Fix coherence for VMbus and PCI pass-thru devices in Hyper-V VM
Subsystems (6)
cover Fix coherence for VMbus and PCI pass-thru devices in Hyper-V VM
[Resend to fix an email address typo for Bjorn Helgaas]
Hyper-V VMs have VMbus synthetic devices and PCI pass-thru devices that are added dynamically via the VMbus protocol and are not represented in the ACPI DSDT. Only the top level VMbus node exists in the DSDT. As such, on ARM64 these devices don't pick up coherence information and default to not hardware coherent. This results in extra software coherence management overhead since the synthetic devices are always hardware coherent. PCI pass-thru devices are also hardware coherent in all current usage scenarios.
Fix this by propagating coherence information from the top level VMbus node in the DSDT to all VMbus synthetic devices and PCI pass-thru devices. While smaller granularity of control would be better, basing on the VMbus node in the DSDT gives as escape path if a future scenario arises with devices that are not hardware coherent.
The first two patches are prep to allow manipulating device coherence from a module (since the VMbus driver can be built as a module) and from architecture independent code without having a bunch of #ifdef's.
The third patch propagates the VMbus node coherence to VMbus synthetic devices.
The fourth patch propagates the coherence to PCI pass-thru devices.
Michael Kelley (4):
ACPI: scan: Export acpi_get_dma_attr()
dma-mapping: Add wrapper function to set dma_coherent
Drivers: hv: vmbus: Propagate VMbus coherence to each VMbus device
PCI: hv: Propagate coherence from VMbus device to PCI device
drivers/acpi/scan.c | 1 + drivers/hv/vmbus_drv.c | 15 +++++++++++++++ drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- include/linux/dma-map-ops.h | 9 +++++++++ 4 files changed, 38 insertions(+), 4 deletions(-)
Signature · 1 line
1.8.3.1
PATCH RESEND 1/4 ACPI: scan: Export acpi_get_dma_attr()
Export acpi_get_dma_attr() so that it can be used by the Hyper-V VMbus driver, which may be built as a module. The related function acpi_dma_configure_id() is already exported.
--- drivers/acpi/scan.c | 1 + 1 file changed, 1 insertion(+)
drivers/acpi/scan.c +1 -0
@@ -1489,6 +1489,7 @@ enum dev_dma_attr acpi_get_dma_attr(struct acpi_device *adev)else return DEV_DMA_NON_COHERENT ; } +EXPORT_SYMBOL_GPL (acpi_get_dma_attr);/** * acpi_dma_get_range() - Get device DMA parameters.
Signature · 1 line
1.8.3.1
Robin Murphy
reviewerRe: [PATCH 1/4 RESEND] ACPI: scan: Export acpi_get_dma_attr()
Quoted 3 lines · Michael Kelley
Export
acpi_get_dma_attr()so that it can be used by the Hyper-V VMbus driver, which may be built as a module. The related functionacpi_dma_configure_id()is already exported.
No. Use device_get_dma_attr() like everyone else, please.
Robin.
Quoted 17 lines
Signed-off-by Michael Kelley <[email protected]>--- drivers/acpi/scan.c | 1 + 1 file changed, 1 insertion(+)drivers/acpi/scan.c +1
-0@@ -1489,6 +1489,7 @@ enum dev_dma_attr acpi_get_dma_attr(struct acpi_device *adev)else return DEV_DMA_NON_COHERENT ; } +EXPORT_SYMBOL_GPL (acpi_get_dma_attr);/** * acpi_dma_get_range() - Get device DMA parameters.
RE: [PATCH 1/4 RESEND] ACPI: scan: Export acpi_get_dma_attr()
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022 9:31 AM
Quoted 8 lines · Michael Kelley
Quoted 3 lines
Export
acpi_get_dma_attr()so that it can be used by the Hyper-V VMbus driver, which may be built as a module. The related functionacpi_dma_configure_id()is already exported.No. Use
device_get_dma_attr()like everyone else, please.
Got it. That works.
Michael
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
PATCH RESEND 2/4 dma-mapping: Add wrapper function to set dma_coherent
In reply to Michael KelleyAdd a wrapper function to set dma_coherent, avoiding the need for complex #ifdef's when setting it in architecture independent code.
