and better to use pci-stub instead?
>>>>>
>>>>> Not only that, so pci-stub is not enough
>>>>>
>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>> and which is relevant/missing/needed for ARM:
>>>>>
>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>> that from the relevant sysfs entries of the pciback.
>>>>>
>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>> guest domain shuts down)
>>>>>
>>>>> 3. Device reset
>>>>>
>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>
>>>>> Please see [1] and [2]:
>>>>>
>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>
>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>> the rest like vPCI etc.
>>>>>
>>>>> 3. pcifront is not used on Arm
>>>>
>>>> It is neither in x86 PVH/HVM guests.
>>> Didn't know that, thank you for pointing
>>>>
>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>> all the goals above.
>>>>>
>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>
>>>> Doing this split should be done, as the pcifront specific part could be
>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>> be supported.
>>> Agree, that the final solution should have the driver split
>>>>
>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>
>>>> Some kind of this patch, yes. It might look different in case the split
>>>> is done first.
>>>>
>>>> I don't mind doing it in either sequence.
>>>>
>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>
>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>
>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>
>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>
>>> have enough bandwidth for that piece of work at the moment.
>>
>> That's fair and I don't want to scope-creep this simple patch asking for
>> an enormous rework. At the same time I don't think we should enable the
>> whole of pciback on ARM because it would be erroneous and confusing.
As the first stage before the driver is split or ifdef's used - can we take the patch
as is now? In either way we chose this needs to be done, e.g. enable compiling
for other architectures and common code move.
>>
>> I am wonder if there is a simple:
>>
>> if (!xen_pv_domain())
>> return;
>>
>> That we could add in a couple of places in pciback to stop it from
>> initializing the parts we don't care about. Something along these lines
>> (untested and probably incomplete).
>>
>> What do you guys think?
>
> Uh no, not in this way, please. This will kill pci passthrough on x86
> with dom0 running as PVH. I don't think this is working right now, but
> adding more code making it even harder to work should be avoided.
>
>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>> index da34ce85dc88..991ba0a9b359 100644
>> --- a/drivers/xen/xen-pciback/xenbus.c
>> +++ b/drivers/xen/xen-pciback/xenbus.c
>> @@ -15,6 +15,7 @@
>> #include <xen/xenbus.h>
>> #include <xen/events.h>
>> #include <xen/pci.h>
>> +#include <xen/xen.h>
>> #include "pciback.h"
>> #define INVALID_EVTCHN_IRQ (-1)
>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>> const struct xenbus_device_id *id)
>> {
>> int err = 0;
>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>> + struct xen_pcibk_device *pdev;
>> +
>> + if (!xen_pv_domain())
>> + return 0;
>> + pdev = alloc_pdev(dev);
>
> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
> early will result in xen_pcibk_xenbus_probe never being called.
>
>> if (pdev == NULL) {
>> err = -ENOMEM;
>> xenbus_dev_fatal(dev, err,
>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>> int __init xen_pcibk_xenbus_register(void)
>> {
>> + if (!xen_pv_domain())
>> + return 0;
>> +
>
> Use #ifdef CONFIG_X86 instead.
The title of this patch says that we want to allow this driver for other archs
and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
right with that respect. Instead, we may want having something like a
dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
or something which is architecture agnostic.
Gating also means that we are not thinking about splitting the backend driver into
two different ones, e.g. one for "common" code and one for PV stuff.
Otherwise this ifdefery won't be needed.
>
>> xen_pcibk_backend = &xen_pcibk_vpci_backend;
>> if (passthrough)
>> xen_pcibk_backend = &xen_pcibk_passthrough_backend;
>> @@ -752,5 +760,7 @@ int __init xen_pcibk_xenbus_register(void)
>> void __exit xen_pcibk_xenbus_unregister(void)
>> {
>> + if (!xen_pv_domain())
>> + return;
>
> #ifdef again.
>
>> xenbus_unregister_driver(&xen_pcibk_driver);
>> }
>>
>
>
> Juergen
On 17.09.21 15:01, Oleksandr Andrushchenko wrote:
> From: Oleksandr Andrushchenko <[email protected]>
>
> Xen-pciback driver was designed to be built for x86 only. But it
> can also be used by other architectures, e.g. Arm.
> Re-structure the driver in a way that it can be built for other
> platforms as well.
>
> Signed-off-by: Oleksandr Andrushchenko <[email protected]>
> Signed-off-by: Anastasiia Lukianenko <[email protected]>
>
> ---
> Tested on Arm and x86.
