x86/pci: Initial commit for new VMD device driver
Subsystems (4)
RFC PATCH v2 1/1 x86/pci: Initial commit for new VMD device driver
The Intel Volume Management Device (VMD) is an integrated endpoint on the platform's PCIe root complex that acts as a host bridge to a secondary PCIe domain. BIOS can reassign one or more root ports to appear within a VMD domain instead of the primary domain.
This driver enumerates and enables the domain using the root bus configuration interface provided by the PCI subsystem. The driver provides configuration space accessor functions (pci_ops), bus and memory resources, a chained MSI irq domain, irq_chip implementation, and dma operations necessary to support the domain through the VMD endpoint's interface.
VMD routes I/O as follows:
1) Configuration Space: BAR 0 ("CFGBAR") of VMD provides the base address and size for configuration space register access to VMD-owned root ports. It works similarly to MMCONFIG for extended configuration space. Bus numbering is independent and does not conflict with the primary domain.
2) MMIO Space: BARs 2 and 4 ("MEMBAR1" and "MEMBAR2") of VMD provide the base address, size, and type for MMIO register access. These addresses are not translated by VMD hardware; they are simply reservations to be distributed to root ports' memory base/limit registers and subdivided among devices downstream.
3) DMA: To interact appropriately with IOMMU, the source ID DMA read and write requests are translated to the bus-device-function of the VMD endpoint. Otherwise, DMA operates normally without VMD-specific address translation.
4) Interrupts: Part of VMD's BAR 4 is reserved for VMD's MSI-X Table and PBA. MSIs from VMD domain devices and ports are remapped to appear if they were issued using one of VMD's MSI-X table entries. Each MSI and MSI-X addresses of VMD-owned devices and ports have a special format where the address refers specific entries in VMD's MSI-X table. As with DMA, the interrupt source id is translated to VMD's bus-device-function.
The driver provides its own MSI and MSI-X configuration functions specific to how MSI messages are used within the VMD domain, and it provides an irq_chip for independent IRQ allocation and to relay interrupts from VMD's interrupt handler to the appropriate device driver's handler.
5) Errors: PCIe error message are intercepted by the root ports normally (e.g. AER), except with VMD, system errors (i.e. firmware first) are disabled by default. AER and hotplug interrupts are translated in the same way as endpoint interrupts.
6) VMD does not support INTx interrupts or IO ports. Devices or drivers requiring these features should either not be placed below VMD-owned root ports, or VMD should be disabled by BIOS for such endpoints.
Contributers to this patch include:
Artur Paszkiewicz <[email protected]> Bryan Veal <[email protected]>
Jon Derrick <[email protected]>
---
v1 -> v2:
The original RFC used custom x86_msi_ops to provide the VMD device specific interrupt setup. This was rejected in favor of a chained irq domain hierarchy, so this version provides that. While it tests out successfully in the limited capacity that I can test this, I honestly don't understand completely how this works, so thank you to Jiang Liu for the guidance!
Perhaps I'm missing a callback, but I don't see how the driver can limit the number of irq's requested with the irq domain way. The allocation is done one at a time instead of at once, so the driver doesn't know at this level how many were originally requested. This isn't terrible as I can circle the irq's back to the beginning if they exceed VMD's MSI-x count.
This version includes the DMA operations required if an IOMMU is used. That feature was omitted from the original RFC. The dma operations are set via a PCI "fixup" if the device is in a VMD provided domain.
All this created a larger in-kernel dependency than before, and it is submitted as a single patch instead of a short series since it is all specific to this driver.