--- include/linux/dma-map-ops.h | 9 +++++++++ 1 file changed, 9 insertions(+)
include/linux/dma-map-ops.h +9 -0
@@ -254,11 +254,20 @@ static inline bool dev_is_dma_coherent(struct device *dev) {return dev->dma_coherent; } +static inline void dev_set_dma_coherent (struct device *dev, +bool coherent) +{ + dev->dma_coherent = coherent; +}#else static inline bool dev_is_dma_coherent (struct device *dev) {return true; } +static inline void dev_set_dma_coherent (struct device *dev, +bool coherent) +{ +}#endif /* CONFIG_ARCH_HAS_DMA_COHERENCE_H */ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
Signature · 6 lines
1.8.3.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
Robin Murphy
reviewerRe: [PATCH 2/4 RESEND] dma-mapping: Add wrapper function to set dma_coherent
Quoted 2 lines · Michael Kelley via iommu
Add a wrapper function to set
dma_coherent, avoiding the need for complex #ifdef's when setting it in architecture independent code.
No. It might happen to work out on the architectures you're looking at, but if Hyper-V were ever to support, say, AArch32 VMs you might see the problem. arch_setup_dma_ops() is the tool for this job.
Robin.
Quoted 30 lines
Signed-off-by Michael Kelley <[email protected]>--- include/linux/dma-map-ops.h | 9 +++++++++ 1 file changed, 9 insertions(+)include/linux/dma-map-ops.h +9
-0@@ -254,11 +254,20 @@ static inline bool dev_is_dma_coherent(struct device *dev) {return dev->dma_coherent; } +static inline void dev_set_dma_coherent (struct device *dev, +bool coherent) +{ + dev->dma_coherent = coherent; +}#else static inline bool dev_is_dma_coherent (struct device *dev) {return true; } +static inline void dev_set_dma_coherent (struct device *dev, +bool coherent) +{ +}#endif /* CONFIG_ARCH_HAS_DMA_COHERENCE_H */ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
RE: [PATCH 2/4 RESEND] dma-mapping: Add wrapper function to set dma_coherent
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022 10:20 AM
Quoted 9 lines · Michael Kelley via iommu
Quoted 2 lines
Add a wrapper function to set
dma_coherent, avoiding the need for complex #ifdef's when setting it in architecture independent code.No. It might happen to work out on the architectures you're looking at, but if Hyper-V were ever to support, say, AArch32 VMs you might see the problem.
arch_setup_dma_ops()is the tool for this job.
OK. There's currently no vIOMMU in a Hyper-V guest, so presumably the
code would call arch_setup_dma_ops() with the dma_base, size, and iommu parameters set to 0 and NULL. This call can then be used in Patch 3 instead of acpi_dma_configure(), and in the Patch 4 hv_dma_configure() function as you suggested. arch_setup_dma_ops() is not exported, so I'd need to wrap it in a Hyper-V specific function in built-in code that is exported.
But at some point in the future if there's a vIOMMU in Hyper-V guests, this approach will need some rework.
Does that make sense? Thanks for your input and suggestions ...
Michael
Robin Murphy
reviewerRe: [PATCH 2/4 RESEND] dma-mapping: Add wrapper function to set dma_coherent
Quoted 18 lines · Michael Kelley (LINUX)
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022 10:20 AM
Quoted 9 lines · Michael Kelley via iommu
Quoted 2 lines
Add a wrapper function to set
dma_coherent, avoiding the need for complex #ifdef's when setting it in architecture independent code.No. It might happen to work out on the architectures you're looking at, but if Hyper-V were ever to support, say, AArch32 VMs you might see the problem.
arch_setup_dma_ops()is the tool for this job.OK. There's currently no vIOMMU in a Hyper-V guest, so presumably the
code would callarch_setup_dma_ops()with thedma_base, size, and iommu parameters set to 0 and NULL. This call can then be used in Patch 3 instead ofacpi_dma_configure(), and in the Patch 4hv_dma_configure()function as you suggested.arch_setup_dma_ops()is not exported, so I'd need to wrap it in a Hyper-V specific function in built-in code that is exported.
Quoted 4 lines
But at some point in the future if there's a vIOMMU in Hyper-V guests, this approach will need some rework.
Does that make sense? Thanks for your input and suggestions ...
Yes, that's essentially what I had in mind. If you did present a vIOMMU to the guest, presumably you'd either have to construct a regular IORT/VIOT, and thus end up adding the root complex to the ACPI namespace too so it can be referenced, at which point it would all get picked up by the standard machinery, or come up with some magic VMBus mechanism that would need a whole load of work to wire up in all the relevant places anyway.
(But please lean extremely heavily towards the former option!)
Thanks,
Robin.