> ---
> arch/x86/include/asm/xen/pci.h | 24 ----------
> arch/x86/pci/xen.c | 74 +----------------------------
> drivers/xen/Kconfig | 2 +-
> drivers/xen/events/events_base.c | 1 +
> drivers/xen/pci.c | 75 ++++++++++++++++++++++++++++++
> drivers/xen/xen-pciback/pci_stub.c | 3 +-
> drivers/xen/xen-pciback/xenbus.c | 2 +-
> include/xen/pci.h | 34 ++++++++++++++
> 8 files changed, 115 insertions(+), 100 deletions(-)
> create mode 100644 include/xen/pci.h
>
> diff --git a/arch/x86/include/asm/xen/pci.h b/arch/x86/include/asm/xen/pci.h
> index 3506d8c598c1..9ff7b49bca08 100644
> --- a/arch/x86/include/asm/xen/pci.h
> +++ b/arch/x86/include/asm/xen/pci.h
> @@ -14,30 +14,6 @@ static inline int pci_xen_hvm_init(void)
> return -1;
> }
> #endif
> -#if defined(CONFIG_XEN_DOM0)
> -int __init pci_xen_initial_domain(void);
Why are you removing this prototype? It is X86 specific.
> -int xen_find_device_domain_owner(struct pci_dev *dev);
> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
> -int xen_unregister_device_domain_owner(struct pci_dev *dev);
> -#else
> -static inline int __init pci_xen_initial_domain(void)
> -{
> - return -1;
> -}
> -static inline int xen_find_device_domain_owner(struct pci_dev *dev)
> -{
> - return -1;
> -}
> -static inline int xen_register_device_domain_owner(struct pci_dev *dev,
> - uint16_t domain)
> -{
> - return -1;
> -}
> -static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
> -{
> - return -1;
> -}
> -#endif
>
> #if defined(CONFIG_PCI_MSI)
> #if defined(CONFIG_PCI_XEN)
> diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
> index 3d41a09c2c14..4a45b0bf9ae4 100644
> --- a/arch/x86/pci/xen.c
> +++ b/arch/x86/pci/xen.c
> @@ -23,6 +23,7 @@
>
> #include <xen/features.h>
> #include <xen/events.h>
> +#include <xen/pci.h>
> #include <asm/xen/pci.h>
> #include <asm/xen/cpuid.h>
> #include <asm/apic.h>
> @@ -583,77 +584,4 @@ int __init pci_xen_initial_domain(void)
> }
> return 0;
> }
> -
> -struct xen_device_domain_owner {
> - domid_t domain;
> - struct pci_dev *dev;
> - struct list_head list;
> -};
> -
> -static DEFINE_SPINLOCK(dev_domain_list_spinlock);
> -static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
> -
> -static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - list_for_each_entry(owner, &dev_domain_list, list) {
> - if (owner->dev == dev)
> - return owner;
> - }
> - return NULL;
> -}
> -
> -int xen_find_device_domain_owner(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> - int domain = -ENODEV;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - owner = find_device(dev);
> - if (owner)
> - domain = owner->domain;
> - spin_unlock(&dev_domain_list_spinlock);
> - return domain;
> -}
> -EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
> -
> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
> - if (!owner)
> - return -ENODEV;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - if (find_device(dev)) {
> - spin_unlock(&dev_domain_list_spinlock);
> - kfree(owner);
> - return -EEXIST;
> - }
> - owner->domain = domain;
> - owner->dev = dev;
> - list_add_tail(&owner->list, &dev_domain_list);
> - spin_unlock(&dev_domain_list_spinlock);
> - return 0;
> -}
> -EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
> -
> -int xen_unregister_device_domain_owner(struct pci_dev *dev)
> -{
> - struct xen_device_domain_owner *owner;
> -
> - spin_lock(&dev_domain_list_spinlock);
> - owner = find_device(dev);
> - if (!owner) {
> - spin_unlock(&dev_domain_list_spinlock);
> - return -ENODEV;
> - }
> - list_del(&owner->list);
> - spin_unlock(&dev_domain_list_spinlock);
> - kfree(owner);
> - return 0;
> -}
> -EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
> #endif
> diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
> index a37eb52fb401..057ddf61ef61 100644
> --- a/drivers/xen/Kconfig
> +++ b/drivers/xen/Kconfig
> @@ -182,7 +182,7 @@ config SWIOTLB_XEN
>
> config XEN_PCIDEV_BACKEND
> tristate "Xen PCI-device backend driver"
> - depends on PCI && X86 && XEN
> + depends on PCI && XEN
> depends on XEN_BACKEND
> default m
> help
> diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c
> index a78704ae3618..35493ff0d146 100644
> --- a/drivers/xen/events/events_base.c
> +++ b/drivers/xen/events/events_base.c
> @@ -65,6 +65,7 @@
> #include <xen/interface/vcpu.h>
> #include <xen/xenbus.h>
> #include <asm/hw_irq.h>
> +#include <xen/pci.h>
This wouldn't be needed if you'd let the pci_xen_initial_domain()
prototype where it has been.