arch/x86/Kconfig | 15 ++ arch/x86/include/asm/vmd.h | 39 +++ arch/x86/kernel/apic/msi.c | 74 ++++++ arch/x86/pci/Makefile | 2 + arch/x86/pci/vmd.c | 594 ++++++++++++++++++++++++++++++++++++++++++++ kernel/irq/chip.c | 1 + kernel/irq/irqdomain.c | 3 + 7 files changed, 728 insertions(+) create mode 100644 arch/x86/include/asm/vmd.h create mode 100644 arch/x86/pci/vmd.c
arch/x86/Kconfig +15 -0
@@ -2631,6 +2631,21 @@ config PMC_ATOM def_bool y depends on PCI +config HAVE_VMDDEV + bool + +config VMDDEV + depends on PCI && PCI_DOMAINS && PCI_MSI && GENERIC_MSI_IRQ_DOMAIN && IRQ_DOMAIN_HIERARCHY + tristate "Volume Management Device Driver" + default N + select HAVE_VMDDEV + ---help--- + Adds support for the Intel Volume Manage Device (VMD). VMD is + a secondary PCI host bridge that allows PCI Express root ports, + and devices attached to them, to be removed from the default PCI + domain and placed within the VMD domain. If your system provides + one of these and has devices attached to it, say "Y". + source "net/Kconfig" source "drivers/Kconfig"
arch/x86/include/asm/vmd.h new file +39 -0
@@ -0,0 +1,39 @@ +#ifndef _ASM_X86_VMD_H +#define _ASM_X86_VMD_H + +#ifdef CONFIG_HAVE_VMDDEV +#include <linux/list.h> +struct vmd_irq_list ; + +struct vmd_dev { +struct pci_dev *dev; + +spinlock_t cfg_lock; +char __iomem *cfgbar; + +int msix_count; +struct msix_entry *msix_entries; +struct vmd_irq_list *irqs; + +struct pci_sysdata sysdata; +struct pci_bus *bus; +struct irq_domain *irq_domain; + +struct list_head node; +#ifdef CONFIG_X86_DEV_DMA_OPS +struct dma_map_ops dma_ops; +#endif +}; + +struct irq_domain *arch_create_vmd_msi_irq_domain (struct pci_dev *dev, +const struct irq_domain_ops *domain_ops); +#ifdef CONFIG_X86_DEV_DMA_OPS +void vmd_add_dma_device (struct vmd_dev *vmd); +void vmd_del_dma_device (struct vmd_dev *vmd); +#else +static void vmd_add_dma_device (struct vmd_dev *vmd) {} +static void vmd_del_dma_device (struct vmd_dev *vmd) {} +#endif + +#endif +#endif /* _ASM_X86_VMD_H */
arch/x86/kernel/apic/msi.c +74 -0
86 lines not on this page · raw message
arch/x86/pci/Makefile +2 -0
@@ -23,6 +23,8 @@ obj-y += bus_numa.o obj-$(CONFIG_AMD_NB) += amd_bus.o obj-$(CONFIG_PCI_CNB20LE_QUIRK) += broadcom_bus.o +obj-$(CONFIG_VMDDEV) += vmd.o + ifeq ($(CONFIG_PCI_DEBUG),y) EXTRA_CFLAGS += -DDEBUG endif
arch/x86/pci/vmd.c new file +594 -0
594 lines not on this page · raw message
kernel/irq/chip.c +1 -0
@@ -1062,3 +1062,4 @@ int irq_chip_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)return 0 ; } +EXPORT_SYMBOL_GPL (irq_chip_compose_msi_msg);
kernel/irq/irqdomain.c +3 -0
@@ -910,6 +910,7 @@ struct irq_data *irq_domain_get_irq_data(struct irq_domain *domain,return NULL ; } +EXPORT_SYMBOL_GPL (irq_domain_get_irq_data);/** * irq_domain_set_hwirq_and_chip - Set hwirq and irqchip of @virq at @domain
@@ -934,6 +935,7 @@ int irq_domain_set_hwirq_and_chip(struct irq_domain *domain, unsigned int virq,return 0 ; } +EXPORT_SYMBOL_GPL (irq_domain_set_hwirq_and_chip);/** * irq_domain_set_info - Set the complete data for a @virq in @domain
@@ -1240,6 +1242,7 @@ struct irq_data *irq_domain_get_irq_data(struct irq_domain *domain,return (irq_data && irq_data->domain == domain) ? irq_data :NULL ; } +EXPORT_SYMBOL_GPL (irq_domain_get_irq_data);/** * irq_domain_set_info - Set the complete data for a @virq in @domain
Signature · 1 line
1.7.10.4
arch/x86/Kconfig config PMC_ATOM · line 2647
Quoted 11 lines · Keith Busch
arch/x86/Kconfig +11
-0+config VMDDEV + depends on PCI && PCI_DOMAINS && PCI_MSI && GENERIC_MSI_IRQ_DOMAIN && IRQ_DOMAIN_HIERARCHY + tristate "Volume Management Device Driver" + default N + select HAVE_VMDDEV + ---help--- + Adds support for the Intel Volume Manage Device (VMD). VMD is + a secondary PCI host bridge that allows PCI Express root ports, + and devices attached to them, to be removed from the default PCI + domain and placed within the VMD domain. If your system provides + one of these and has devices attached to it, say "Y".