RE: [PATCH 2/4 RESEND] dma-mapping: Add wrapper function to set dma_coherent
From: Robin Murphy <[email protected]> Sent: Friday, March 18, 2022 4:08 AM
Quoted 34 lines · Michael Kelley (LINUX)
Quoted 1 line
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022
10:20 AM
Quoted 17 lines
Quoted 9 lines · Michael Kelley via iommu
Quoted 2 lines
Add a wrapper function to set
dma_coherent, avoiding the need for complex #ifdef's when setting it in architecture independent code.No. It might happen to work out on the architectures you're looking at, but if Hyper-V were ever to support, say, AArch32 VMs you might see the problem.
arch_setup_dma_ops()is the tool for this job.OK. There's currently no vIOMMU in a Hyper-V guest, so presumably the
code would callarch_setup_dma_ops()with thedma_base, size, and iommu parameters set to 0 and NULL. This call can then be used in Patch 3 instead ofacpi_dma_configure(), and in the Patch 4hv_dma_configure()function as you suggested.arch_setup_dma_ops()is not exported, so I'd need to wrap it in a Hyper-V specific function in built-in code that is exported.Quoted 4 lines
But at some point in the future if there's a vIOMMU in Hyper-V guests, this approach will need some rework.
Does that make sense? Thanks for your input and suggestions ...
Yes, that's essentially what I had in mind. If you did present a vIOMMU to the guest, presumably you'd either have to construct a regular IORT/VIOT, and thus end up adding the root complex to the ACPI namespace too so it can be referenced, at which point it would all get picked up by the standard machinery, or come up with some magic VMBus mechanism that would need a whole load of work to wire up in all the relevant places anyway.
Quoted 1 line
(But please lean extremely heavily towards the former option!)
Agreed!
Quoted 3 lines
Thanks,
Robin.
PATCH RESEND 3/4 Drivers: hv: vmbus: Propagate VMbus coherence to each VMbus device
In reply to Michael KelleyVMbus synthetic devices are not represented in the ACPI DSDT -- only the top level VMbus device is represented. As a result, on ARM64 coherence information in the _CCA method is not specified for synthetic devices, so they default to not hardware coherent. Drivers for some of these synthetic devices have been recently updated to use the standard DMA APIs, and they are incurring extra overhead of unneeded software coherence management.
Fix this by propagating coherence information from the VMbus node in ACPI to the individual synthetic devices. There's no effect on x86/x64 where devices are always hardware coherent.
--- drivers/hv/vmbus_drv.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)
drivers/hv/vmbus_drv.c +15 -0
@@ -904,6 +904,21 @@ static int vmbus_probe(struct device *child_device)drv_to_hv_drv (child_device->driver);struct hv_device *dev =device_to_hv_device (child_device);const struct hv_vmbus_device_id *dev_id; +enum dev_dma_attr coherent; + +/* +* On ARM64, propagate the DMA coherence setting from the top level +* VMbus ACPI device to the child VMbus device being added here. +* Older Hyper-V ARM64 versions don't set the _CCA method on the +* top level VMbus ACPI device as they should. Treat these cases +* as DMA coherent since that's the assumption made by Hyper-V. +* +* On x86/x64 these calls assume coherence and have no effect. +*/ + coherent =acpi_get_dma_attr (hv_acpi_dev); +if (coherent ==DEV_DMA_NOT_SUPPORTED ) + coherent =DEV_DMA_COHERENT ; +acpi_dma_configure (child_device, coherent); dev_id =hv_vmbus_get_id (drv, dev);if (drv->probe) {
Signature · 6 lines
1.8.3.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
Robin Murphy
reviewerRe: [PATCH 3/4 RESEND] Drivers: hv: vmbus: Propagate VMbus coherence to each VMbus device
Quoted 30 lines · Michael Kelley via iommu
VMbus synthetic devices are not represented in the ACPI DSDT -- only the top level VMbus device is represented. As a result, on ARM64 coherence information in the
_CCAmethod is not specified for synthetic devices, so they default to not hardware coherent. Drivers for some of these synthetic devices have been recently updated to use the standard DMA APIs, and they are incurring extra overhead of unneeded software coherence management.Fix this by propagating coherence information from the VMbus node in ACPI to the individual synthetic devices. There's no effect on x86/x64 where devices are always hardware coherent.