Juergen
On 21.09.21 10:54, Juergen Gross wrote:
> On 17.09.21 15:01, Oleksandr Andrushchenko wrote:
>> From: Oleksandr Andrushchenko <[email protected]>
>>
>> Xen-pciback driver was designed to be built for x86 only. But it
>> can also be used by other architectures, e.g. Arm.
>> Re-structure the driver in a way that it can be built for other
>> platforms as well.
>>
>> Signed-off-by: Oleksandr Andrushchenko <[email protected]>
>> Signed-off-by: Anastasiia Lukianenko <[email protected]>
>>
>> ---
>> Tested on Arm and x86.
>> ---
>> arch/x86/include/asm/xen/pci.h | 24 ----------
>> arch/x86/pci/xen.c | 74 +----------------------------
>> drivers/xen/Kconfig | 2 +-
>> drivers/xen/events/events_base.c | 1 +
>> drivers/xen/pci.c | 75 ++++++++++++++++++++++++++++++
>> drivers/xen/xen-pciback/pci_stub.c | 3 +-
>> drivers/xen/xen-pciback/xenbus.c | 2 +-
>> include/xen/pci.h | 34 ++++++++++++++
>> 8 files changed, 115 insertions(+), 100 deletions(-)
>> create mode 100644 include/xen/pci.h
>>
>> diff --git a/arch/x86/include/asm/xen/pci.h b/arch/x86/include/asm/xen/pci.h
>> index 3506d8c598c1..9ff7b49bca08 100644
>> --- a/arch/x86/include/asm/xen/pci.h
>> +++ b/arch/x86/include/asm/xen/pci.h
>> @@ -14,30 +14,6 @@ static inline int pci_xen_hvm_init(void)
>> return -1;
>> }
>> #endif
>> -#if defined(CONFIG_XEN_DOM0)
>> -int __init pci_xen_initial_domain(void);
>
> Why are you removing this prototype? It is X86 specific.
Indeed it is. Will not remove it form x86 code
>
>> -int xen_find_device_domain_owner(struct pci_dev *dev);
>> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain);
>> -int xen_unregister_device_domain_owner(struct pci_dev *dev);
>> -#else
>> -static inline int __init pci_xen_initial_domain(void)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_find_device_domain_owner(struct pci_dev *dev)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_register_device_domain_owner(struct pci_dev *dev,
>> - uint16_t domain)
>> -{
>> - return -1;
>> -}
>> -static inline int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> -{
>> - return -1;
>> -}
>> -#endif
>> #if defined(CONFIG_PCI_MSI)
>> #if defined(CONFIG_PCI_XEN)
>> diff --git a/arch/x86/pci/xen.c b/arch/x86/pci/xen.c
>> index 3d41a09c2c14..4a45b0bf9ae4 100644
>> --- a/arch/x86/pci/xen.c
>> +++ b/arch/x86/pci/xen.c
>> @@ -23,6 +23,7 @@
>> #include <xen/features.h>
>> #include <xen/events.h>
>> +#include <xen/pci.h>
>> #include <asm/xen/pci.h>
>> #include <asm/xen/cpuid.h>
>> #include <asm/apic.h>
>> @@ -583,77 +584,4 @@ int __init pci_xen_initial_domain(void)
>> }
>> return 0;
>> }
>> -
>> -struct xen_device_domain_owner {
>> - domid_t domain;
>> - struct pci_dev *dev;
>> - struct list_head list;
>> -};
>> -
>> -static DEFINE_SPINLOCK(dev_domain_list_spinlock);
>> -static struct list_head dev_domain_list = LIST_HEAD_INIT(dev_domain_list);
>> -
>> -static struct xen_device_domain_owner *find_device(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - list_for_each_entry(owner, &dev_domain_list, list) {
>> - if (owner->dev == dev)
>> - return owner;
>> - }
>> - return NULL;
>> -}
>> -
>> -int xen_find_device_domain_owner(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> - int domain = -ENODEV;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - owner = find_device(dev);
>> - if (owner)
>> - domain = owner->domain;
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return domain;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_find_device_domain_owner);
>> -
>> -int xen_register_device_domain_owner(struct pci_dev *dev, uint16_t domain)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - owner = kzalloc(sizeof(struct xen_device_domain_owner), GFP_KERNEL);
>> - if (!owner)
>> - return -ENODEV;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - if (find_device(dev)) {
>> - spin_unlock(&dev_domain_list_spinlock);
>> - kfree(owner);
>> - return -EEXIST;
>> - }
>> - owner->domain = domain;
>> - owner->dev = dev;
>> - list_add_tail(&owner->list, &dev_domain_list);
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return 0;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_register_device_domain_owner);