So what this text does not explain is why does the user care? What tangible benefits does this feature offer to users?
Thanks,
Ingo
Quoted 15 lines · Ingo Molnar
arch/x86/Kconfig config PMC_ATOM · line 2647
Quoted 11 lines · Keith Busch
arch/x86/Kconfig +11
-0+config VMDDEV + depends on PCI && PCI_DOMAINS && PCI_MSI && GENERIC_MSI_IRQ_DOMAIN && IRQ_DOMAIN_HIERARCHY + tristate "Volume Management Device Driver" + default N + select HAVE_VMDDEV + ---help--- + Adds support for the Intel Volume Manage Device (VMD). VMD is + a secondary PCI host bridge that allows PCI Express root ports, + and devices attached to them, to be removed from the default PCI + domain and placed within the VMD domain. If your system provides + one of these and has devices attached to it, say "Y".So what this text does not explain is why does the user care? What tangible benefits does this feature offer to users?
Hi Ingo,
The immediate benefit is that devices on VMD domains do not use resources on the default PCI domain, so we have more than the 256 buses available.
Quoted 17 lines · Keith Busch
Quoted 15 lines · Ingo Molnar
arch/x86/Kconfig config PMC_ATOM · line 2647
Quoted 11 lines · Keith Busch
arch/x86/Kconfig +11
-0+config VMDDEV + depends on PCI && PCI_DOMAINS && PCI_MSI && GENERIC_MSI_IRQ_DOMAIN && IRQ_DOMAIN_HIERARCHY + tristate "Volume Management Device Driver" + default N + select HAVE_VMDDEV + ---help--- + Adds support for the Intel Volume Manage Device (VMD). VMD is + a secondary PCI host bridge that allows PCI Express root ports, + and devices attached to them, to be removed from the default PCI + domain and placed within the VMD domain. If your system provides + one of these and has devices attached to it, say "Y".So what this text does not explain is why does the user care? What tangible benefits does this feature offer to users?
Quoted 4 lines
Hi Ingo,
The immediate benefit is that devices on VMD domains do not use resources on the default PCI domain, so we have more than the 256 buses available.
Would be nice to incorporate that information in the help text and in the changelog.
Thanks,
Ingo
Quoted 6 lines · Ingo Molnar
Quoted 2 lines · Keith Busch
The immediate benefit is that devices on VMD domains do not use resources on the default PCI domain, so we have more than the 256 buses available.
Would be nice to incorporate that information in the help text and in the changelog.
Fair enough. Perhaps we thought the "secondary domain" description implied that information, but maybe we've been doing this too long. :)
Thanks for the feedback.