Signed-off-by Michael Kelley <[email protected]>--- drivers/hv/vmbus_drv.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)drivers/hv/vmbus_drv.c +5
-0@@ -904,6 +904,21 @@ static int vmbus_probe(struct device *child_device)drv_to_hv_drv (child_device->driver);struct hv_device *dev =device_to_hv_device (child_device);const struct hv_vmbus_device_id *dev_id; +enum dev_dma_attr coherent; + +/* +* On ARM64, propagate the DMA coherence setting from the top level +* VMbus ACPI device to the child VMbus device being added here.
drivers/hv/vmbus_drv.c static int vmbus_probe(struct device *child_device) · line 921
Quoted 10 lines
drivers/hv/vmbus_drv.c +10
-0+* Older Hyper-V ARM64 versions don't set the _CCA method on the +* top level VMbus ACPI device as they should. Treat these cases +* as DMA coherent since that's the assumption made by Hyper-V. +* +* On x86/x64 these calls assume coherence and have no effect. +*/ + coherent =acpi_get_dma_attr (hv_acpi_dev); +if (coherent ==DEV_DMA_NOT_SUPPORTED ) + coherent =DEV_DMA_COHERENT ; +acpi_dma_configure (child_device, coherent);
acpi_dma_configure is for devices represented in ACPI. The commit message implies that these VMBus devices aren't represented in ACPI. What gives?
Robin.
Quoted 3 lines
drivers/hv/vmbus_drv.c +0
-0dev_id =hv_vmbus_get_id (drv, dev);if (drv->probe) {
PATCH RESEND 4/4 PCI: hv: Propagate coherence from VMbus device to PCI device
In reply to Michael KelleyPCI pass-thru devices in a Hyper-V VM are represented as a VMBus device and as a PCI device. The coherence of the VMbus device is set based on the VMbus node in ACPI, but the PCI device has no ACPI node and defaults to not hardware coherent. This results in extra software coherence management overhead on ARM64 when devices are hardware coherent.
Fix this by propagating the coherence of the VMbus device to the PCI device. There's no effect on x86/x64 where devices are always hardware coherent.
--- drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)
drivers/pci/controller/pci-hyperv.c +13 -4
@@ -49,6 +49,7 @@#include <linux/refcount.h> #include <linux/irqdomain.h> #include <linux/acpi.h> +#include <linux/dma-map-ops.h> #include <asm/mshyperv.h> /*
@@ -2142,9 +2143,9 @@ static void hv_pci_remove_slots(struct hv_pcibus_device *hbus) }/* -* Set NUMA node for the devices on the bus +* Set NUMA node and DMA coherence for the devices on the bus */ -static void hv_pci_assign_numa_node (struct hv_pcibus_device *hbus) +static void hv_pci_assign_properties (struct hv_pcibus_device *hbus) {struct pci_dev *dev;struct pci_bus *bus = hbus->bridge->bus;
@@ -2167,6 +2168,14 @@ static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)numa_map_to_online_node ( hv_dev->desc.virtual_numa_node)); +/* +* On ARM64, propagate the DMA coherence from the VMbus device +* to the corresponding PCI device. On x86/x64, these calls +* have no effect because DMA is always hardware coherent. +*/ +dev_set_dma_coherent (&dev->dev, +dev_is_dma_coherent (&hbus->hdev->device)); +put_pcichild (hv_dev); } }
@@ -2191,7 +2200,7 @@ static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)return error;pci_lock_rescan_remove (); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);pci_bus_assign_resources (bridge->bus);hv_pci_assign_slots (hbus);pci_bus_add_devices (bridge->bus);
@@ -2458,7 +2467,7 @@ static void pci_devices_present_work(struct work_struct *work)*/ pci_lock_rescan_remove ();pci_scan_child_bus (hbus->bridge->bus); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);hv_pci_assign_slots (hbus);pci_unlock_rescan_remove ();break ;
Signature · 6 lines
1.8.3.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
Robin Murphy
reviewerRe: [PATCH 4/4 RESEND] PCI: hv: Propagate coherence from VMbus device to PCI device
Quoted 41 lines · Michael Kelley via iommu
PCI pass-thru devices in a Hyper-V VM are represented as a VMBus device and as a PCI device. The coherence of the VMbus device is set based on the VMbus node in ACPI, but the PCI device has no ACPI node and defaults to not hardware coherent. This results in extra software coherence management overhead on ARM64 when devices are hardware coherent.
Fix this by propagating the coherence of the VMbus device to the PCI device. There's no effect on x86/x64 where devices are always hardware coherent.