>> -
>> -int xen_unregister_device_domain_owner(struct pci_dev *dev)
>> -{
>> - struct xen_device_domain_owner *owner;
>> -
>> - spin_lock(&dev_domain_list_spinlock);
>> - owner = find_device(dev);
>> - if (!owner) {
>> - spin_unlock(&dev_domain_list_spinlock);
>> - return -ENODEV;
>> - }
>> - list_del(&owner->list);
>> - spin_unlock(&dev_domain_list_spinlock);
>> - kfree(owner);
>> - return 0;
>> -}
>> -EXPORT_SYMBOL_GPL(xen_unregister_device_domain_owner);
>> #endif
>> diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
>> index a37eb52fb401..057ddf61ef61 100644
>> --- a/drivers/xen/Kconfig
>> +++ b/drivers/xen/Kconfig
>> @@ -182,7 +182,7 @@ config SWIOTLB_XEN
>> config XEN_PCIDEV_BACKEND
>> tristate "Xen PCI-device backend driver"
>> - depends on PCI && X86 && XEN
>> + depends on PCI && XEN
>> depends on XEN_BACKEND
>> default m
>> help
>> diff --git a/drivers/xen/events/events_base.c b/drivers/xen/events/events_base.c
>> index a78704ae3618..35493ff0d146 100644
>> --- a/drivers/xen/events/events_base.c
>> +++ b/drivers/xen/events/events_base.c
>> @@ -65,6 +65,7 @@
>> #include <xen/interface/vcpu.h>
>> #include <xen/xenbus.h>
>> #include <asm/hw_irq.h>
>> +#include <xen/pci.h>
>
> This wouldn't be needed if you'd let the pci_xen_initial_domain()
> prototype where it has been.
Sure, will leave it where it was
>
>
> Juergen
On Tue, 21 Sep 2021, Oleksandr Andrushchenko wrote:
> On 21.09.21 10:09, Juergen Gross wrote:
> > On 21.09.21 09:00, Oleksandr Andrushchenko wrote:
> >>
> >> On 21.09.21 09:49, Juergen Gross wrote:
> >>> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
> >>>>
> >>>> On 21.09.21 09:07, Juergen Gross wrote:
> >>>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
> >>>>>>
> >>>>>> On 21.09.21 08:20, Juergen Gross wrote:
> >>>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
> >>>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
> >>>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
> >>>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
> >>>>>>>>>>> Hello, Stefano!
> >>>>>>>>>>>
> >>>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
> >>>>>>>>>>>> Hi Oleksandr,
> >>>>>>>>>>>>
> >>>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
> >>>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
> >>>>>>>>>>> Not only that
> >>>>>>>>>>>>
> >>>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
> >>>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
> >>>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
> >>>>>>>>>>>> the same time.
> >>>>>>>>>>> Correct, it is not used
> >>>>>>>>>>>>
> >>>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
> >>>>>>>>>>>> possible and better to use pci-stub instead?
> >>>>>>>>>>>
> >>>>>>>>>>> Not only that, so pci-stub is not enough
> >>>>>>>>>>>
> >>>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
> >>>>>>>>>>> and which is relevant/missing/needed for ARM:
> >>>>>>>>>>>
> >>>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
> >>>>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
> >>>>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
> >>>>>>>>>>> that from the relevant sysfs entries of the pciback.
> >>>>>>>>>>>
> >>>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
> >>>>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
> >>>>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
> >>>>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
> >>>>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
> >>>>>>>>>>> guest domain shuts down)
> >>>>>>>>>>>
> >>>>>>>>>>> 3. Device reset
> >>>>>>>>>>>
> >>>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
> >>>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
> >>>>>>>>>>>
> >>>>>>>>>>> Please see [1] and [2]:
> >>>>>>>>>>>
> >>>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
> >>>>>>>>>>>
> >>>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
> >>>>>>>>>>> the rest like vPCI etc.
> >>>>>>>>>>>
> >>>>>>>>>>> 3. pcifront is not used on Arm
> >>>>>>>>>>
> >>>>>>>>>> It is neither in x86 PVH/HVM guests.