kbuild test robot
build botconfig: i386-randconfig-h1-10040721 (attached as .config) reproduce:
Hi Keith,
[auto build test results on v4.3-rc3 -- if it's inappropriate base, please ignore]
config: i386-randconfig-h1-10040721 (attached as .config) reproduce:
# save the attached .config to linux build tree
make ARCH=i386 All warnings (new ones prefixed by >>):
In file included from arch/x86/pci/vmd.c:25:0:
>> arch/x86/include/asm/vmd.h:34:13: warning: 'vmd_add_dma_device' defined but not used [-Wunused-function]
static void vmd_add_dma_device(struct vmd_dev *vmd) {}
^
>> arch/x86/include/asm/vmd.h:35:13: warning: 'vmd_del_dma_device' defined but not used [-Wunused-function]
static void vmd_del_dma_device(struct vmd_dev *vmd) {}
^vim +/vmd_add_dma_device +34 arch/x86/include/asm/vmd.h
28 struct irq_domain *arch_create_vmd_msi_irq_domain(struct pci_dev *dev,
29 const struct irq_domain_ops *domain_ops);
30 #ifdef CONFIG_X86_DEV_DMA_OPS
31 void vmd_add_dma_device(struct vmd_dev *vmd);
32 void vmd_del_dma_device(struct vmd_dev *vmd);
33 #else
> 34 static void vmd_add_dma_device(struct vmd_dev *vmd) {}
> 35 static void vmd_del_dma_device(struct vmd_dev *vmd) {}
36 #endif
37
38 #endif---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
Hi Keith,
I have a few comments scattered below.
Quoted 227 lines · Keith Busch
The Intel Volume Management Device (VMD) is an integrated endpoint on the platform's PCIe root complex that acts as a host bridge to a secondary PCIe domain. BIOS can reassign one or more root ports to appear within a VMD domain instead of the primary domain.
This driver enumerates and enables the domain using the root bus configuration interface provided by the PCI subsystem. The driver provides configuration space accessor functions (
pci_ops), bus and memory resources, a chained MSI irq domain,irq_chipimplementation, and dma operations necessary to support the domain through the VMD endpoint's interface.VMD routes I/O as follows:
1) Configuration Space: BAR 0 ("CFGBAR") of VMD provides the base address and size for configuration space register access to VMD-owned root ports. It works similarly to MMCONFIG for extended configuration space. Bus numbering is independent and does not conflict with the primary domain.
2) MMIO Space: BARs 2 and 4 ("MEMBAR1" and "MEMBAR2") of VMD provide the base address, size, and type for MMIO register access. These addresses are not translated by VMD hardware; they are simply reservations to be distributed to root ports' memory base/limit registers and subdivided among devices downstream.
3) DMA: To interact appropriately with IOMMU, the source ID DMA read and write requests are translated to the bus-device-function of the VMD endpoint. Otherwise, DMA operates normally without VMD-specific address translation.
4) Interrupts: Part of VMD's BAR 4 is reserved for VMD's MSI-X Table and PBA. MSIs from VMD domain devices and ports are remapped to appear if they were issued using one of VMD's MSI-X table entries. Each MSI and MSI-X addresses of VMD-owned devices and ports have a special format where the address refers specific entries in VMD's MSI-X table. As with DMA, the interrupt source id is translated to VMD's bus-device-function.
The driver provides its own MSI and MSI-X configuration functions specific to how MSI messages are used within the VMD domain, and it provides an
irq_chipfor independent IRQ allocation and to relay interrupts from VMD's interrupt handler to the appropriate device driver's handler.5) Errors: PCIe error message are intercepted by the root ports normally (e.g. AER), except with VMD, system errors (i.e. firmware first) are disabled by default. AER and hotplug interrupts are translated in the same way as endpoint interrupts.
6) VMD does not support INTx interrupts or IO ports. Devices or drivers requiring these features should either not be placed below VMD-owned root ports, or VMD should be disabled by BIOS for such endpoints.
Contributers to this patch include:
Artur Paszkiewicz <[email protected]> Bryan Veal <[email protected]>
Jon Derrick <[email protected]>Signed-off-by Keith Busch <[email protected]>---
v1 -> v2:The original RFC used custom
x86_msi_opsto provide the VMD device specific interrupt setup. This was rejected in favor of a chained irq domain hierarchy, so this version provides that. While it tests out successfully in the limited capacity that I can test this, I honestly don't understand completely how this works, so thank you to Jiang Liu for the guidance!Perhaps I'm missing a callback, but I don't see how the driver can limit the number of irq's requested with the irq domain way. The allocation is done one at a time instead of at once, so the driver doesn't know at this level how many were originally requested. This isn't terrible as I can circle the irq's back to the beginning if they exceed VMD's MSI-x count.