Signed-off-by Michael Kelley <[email protected]>--- drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)drivers/pci/controller/pci-hyperv.c +3
-2@@ -49,6 +49,7 @@#include <linux/refcount.h> #include <linux/irqdomain.h> #include <linux/acpi.h> +#include <linux/dma-map-ops.h> #include <asm/mshyperv.h> /* @@ -2142,9 +2143,9 @@ static void hv_pci_remove_slots(struct hv_pcibus_device *hbus) }/* -* Set NUMA node for the devices on the bus +* Set NUMA node and DMA coherence for the devices on the bus */ -static void hv_pci_assign_numa_node (struct hv_pcibus_device *hbus) +static void hv_pci_assign_properties (struct hv_pcibus_device *hbus) {struct pci_dev *dev;struct pci_bus *bus = hbus->bridge->bus;
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus) · line 2177
Quoted 10 lines
drivers/pci/controller/pci-hyperv.c +7
-0@@ -2167,6 +2168,14 @@ static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)numa_map_to_online_node ( hv_dev->desc.virtual_numa_node)); +/* +* On ARM64, propagate the DMA coherence from the VMbus device +* to the corresponding PCI device. On x86/x64, these calls +* have no effect because DMA is always hardware coherent. +*/ +dev_set_dma_coherent (&dev->dev, +dev_is_dma_coherent (&hbus->hdev->device));
Eww... if you really have to do this, I'd prefer to see a proper hv_dma_configure() helper implemented and wired up to pci_dma_configure(). Although since it's a generic property I guess at worst pci_dma_configure could perhaps propagate coherency from the host bridge to its children by itself in the absence of any other firmware info. And it's built-in so could use arch_setup_dma_ops() like everyone else.
Robin.
Quoted 22 lines
drivers/pci/controller/pci-hyperv.c +3
-2+put_pcichild (hv_dev); } }@@ -2191,7 +2200,7 @@ static int create_root_hv_pci_bus(struct hv_pcibus_device *hbus)return error;pci_lock_rescan_remove (); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);pci_bus_assign_resources (bridge->bus);hv_pci_assign_slots (hbus);pci_bus_add_devices (bridge->bus);@@ -2458,7 +2467,7 @@ static void pci_devices_present_work(struct work_struct *work)*/ pci_lock_rescan_remove ();pci_scan_child_bus (hbus->bridge->bus); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);hv_pci_assign_slots (hbus);pci_unlock_rescan_remove ();break ;
RE: [PATCH 4/4 RESEND] PCI: hv: Propagate coherence from VMbus device to PCI device
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022 10:15 AM
Quoted 48 lines · Michael Kelley via iommu
drivers/pci/controller/pci-hyperv.c static void hv_pci_remove_slots(struct hv_pcibus_device
Quoted 29 lines
PCI pass-thru devices in a Hyper-V VM are represented as a VMBus device and as a PCI device. The coherence of the VMbus device is set based on the VMbus node in ACPI, but the PCI device has no ACPI node and defaults to not hardware coherent. This results in extra software coherence management overhead on ARM64 when devices are hardware coherent.
Fix this by propagating the coherence of the VMbus device to the PCI device. There's no effect on x86/x64 where devices are always hardware coherent.
Signed-off-by Michael Kelley <[email protected]>--- drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)drivers/pci/controller/pci-hyperv.c +1
-0@@ -49,6 +49,7 @@#include <linux/refcount.h> #include <linux/irqdomain.h> #include <linux/acpi.h> +#include <linux/dma-map-ops.h> #include <asm/mshyperv.h> /* @@ -2142,9 +2143,9 @@ static void hv_pci_remove_slots(struct hv_pcibus_device*hbus)
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct
Quoted 12 lines
drivers/pci/controller/pci-hyperv.c +2
-2}/* -* Set NUMA node for the devices on the bus +* Set NUMA node and DMA coherence for the devices on the bus */ -static void hv_pci_assign_numa_node (struct hv_pcibus_device *hbus) +static void hv_pci_assign_properties (struct hv_pcibus_device *hbus) {struct pci_dev *dev;struct pci_bus *bus = hbus->bridge->bus;@@ -2167,6 +2168,14 @@ static void hv_pci_assign_numa_node(struct
hv_pcibus_device*hbus)Quoted 3 lines
drivers/pci/controller/pci-hyperv.c +0
-0numa_map_to_online_node ( hv_dev->desc.virtual_numa_node));
Quoted 16 lines
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus) · line 2177
Quoted 7 lines
drivers/pci/controller/pci-hyperv.c +7
-0+/* +* On ARM64, propagate the DMA coherence from the VMbus device +* to the corresponding PCI device. On x86/x64, these calls +* have no effect because DMA is always hardware coherent. +*/ +dev_set_dma_coherent (&dev->dev, +dev_is_dma_coherent (&hbus->hdev->device));Eww... if you really have to do this, I'd prefer to see a proper
hv_dma_configure()helper implemented and wired up topci_dma_configure(). Although since it's a generic property I guess at worstpci_dma_configurecould perhaps propagate coherency from the host bridge to its children by itself in the absence of any other firmware info. And it's built-in so could usearch_setup_dma_ops()like everyone else.