> >>>>>>>>> Didn't know that, thank you for pointing
> >>>>>>>>>>
> >>>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
> >>>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
> >>>>>>>>>>> all the goals above.
> >>>>>>>>>>>
> >>>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
> >>>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
> >>>>>>>>>>
> >>>>>>>>>> Doing this split should be done, as the pcifront specific part could be
> >>>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
> >>>>>>>>>> be supported.
> >>>>>>>>> Agree, that the final solution should have the driver split
> >>>>>>>>>>
> >>>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
> >>>>>>>>>>
> >>>>>>>>>> Some kind of this patch, yes. It might look different in case the split
> >>>>>>>>>> is done first.
> >>>>>>>>>>
> >>>>>>>>>> I don't mind doing it in either sequence.
> >>>>>>>>>>
> >>>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
> >>>>>>>>>
> >>>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
> >>>>>>>>>
> >>>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
> >>>>>>>>>
> >>>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
> >>>>>>>>>
> >>>>>>>>> have enough bandwidth for that piece of work at the moment.
> >>>>>>>>
> >>>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
> >>>>>>>> an enormous rework. At the same time I don't think we should enable the
> >>>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
> >>>>>>
> >>>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
> >>>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
> >>>>>> for other architectures and common code move.
> >>>>>
> >>>>> Fine with me in principle. I need to take a more thorough look
> >>>>> at the patch, though.
> >>>> Of course
> >>>>>
> >>>>>>
> >>>>>>>>
> >>>>>>>> I am wonder if there is a simple:
> >>>>>>>>
> >>>>>>>> if (!xen_pv_domain())
> >>>>>>>> return;
> >>>>>>>>
> >>>>>>>> That we could add in a couple of places in pciback to stop it from
> >>>>>>>> initializing the parts we don't care about. Something along these lines
> >>>>>>>> (untested and probably incomplete).
> >>>>>>>>
> >>>>>>>> What do you guys think?
> >>>>>>>
> >>>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
> >>>>>>> with dom0 running as PVH. I don't think this is working right now, but
> >>>>>>> adding more code making it even harder to work should be avoided.
> >>>>>>>
> >>>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
> >>>>>>>> index da34ce85dc88..991ba0a9b359 100644
> >>>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
> >>>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
> >>>>>>>> @@ -15,6 +15,7 @@
> >>>>>>>> #include <xen/xenbus.h>
> >>>>>>>> #include <xen/events.h>
> >>>>>>>> #include <xen/pci.h>
> >>>>>>>> +#include <xen/xen.h>
> >>>>>>>> #include "pciback.h"
> >>>>>>>> #define INVALID_EVTCHN_IRQ (-1)
> >>>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
> >>>>>>>> const struct xenbus_device_id *id)
> >>>>>>>> {
> >>>>>>>> int err = 0;
> >>>>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
> >>>>>>>> + struct xen_pcibk_device *pdev;
> >>>>>>>> +
> >>>>>>>> + if (!xen_pv_domain())
> >>>>>>>> + return 0;
> >>>>>>>> + pdev = alloc_pdev(dev);
> >>>>>>>
> >>>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
> >>>>>>> early will result in xen_pcibk_xenbus_probe never being called.
> >>>>>>>
> >>>>>>>> if (pdev == NULL) {
> >>>>>>>> err = -ENOMEM;
> >>>>>>>> xenbus_dev_fatal(dev, err,
> >>>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
> >>>>>>>> int __init xen_pcibk_xenbus_register(void)
> >>>>>>>> {
> >>>>>>>> + if (!xen_pv_domain())
> >>>>>>>> + return 0;
> >>>>>>>> +
> >>>>>>>
> >>>>>>> Use #ifdef CONFIG_X86 instead.
> >>>>>>
> >>>>>> The title of this patch says that we want to allow this driver for other archs
> >>>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
> >>>>>> right with that respect. Instead, we may want having something like a
> >>>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
> >>>>>> or something which is architecture agnostic.
> >>>>>
> >>>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
> >>>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
> >>>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
> >>>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
> >>>>> wanted. The splitting of the driver can still be done later.
> >>>>
> >>>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
> >>>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
> >>>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
> >>>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
> >>>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
> >>>
> >>> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
> >>> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
> >>> not set (this will be the case on Arm).
> >>
> >> But this will require that the existing kernel configurations out there have to additionally define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
> >> the things while doing the split/re-work.
> >
> > By letting XEN_PCIDEV_BACKEND select XEN_PCI_STUB this won't happen.
> Indeed
> >
> >> I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
> >> may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
> >> I am not quite sure about this though.