This version includes the DMA operations required if an IOMMU is used. That feature was omitted from the original RFC. The dma operations are set via a PCI "fixup" if the device is in a VMD provided domain.
All this created a larger in-kernel dependency than before, and it is submitted as a single patch instead of a short series since it is all specific to this driver.
arch/x86/Kconfig | 15 ++ arch/x86/include/asm/vmd.h | 39 +++ arch/x86/kernel/apic/msi.c | 74 ++++++ arch/x86/pci/Makefile | 2 + arch/x86/pci/vmd.c | 594 ++++++++++++++++++++++++++++++++++++++++++++ kernel/irq/chip.c | 1 + kernel/irq/irqdomain.c | 3 + 7 files changed, 728 insertions(+) create mode 100644 arch/x86/include/asm/vmd.h create mode 100644 arch/x86/pci/vmd.carch/x86/Kconfig +15
-0@@ -2631,6 +2631,21 @@ config PMC_ATOM def_bool y depends on PCI +config HAVE_VMDDEV + bool + +config VMDDEV + depends on PCI && PCI_DOMAINS && PCI_MSI && GENERIC_MSI_IRQ_DOMAIN && IRQ_DOMAIN_HIERARCHY + tristate "Volume Management Device Driver" + default N + select HAVE_VMDDEV + ---help--- + Adds support for the Intel Volume Manage Device (VMD). VMD is + a secondary PCI host bridge that allows PCI Express root ports, + and devices attached to them, to be removed from the default PCI + domain and placed within the VMD domain. If your system provides + one of these and has devices attached to it, say "Y". + source "net/Kconfig" source "drivers/Kconfig"arch/x86/include/asm/vmd.h new file +39
-0@@ -0,0 +1,39 @@ +#ifndef _ASM_X86_VMD_H +#define _ASM_X86_VMD_H + +#ifdef CONFIG_HAVE_VMDDEV +#include <linux/list.h> +struct vmd_irq_list ; + +struct vmd_dev { +struct pci_dev *dev; + +spinlock_t cfg_lock; +char __iomem *cfgbar; + +int msix_count; +struct msix_entry *msix_entries; +struct vmd_irq_list *irqs; + +struct pci_sysdata sysdata; +struct pci_bus *bus; +struct irq_domain *irq_domain; + +struct list_head node; +#ifdef CONFIG_X86_DEV_DMA_OPS +struct dma_map_ops dma_ops; +#endif +}; + +struct irq_domain *arch_create_vmd_msi_irq_domain (struct pci_dev *dev, +const struct irq_domain_ops *domain_ops); +#ifdef CONFIG_X86_DEV_DMA_OPS +void vmd_add_dma_device (struct vmd_dev *vmd); +void vmd_del_dma_device (struct vmd_dev *vmd); +#else +static void vmd_add_dma_device (struct vmd_dev *vmd) {} +static void vmd_del_dma_device (struct vmd_dev *vmd) {} +#endif + +#endif +#endif /* _ASM_X86_VMD_H */ arch/x86/kernel/apic/msi.c +45
-0@@ -22,6 +22,7 @@#include <asm/hw_irq.h> #include <asm/apic.h> #include <asm/irq_remapping.h> +#include <asm/vmd.h> static struct irq_domain *msi_default_domain;@@ -134,6 +135,79 @@ static struct msi_domain_info pci_msi_domain_info = { .handler_name = "edge", }; +#ifdef CONFIG_HAVE_VMDDEV +static struct irq_chip pci_chained_msi_controller = { + .name = "PCI-MSI-chained", +}; + +static struct msi_domain_info pci_chained_msi_domain_info = { + .flags = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS | + MSI_FLAG_PCI_MSIX, + .ops = &pci_msi_domain_ops, + .chip = &pci_chained_msi_controller, +}; + +struct irq_domain *arch_create_vmd_msi_irq_domain(struct pci_dev *dev, + const struct irq_domain_ops *domain_ops) +{ + int count = 0; + struct msi_desc *msi_desc; + struct irq_domain *vmd_irqdomain, *msi_irqdomain; + + if (!dev->msix_enabled) + return NULL; + + for_each_pci_msi_entry(msi_desc, dev) + count += msi_desc->nvec_used; + vmd_irqdomain = irq_domain_add_linear(NULL, count, + domain_ops, dev); + if (!vmd_irqdomain) + return NULL; + msi_irqdomain = pci_msi_create_irq_domain(NULL, &pci_chained_msi_domain_info, + vmd_irqdomain); + if (!msi_irqdomain) + irq_domain_remove(vmd_irqdomain); + return msi_irqdomain; +} +EXPORT_SYMBOL_GPL(arch_create_vmd_msi_irq_domain); + +#ifdef CONFIG_X86_DEV_DMA_OPS +DEFINE_SPINLOCK(vmd_dev_list_lock); +LIST_HEAD(vmd_dma_list); + +static void vmd_dma_set_ops(struct pci_dev *pdev) +{ + struct vmd_dev *vmd; +
arch/x86/kernel/apic/msi.c static struct msi_domain_info pci_msi_domain_info = { · line 191
Quoted 10 lines
arch/x86/kernel/apic/msi.c +10
-0+spin_lock (&vmd_dev_list_lock); +list_for_each_entry (vmd, &vmd_dma_list, node) { +if (pci_domain_nr (pdev->bus) == vmd->sysdata.domain) { + pdev->dev.archdata.dma_ops = &vmd->dma_ops; +break ; + } + } +spin_unlock (&vmd_dev_list_lock); +} +DECLARE_PCI_FIXUP_ENABLE (PCI_ANY_ID ,PCI_ANY_ID , vmd_dma_set_ops);
This seems like sort of a sledgehammer approach. We already have five different places where we set up dev->archdata.dma_ops, and this adds a sixth:
pci_device_add
of_pci_dma_configure
of_dma_configure
arch_setup_dma_ops
dev->archdata.dma_ops = ... # (1) arm64
pcibios_add_device
pcibios_setup_device
set_dma_ops
dev->archdata.dma_ops = ... # (2) powerpc
device_add
platform_notify
acpi_platform_notify
acpi_bind_one
arch_setup_dma_ops
dev->archdata.dma_ops = ... # (3) arm64 pci_enable_device
pci_enable_device_flags
do_pci_enable_device
pcibios_enable_device
pcibios_plat_dev_init
dev->dev.archdata.dma_ops = ... # (4) mips
pci_fixup_device(pci_fixup_enable, dev)
vmd_dma_set_ops
pdev->dev.archdata.dma_ops = ... # (6) x86 VMD calgary_init # (5) x86 (calgary)
for_each_pci_dev
dev->dev.archdata.dma_ops = ...We can discard the Calgary approach; that is clearly wrong because it doesn't work at all for hotplug. The pci_device_add() path seems to be the winner, so I'd suggest doing something in pcibios_add_device().
I know it will still be a little ugly because there will be some VMD-specific stuff in the x86 pcibios_add_device(). But somebody is working on replacing some of the pcibios_*() interfaces with a set of host bridge function pointers, and that could end up being a clean way to handle this. You already have a struct pci_host_bridge for the VMD device (allocated implicitly in pci_create_root_bus()), so we could someday pass in VMD-specific host bridge ops that would do the setup.