I'm not seeing an existing mechanism to provide a "helper" or override
of pci_dma_configure(). Could you elaborate? Or is this something
that needs to be created?
Michael
Quoted 18 lines
Robin.
drivers/pci/controller/pci-hyperv.c static int create_root_hv_pci_bus(struct hv_pcibus_device
Quoted 5 lines
drivers/pci/controller/pci-hyperv.c +1
-0+put_pcichild (hv_dev); } }@@ -2191,7 +2200,7 @@ static int create_root_hv_pci_bus(struct hv_pcibus_device*hbus)
drivers/pci/controller/pci-hyperv.c static void pci_devices_present_work(struct work_struct
Quoted 9 lines
drivers/pci/controller/pci-hyperv.c +1
-1return error;pci_lock_rescan_remove (); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);pci_bus_assign_resources (bridge->bus);hv_pci_assign_slots (hbus);pci_bus_add_devices (bridge->bus);@@ -2458,7 +2467,7 @@ static void pci_devices_present_work(struct work_struct*work)
Quoted 8 lines
drivers/pci/controller/pci-hyperv.c static void pci_devices_present_work(struct work_struct *work) · line 2473
Quoted 8 lines
drivers/pci/controller/pci-hyperv.c +1
-1*/ pci_lock_rescan_remove ();pci_scan_child_bus (hbus->bridge->bus); -hv_pci_assign_numa_node (hbus); +hv_pci_assign_properties (hbus);hv_pci_assign_slots (hbus);pci_unlock_rescan_remove ();break ;
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Robin Murphy
reviewerRe: [PATCH 4/4 RESEND] PCI: hv: Propagate coherence from VMbus device to PCI device
Quoted 66 lines · Michael Kelley (LINUX)
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022 10:15 AM
Quoted 48 lines · Michael Kelley via iommu
drivers/pci/controller/pci-hyperv.c static void hv_pci_remove_slots(struct hv_pcibus_device
Quoted 29 lines
PCI pass-thru devices in a Hyper-V VM are represented as a VMBus device and as a PCI device. The coherence of the VMbus device is set based on the VMbus node in ACPI, but the PCI device has no ACPI node and defaults to not hardware coherent. This results in extra software coherence management overhead on ARM64 when devices are hardware coherent.
Fix this by propagating the coherence of the VMbus device to the PCI device. There's no effect on x86/x64 where devices are always hardware coherent.
Signed-off-by Michael Kelley <[email protected]>--- drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)drivers/pci/controller/pci-hyperv.c +1
-0@@ -49,6 +49,7 @@#include <linux/refcount.h> #include <linux/irqdomain.h> #include <linux/acpi.h> +#include <linux/dma-map-ops.h> #include <asm/mshyperv.h> /* @@ -2142,9 +2143,9 @@ static void hv_pci_remove_slots(struct hv_pcibus_device*hbus)
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct
Quoted 12 lines
drivers/pci/controller/pci-hyperv.c +2
-2}/* -* Set NUMA node for the devices on the bus +* Set NUMA node and DMA coherence for the devices on the bus */ -static void hv_pci_assign_numa_node (struct hv_pcibus_device *hbus) +static void hv_pci_assign_properties (struct hv_pcibus_device *hbus) {struct pci_dev *dev;struct pci_bus *bus = hbus->bridge->bus;@@ -2167,6 +2168,14 @@ static void hv_pci_assign_numa_node(struct
hv_pcibus_device*hbus)Quoted 3 lines
drivers/pci/controller/pci-hyperv.c +0
-0numa_map_to_online_node ( hv_dev->desc.virtual_numa_node));Quoted 16 lines
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct · line 2177
Quoted 7 lines
drivers/pci/controller/pci-hyperv.c +7
-0+/* +* On ARM64, propagate the DMA coherence from the VMbus device +* to the corresponding PCI device. On x86/x64, these calls +* have no effect because DMA is always hardware coherent. +*/ +dev_set_dma_coherent (&dev->dev, +dev_is_dma_coherent (&hbus->hdev->device));Eww... if you really have to do this, I'd prefer to see a proper
hv_dma_configure()helper implemented and wired up topci_dma_configure(). Although since it's a generic property I guess at worstpci_dma_configurecould perhaps propagate coherency from the host bridge to its children by itself in the absence of any other firmware info. And it's built-in so could usearch_setup_dma_ops()like everyone else.