> >
> > This would be a very weird semantics of XEN_PCI_STUB, as the stub part
> > is needed on X86 and on Arm.
> >
> > Gating could even be done with Stefano's patch just by replacing his
> > "!xen_pv_domain()" tests with "!IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND)".
>
> Makes sense.
>
> Another question if we do not want the code to be compiled or not executed?
>
> If the later then we can define something like:
>
> bool need_pv_part(void)
>
> {
>
> return IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND);
>
> }
>
> and then just use need_pv_part() for the checks where it is needed.
>
> This allows avoiding multiple ifdef's through the code
This is even better.
For my clarity, Oleksandr, are you OK with adding a few need_pv_part()
checks through the code as part of this series so that the PV PCI
backend is not initialized?
I don't have a good test environment ready for this, so I cannot really
volunteer myself.
I would prefer if we made this change as part of this series so that the
PV PCI backend features doesn't get enabled on ARM, not even temporarily.
On 21.09.21 23:44, Stefano Stabellini wrote:
> On Tue, 21 Sep 2021, Oleksandr Andrushchenko wrote:
>> On 21.09.21 10:09, Juergen Gross wrote:
>>> On 21.09.21 09:00, Oleksandr Andrushchenko wrote:
>>>> On 21.09.21 09:49, Juergen Gross wrote:
>>>>> On 21.09.21 08:38, Oleksandr Andrushchenko wrote:
>>>>>> On 21.09.21 09:07, Juergen Gross wrote:
>>>>>>> On 21.09.21 07:51, Oleksandr Andrushchenko wrote:
>>>>>>>> On 21.09.21 08:20, Juergen Gross wrote:
>>>>>>>>> On 21.09.21 01:16, Stefano Stabellini wrote:
>>>>>>>>>> On Mon, 20 Sep 2021, Oleksandr Andrushchenko wrote:
>>>>>>>>>>> On 20.09.21 14:30, Juergen Gross wrote:
>>>>>>>>>>>> On 20.09.21 07:23, Oleksandr Andrushchenko wrote:
>>>>>>>>>>>>> Hello, Stefano!
>>>>>>>>>>>>>
>>>>>>>>>>>>> On 18.09.21 00:45, Stefano Stabellini wrote:
>>>>>>>>>>>>>> Hi Oleksandr,
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Why do you want to enable pciback on ARM? Is it only to "disable" a PCI
>>>>>>>>>>>>>> device in Dom0 so that it can be safely assigned to a DomU?
>>>>>>>>>>>>> Not only that
>>>>>>>>>>>>>> I am asking because actually I don't think we want to enable the PV PCI
>>>>>>>>>>>>>> backend feature of pciback on ARM, right? That would clash with the PCI
>>>>>>>>>>>>>> assignment work you have been doing in Xen. They couldn't both work at
>>>>>>>>>>>>>> the same time.
>>>>>>>>>>>>> Correct, it is not used
>>>>>>>>>>>>>> If we only need pciback to "park" a device in Dom0, wouldn't it be
>>>>>>>>>>>>>> possible and better to use pci-stub instead?
>>>>>>>>>>>>> Not only that, so pci-stub is not enough
>>>>>>>>>>>>>
>>>>>>>>>>>>> The functionality which is implemented by the pciback and the toolstack
>>>>>>>>>>>>> and which is relevant/missing/needed for ARM:
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1. pciback is used as a database for assignable PCI devices, e.g. xl
>>>>>>>>>>>>> pci-assignable-{add|remove|list} manipulates that list. So, whenever the
>>>>>>>>>>>>> toolstack needs to know which PCI devices can be passed through it reads
>>>>>>>>>>>>> that from the relevant sysfs entries of the pciback.
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2. pciback is used to hold the unbound PCI devices, e.g. when passing through
>>>>>>>>>>>>> a PCI device it needs to be unbound from the relevant device driver and bound
>>>>>>>>>>>>> to pciback (strictly speaking it is not required that the device is bound to
>>>>>>>>>>>>> pciback, but pciback is again used as a database of the passed through PCI
>>>>>>>>>>>>> devices, so we can re-bind the devices back to their original drivers when
>>>>>>>>>>>>> guest domain shuts down)
>>>>>>>>>>>>>
>>>>>>>>>>>>> 3. Device reset
>>>>>>>>>>>>>
>>>>>>>>>>>>> We have previously discussed on xen-devel ML possible solutions to that as from the
>>>>>>>>>>>>> above we see that pciback functionality is going to be only partially used on Arm.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Please see [1] and [2]:
>>>>>>>>>>>>>
>>>>>>>>>>>>> 1. It is not acceptable to manage the assignable list in Xen itself
>>>>>>>>>>>>>
>>>>>>>>>>>>> 2. pciback can be split into two parts: PCI assignable/bind/reset handling and
>>>>>>>>>>>>> the rest like vPCI etc.