Quoted 134 lines
arch/x86/kernel/apic/msi.c +19
-0+ +void vmd_add_dma_device (struct vmd_dev *vmd) +{ +spin_lock (&vmd_dev_list_lock); +list_add (&vmd->node, &vmd_dma_list); +spin_unlock (&vmd_dev_list_lock); +} +EXPORT_SYMBOL_GPL (vmd_add_dma_device); + +void vmd_del_dma_device (struct vmd_dev *vmd) +{ +spin_lock (&vmd_dev_list_lock); +list_del_init (&vmd->node); +spin_unlock (&vmd_dev_list_lock); +} +EXPORT_SYMBOL_GPL (vmd_del_dma_device); +#endif +#endif +void arch_init_msi_domain (struct irq_domain *parent) {if (disable_apic)arch/x86/pci/Makefile +2
-0@@ -23,6 +23,8 @@ obj-y += bus_numa.o obj-$(CONFIG_AMD_NB) += amd_bus.o obj-$(CONFIG_PCI_CNB20LE_QUIRK) += broadcom_bus.o +obj-$(CONFIG_VMDDEV) += vmd.o + ifeq ($(CONFIG_PCI_DEBUG),y) EXTRA_CFLAGS += -DDEBUG endifarch/x86/pci/vmd.c new file +93
-0@@ -0,0 +1,594 @@ +/* +* Volume Management Device driver +* Copyright (c) 2015, Intel Corporation. +* +* This program is free software; you can redistribute it and/or modify it +* under the terms and conditions of the GNU General Public License, +* version 2, as published by the Free Software Foundation. +* +* This program is distributed in the hope it will be useful, but WITHOUT +* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or +* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for +* more details. +*/ + +#include <linux/device.h> +#include <linux/interrupt.h> +#include <linux/irq.h> +#include <linux/irqchip/chained_irq.h> +#include <linux/kernel.h> +#include <linux/module.h> +#include <linux/pci.h> +#include <asm/irqdomain.h> +#include <asm/hpet.h> +#include <asm/msidef.h> +#include <asm/vmd.h> + +struct vmd_irq { +struct list_head node; +struct vmd_irq_list *irq; +unsigned int virq; +}; + +struct vmd_irq_list { +struct list_head irq_list; +spinlock_t irq_lock; +unsigned int vmd_vector; +unsigned int index; +}; + +static inline struct vmd_dev *vmd_from_bus (struct pci_bus *bus) +{ + return container_of(bus->sysdata, struct vmd_dev, sysdata); +} + +static void vmd_msi_compose_msg (struct irq_data *data,struct msi_msg *msg) +{ +struct vmd_irq *vmdirq = data->chip_data; + + msg->address_hi =MSI_ADDR_BASE_HI ; + msg->address_lo =MSI_ADDR_BASE_LO |MSI_ADDR_DEST_ID (vmdirq->irq->index); + msg->data =MSI_DATA_VECTOR (vmdirq->irq->vmd_vector); +} + +static void vmd_msi_write_msg (struct irq_data *data,struct msi_msg *msg) +{ +pci_write_msi_msg (data->irq, msg); +} + +/* +* Function is required when using irq hierarchy domains, but we don't have a +* good way to not create conflicts with other devices sharing the same vector. +*/ +int vmd_irq_set_affinity (struct irq_data *data,const struct cpumask *dest, +bool force) +{ +return 0 ; +} + +static struct irq_chip vmd_chip = { + .name ="VMD-MSI" , + .irq_compose_msi_msg = vmd_msi_compose_msg, + .irq_write_msi_msg = vmd_msi_write_msg, + .irq_set_affinity = vmd_irq_set_affinity, +}; + +static void vmd_msi_free_irqs (struct irq_domain *domain, +unsigned int virq,unsigned int nr_irqs) +{ +struct vmd_irq *vmdirq; +struct irq_data *irq_data; + + irq_data =irq_domain_get_irq_data (domain, virq); +if (!irq_data) +return ; + vmdirq = irq_data->chip_data; +kfree (vmdirq); +} + +static int vmd_msi_alloc_irqs (struct irq_domain *domain,unsigned int virq, +unsigned int nr_irqs,void *arg) +{ +struct vmd_dev *vmd; +struct vmd_irq *vmdirq;
arch/x86/pci/vmd.c line 103
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+struct irq_data *irq_data; +struct pci_dev *dev = domain->host_data; + + irq_data =irq_domain_get_irq_data (domain, virq); +if (!irq_data) +return -EINVAL ; + + vmd =pci_get_drvdata (dev); +if (!vmd) +return -EINVAL ;
"vmd == NULL" is impossible, isn't it? If so, remove the test.