Quoted 3 lines
I'm not seeing an existing mechanism to provide a "helper" or override
ofpci_dma_configure(). Could you elaborate? Or is this something
that needs to be created?
I mean something like the diff below (other #includes omitted for clarity). Essentially if VMBus has its own way of describing parts of the system, then for those parts it's nice if it could fit into the same abstractions we use for firmware-based system description.
Cheers,
Robin.
----->8-----
drivers/pci/pci-driver.c +3 -0
@@ -20,6 +20,7 @@#include <linux/of_device.h> #include <linux/acpi.h> #include <linux/dma-map-ops.h> +#include <linux/hyperv.h> #include "pci.h" #include "pcie/portdrv.h"
@@ -1602,6 +1603,8 @@ static int pci_dma_configure(struct device *dev)struct acpi_device *adev =to_acpi_device_node (bridge->fwnode); ret =acpi_dma_configure (dev,acpi_get_dma_attr (adev)); + }else if (is_vmbus_dev (bridge)) { + ret =hv_dma_configure (dev,device_get_dma_attr (bridge)); }pci_put_host_bridge_device (bridge);
include/linux/hyperv.h +15 -0
@@ -1764,4 +1764,19 @@ static inline unsigned long virt_to_hvpfn(void *addr)#define HVPFN_DOWN (x) ((x) >> HV_HYP_PAGE_SHIFT)#define page_to_hvpfn (page) (page_to_pfn(page) * NR_HV_HYP_PAGES_IN_PAGE) +static inline bool is_vmbus_dev (struct device *dev) +{ +/* +* dev->bus == &hv_bus would break when the caller is built-in +* and CONFIG_HYPERV=m, so look for it by name instead. +*/ +return !strcmp (dev->bus->name,"vmbus" ); +} + +static inline int hv_dma_configure (struct device *dev,enum dev_dma_attr attr) +{ +arch_setup_dma_ops (dev,0 ,0 ,NULL , attr ==DEV_DMA_COHERENT ); +return 0 ; +} +#endif /* _HYPERV_H */
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
RE: [PATCH 4/4 RESEND] PCI: hv: Propagate coherence from VMbus device to PCI device
From: Robin Murphy <[email protected]> Sent: Friday, March 18, 2022 3:58 AM
Quoted 69 lines · Michael Kelley (LINUX)
Quoted 1 line
From: Robin Murphy <[email protected]> Sent: Thursday, March 17, 2022
10:15 AM
Quoted 65 lines
Quoted 64 lines · Michael Kelley via iommu
drivers/pci/controller/pci-hyperv.c static void hv_pci_remove_slots(struct hv_pcibus_device
Quoted 29 lines
PCI pass-thru devices in a Hyper-V VM are represented as a VMBus device and as a PCI device. The coherence of the VMbus device is set based on the VMbus node in ACPI, but the PCI device has no ACPI node and defaults to not hardware coherent. This results in extra software coherence management overhead on ARM64 when devices are hardware coherent.
Fix this by propagating the coherence of the VMbus device to the PCI device. There's no effect on x86/x64 where devices are always hardware coherent.