>>>>>>>>>>>>>
>>>>>>>>>>>>> 3. pcifront is not used on Arm
>>>>>>>>>>>> It is neither in x86 PVH/HVM guests.
>>>>>>>>>>> Didn't know that, thank you for pointing
>>>>>>>>>>>>> So, limited use of the pciback is one of the bricks used to enable PCI passthrough
>>>>>>>>>>>>> on Arm. It was enough to just re-structure the driver and have it run on Arm to achieve
>>>>>>>>>>>>> all the goals above.
>>>>>>>>>>>>>
>>>>>>>>>>>>> If we still think it is desirable to break the pciback driver into "common" and "pcifront specific"
>>>>>>>>>>>>> parts then it can be done, yet the patch is going to be the very first brick in that building.
>>>>>>>>>>>> Doing this split should be done, as the pcifront specific part could be
>>>>>>>>>>>> omitted on x86, too, in case no PV guests using PCI passthrough have to
>>>>>>>>>>>> be supported.
>>>>>>>>>>> Agree, that the final solution should have the driver split
>>>>>>>>>>>>> So, I think this patch is still going to be needed besides which direction we take.
>>>>>>>>>>>> Some kind of this patch, yes. It might look different in case the split
>>>>>>>>>>>> is done first.
>>>>>>>>>>>>
>>>>>>>>>>>> I don't mind doing it in either sequence.
>>>>>>>>>>>>
>>>>>>>>>>> With this patch we have Arm on the same page as the above mentioned x86 guests,
>>>>>>>>>>>
>>>>>>>>>>> e.g. the driver has unused code, but yet allows Arm to function now.
>>>>>>>>>>>
>>>>>>>>>>> At this stage of PCI passthrough on Arm it is yet enough. Long term, when
>>>>>>>>>>>
>>>>>>>>>>> the driver gets split, Arm will benefit from that split too, but unfortunately I do not
>>>>>>>>>>>
>>>>>>>>>>> have enough bandwidth for that piece of work at the moment.
>>>>>>>>>> That's fair and I don't want to scope-creep this simple patch asking for
>>>>>>>>>> an enormous rework. At the same time I don't think we should enable the
>>>>>>>>>> whole of pciback on ARM because it would be erroneous and confusing.
>>>>>>>> As the first stage before the driver is split or ifdef's used - can we take the patch
>>>>>>>> as is now? In either way we chose this needs to be done, e.g. enable compiling
>>>>>>>> for other architectures and common code move.
>>>>>>> Fine with me in principle. I need to take a more thorough look
>>>>>>> at the patch, though.
>>>>>> Of course
>>>>>>>>>> I am wonder if there is a simple:
>>>>>>>>>>
>>>>>>>>>> if (!xen_pv_domain())
>>>>>>>>>> return;
>>>>>>>>>>
>>>>>>>>>> That we could add in a couple of places in pciback to stop it from
>>>>>>>>>> initializing the parts we don't care about. Something along these lines
>>>>>>>>>> (untested and probably incomplete).
>>>>>>>>>>
>>>>>>>>>> What do you guys think?
>>>>>>>>> Uh no, not in this way, please. This will kill pci passthrough on x86
>>>>>>>>> with dom0 running as PVH. I don't think this is working right now, but
>>>>>>>>> adding more code making it even harder to work should be avoided.
>>>>>>>>>
>>>>>>>>>> diff --git a/drivers/xen/xen-pciback/xenbus.c b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> index da34ce85dc88..991ba0a9b359 100644
>>>>>>>>>> --- a/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> +++ b/drivers/xen/xen-pciback/xenbus.c
>>>>>>>>>> @@ -15,6 +15,7 @@
>>>>>>>>>> #include <xen/xenbus.h>
>>>>>>>>>> #include <xen/events.h>
>>>>>>>>>> #include <xen/pci.h>
>>>>>>>>>> +#include <xen/xen.h>
>>>>>>>>>> #include "pciback.h"
>>>>>>>>>> #define INVALID_EVTCHN_IRQ (-1)
>>>>>>>>>> @@ -685,8 +686,12 @@ static int xen_pcibk_xenbus_probe(struct xenbus_device *dev,
>>>>>>>>>> const struct xenbus_device_id *id)
>>>>>>>>>> {
>>>>>>>>>> int err = 0;
>>>>>>>>>> - struct xen_pcibk_device *pdev = alloc_pdev(dev);
>>>>>>>>>> + struct xen_pcibk_device *pdev;
>>>>>>>>>> +
>>>>>>>>>> + if (!xen_pv_domain())
>>>>>>>>>> + return 0;
>>>>>>>>>> + pdev = alloc_pdev(dev);
>>>>>>>>> This hunk isn't needed, as with bailing out of xen_pcibk_xenbus_register
>>>>>>>>> early will result in xen_pcibk_xenbus_probe never being called.