Signed-off-by Michael Kelley <[email protected]>--- drivers/pci/controller/pci-hyperv.c | 17 +++++++++++++---- 1 file changed, 13 insertions(+), 4 deletions(-)drivers/pci/controller/pci-hyperv.c +1
-0@@ -49,6 +49,7 @@#include <linux/refcount.h> #include <linux/irqdomain.h> #include <linux/acpi.h> +#include <linux/dma-map-ops.h> #include <asm/mshyperv.h> /* @@ -2142,9 +2143,9 @@ static void hv_pci_remove_slots(struct hv_pcibus_device*hbus)
drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct
Quoted 12 lines
drivers/pci/controller/pci-hyperv.c +2
-2}/* -* Set NUMA node for the devices on the bus +* Set NUMA node and DMA coherence for the devices on the bus */ -static void hv_pci_assign_numa_node (struct hv_pcibus_device *hbus) +static void hv_pci_assign_properties (struct hv_pcibus_device *hbus) {struct pci_dev *dev;struct pci_bus *bus = hbus->bridge->bus;@@ -2167,6 +2168,14 @@ static void hv_pci_assign_numa_node(struct
hv_pcibus_device*hbus)drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct · line 2177
Quoted 10 lines
drivers/pci/controller/pci-hyperv.c +7
-0numa_map_to_online_node ( hv_dev->desc.virtual_numa_node)); +/* +* On ARM64, propagate the DMA coherence from the VMbus device +* to the corresponding PCI device. On x86/x64, these calls +* have no effect because DMA is always hardware coherent. +*/ +dev_set_dma_coherent (&dev->dev, +dev_is_dma_coherent (&hbus->hdev->device));Eww... if you really have to do this, I'd prefer to see a proper
hv_dma_configure()helper implemented and wired up topci_dma_configure(). Although since it's a generic property I guess at worstpci_dma_configurecould perhaps propagate coherency from the host bridge to its children by itself in the absence of any other firmware info. And it's built-in so could usearch_setup_dma_ops()like everyone else.
Quoted 8 lines
Quoted 3 lines
I'm not seeing an existing mechanism to provide a "helper" or override
ofpci_dma_configure(). Could you elaborate? Or is this something
that needs to be created?I mean something like the diff below (other #includes omitted for clarity). Essentially if VMBus has its own way of describing parts of the system, then for those parts it's nice if it could fit into the same abstractions we use for firmware-based system description.
OK, got it. Adding the VMbus special case into pci_dma_configure() would work. But after investigating further, let me throw out two other possible solutions as well.
1) It's easy to make the top-level VMbus node in the DSDT be the ACPI companion for the pci_host bridge. The function pcibios_root_bridge_prepare() will do this if the pci-hyperv.c driver sets sysdata->parent to that top-level VMbus node. Then pci_dma_configure()works as is. Also, commit 7d40c0f70 that special cases pcibios_root_bridge_prepare() for Hyper-V becomes
unnecessary. I've coded this approach and it seems to work, but
I don't know all the implications of making that VMbus node be the ACPI companion, potentially for multiple pci_host bridges.
2) Since there's no vIOMMU and we don't know how it will be described if there should be one in the future, it's a bit premature to try to set things up "correctly" now to handle it. A simple approach is to set dma_default_coherent to true if the top-level VMbus node in the DSDT indicates coherent. No other changes are needed. If there's a vIOMMU at some later time, then we add the use of arch_setup_dma_ops() for each device.
Any thoughts on these two approaches vs. adding the VMbus special case into pci_dma_configure()? My preference would be to avoid adding that special case if one of the other two approaches is reasonable.
Michael
Quoted 52 lines
Cheers,
Robin.----->8-----
drivers/pci/pci-driver.c +3
-0@@ -20,6 +20,7 @@#include <linux/of_device.h> #include <linux/acpi.h> #include <linux/dma-map-ops.h> +#include <linux/hyperv.h> #include "pci.h" #include "pcie/portdrv.h" @@ -1602,6 +1603,8 @@ static int pci_dma_configure(struct device *dev)struct acpi_device *adev =to_acpi_device_node (bridge->fwnode); ret =acpi_dma_configure (dev,acpi_get_dma_attr (adev)); + }else if (is_vmbus_dev (bridge)) { + ret =hv_dma_configure (dev,device_get_dma_attr (bridge)); }pci_put_host_bridge_device (bridge);include/linux/hyperv.h +10
-0@@ -1764,4 +1764,19 @@ static inline unsigned long virt_to_hvpfn(void *addr)#define HVPFN_DOWN (x) ((x) >> HV_HYP_PAGE_SHIFT)#define page_to_hvpfn (page) (page_to_pfn(page) * NR_HV_HYP_PAGES_IN_PAGE) +static inline bool is_vmbus_dev(struct device *dev) +{ +/* +* dev->bus == &hv_bus would break when the caller is built-in +* and CONFIG_HYPERV=m, so look for it by name instead. +*/ +return !strcmp (dev->bus->name,"vmbus" ); +} + +static inline int hv_dma_configure(struct device *dev, enum
dev_dma_attrattr)include/linux/hyperv.h +5
-0+{ +arch_setup_dma_ops (dev,0 ,0 ,NULL , attr ==DEV_DMA_COHERENT ); +return 0 ; +} +#endif /* _HYPERV_H */