>>>>>>>>>
>>>>>>>>>> if (pdev == NULL) {
>>>>>>>>>> err = -ENOMEM;
>>>>>>>>>> xenbus_dev_fatal(dev, err,
>>>>>>>>>> @@ -743,6 +748,9 @@ const struct xen_pcibk_backend *__read_mostly xen_pcibk_backend;
>>>>>>>>>> int __init xen_pcibk_xenbus_register(void)
>>>>>>>>>> {
>>>>>>>>>> + if (!xen_pv_domain())
>>>>>>>>>> + return 0;
>>>>>>>>>> +
>>>>>>>>> Use #ifdef CONFIG_X86 instead.
>>>>>>>> The title of this patch says that we want to allow this driver for other archs
>>>>>>>> and now we want to introduce "#ifdef CONFIG_X86" which doesn't sound
>>>>>>>> right with that respect. Instead, we may want having something like a
>>>>>>>> dedicated gate for this, e.g. "#ifdef CONFIG_XEN_PCIDEV_BACKEND_SUPP_PV"
>>>>>>>> or something which is architecture agnostic.
>>>>>>> Something like that, yes. But I'd rather use CONFIG_XEN_PCIDEV_BACKEND
>>>>>>> acting as this gate and introduce CONFIG_XEN_PCI_STUB for the stub
>>>>>>> functionality needed on Arm. XEN_PCIDEV_BACKEND would depend on X86 and
>>>>>>> select XEN_PCI_STUB, while on Arm XEN_PCI_STUB could be configured if
>>>>>>> wanted. The splitting of the driver can still be done later.
>>>>>> Hm, pciback is now compiled when CONFIG_XEN_PCIDEV_BACKEND is enabled
>>>>>> and we want to skip some parts of its code when CONFIG_XEN_PCI_STUB is set.
>>>>>> So, I imagine that for x86 we just enable CONFIG_XEN_PCIDEV_BACKEND and the
>>>>>> driver compiles in its current state. For Arm we enable both CONFIG_XEN_PCIDEV_BACKEND
>>>>>> and CONFIG_XEN_PCI_STUB, so part of the driver is not compiled.
>>>>> No, I'd rather switch to compiling xen-pciback when CONFIG_XEN_PCI_STUB
>>>>> is set and compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is
>>>>> not set (this will be the case on Arm).
>>>> But this will require that the existing kernel configurations out there have to additionally define CONFIG_XEN_PCI_STUB to get what they had before with simply enabling CONFIG_XEN_PCIDEV_BACKEND. My point was that it is probably desirable not to break
>>>> the things while doing the split/re-work.
>>> By letting XEN_PCIDEV_BACKEND select XEN_PCI_STUB this won't happen.
>> Indeed
>>>> I also thought that "compile only parts of it when CONFIG_XEN_PCIDEV_BACKEND is not set"
>>>> may have more code gated rather than with gating unwanted code with CONFIG_XEN_PCI_STUB.
>>>> I am not quite sure about this though.
>>> This would be a very weird semantics of XEN_PCI_STUB, as the stub part
>>> is needed on X86 and on Arm.
>>>
>>> Gating could even be done with Stefano's patch just by replacing his
>>> "!xen_pv_domain()" tests with "!IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND)".
>> Makes sense.
>>
>> Another question if we do not want the code to be compiled or not executed?
>>
>> If the later then we can define something like:
>>
>> bool need_pv_part(void)
>>
>> {
>>
>> return IS_ENABLED(CONFIG_XEN_PCIDEV_BACKEND);
>>
>> }
>>
>> and then just use need_pv_part() for the checks where it is needed.
>>
>> This allows avoiding multiple ifdef's through the code
> This is even better.
>
> For my clarity, Oleksandr, are you OK with adding a few need_pv_part()
> checks through the code as part of this series so that the PV PCI
> backend is not initialized?
Yes
>
> I don't have a good test environment ready for this, so I cannot really
> volunteer myself.
>
> I would prefer if we made this change as part of this series so that the
> PV PCI backend features doesn't get enabled on ARM, not even temporarily.
Ok, I will push v2 today with the additional patch for PV