Quoted 306 lines
arch/x86/pci/vmd.c new file +306
-0+ + vmdirq =kzalloc (sizeof (*vmdirq),GFP_KERNEL ); +if (!vmdirq) +return -ENOMEM ; + +INIT_LIST_HEAD (&vmdirq->node); + vmdirq->irq = &vmd->irqs[virq % vmd->msix_count]; + vmdirq->virq = virq; + +irq_domain_set_hwirq_and_chip (domain, virq, vmdirq->irq->vmd_vector, &vmd_chip, vmdirq); +irq_set_chip (virq, &vmd_chip); +irq_set_handler_data (virq, vmdirq); +__irq_set_handler (virq, handle_simple_irq,0 ,NULL ); + +return 0 ; +} + +static void vmd_msi_activate (struct irq_domain *domain, +struct irq_data *data) +{ +unsigned long flags; +struct msi_msg msg; +struct vmd_irq *vmdirq = data->chip_data; +struct vmd_irq_list *irq = vmdirq->irq; + +BUG_ON (irq_chip_compose_msi_msg (data, &msg)); + data->chip->irq_write_msi_msg (data, &msg); + +spin_lock_irqsave (&irq->irq_lock, flags); +list_add_tail (&vmdirq->node, &irq->irq_list); +spin_unlock_irqrestore (&irq->irq_lock, flags); +} + +static void vmd_msi_deactivate (struct irq_domain *domain, +struct irq_data *data) +{ +unsigned long flags; +struct msi_msg msg; +struct vmd_irq *vmdirq = data->chip_data; +struct vmd_irq_list *irq = vmdirq->irq; + +memset (&msg,0 ,sizeof (msg)); + data->chip->irq_write_msi_msg (data, &msg); + +spin_lock_irqsave (&irq->irq_lock, flags); +list_del_init (&vmdirq->node); +spin_unlock_irqrestore (&irq->irq_lock, flags); + +} + +static const struct irq_domain_ops vmd_domain_ops = { + .alloc = vmd_msi_alloc_irqs, + .free = vmd_msi_free_irqs, + .activate = vmd_msi_activate, + .deactivate = vmd_msi_deactivate, +}; + +#ifdef CONFIG_X86_DEV_DMA_OPS +static struct device *dev_to_vmd_dev (struct device *dev) +{ +struct pci_dev *pdev =to_pci_dev (dev); +struct vmd_dev *vmd =vmd_from_bus (pdev->bus); +return &vmd->dev->dev; +} + +static void *vmd_alloc (struct device *dev,size_t size,dma_addr_t *addr, +gfp_t flag,struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->alloc (vdev, size, addr, flag, attrs); +} + +static void vmd_free (struct device *dev,size_t size,void *vaddr, +dma_addr_t addr,struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->free (vdev, size, vaddr, addr, attrs); +} + +static int vmd_mmap (struct device *dev,struct vm_area_struct *vma, +void *cpu_addr,dma_addr_t addr, +size_t size,struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->mmap (vdev, vma, cpu_addr, addr, size, + attrs); +} + +static int vmd_get_sgtable (struct device *dev,struct sg_table *sgt, +void *cpu_addr,dma_addr_t addr, +size_t size,struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->get_sgtable (vdev, sgt, cpu_addr, addr, + size, attrs); +} + +static dma_addr_t vmd_map_page (struct device *dev,struct page *page, +unsigned long offset,size_t size, +enum dma_data_direction dir, +struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->map_page (vdev, page, offset, size, dir, + attrs); +} + +static void vmd_unmap_page (struct device *dev,dma_addr_t addr,size_t size, +enum dma_data_direction dir, +struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->unmap_page (vdev, addr, size, dir, attrs); +} + +static int vmd_map_sg (struct device *dev,struct scatterlist *sg,int nents, +enum dma_data_direction dir, +struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->map_sg (dev, sg, nents, dir, attrs); +} + +static void vmd_unmap_sg (struct device *dev,struct scatterlist *sg,int nents, +enum dma_data_direction dir, +struct dma_attrs *attrs) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->unmap_sg (vdev, sg, nents, dir, attrs); +} + +static void vmd_sync_single_for_cpu (struct device *dev,dma_addr_t addr, +size_t size,enum dma_data_direction dir) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->sync_single_for_cpu (vdev, addr, size, dir); +} + +static void vmd_sync_single_for_device (struct device *dev,dma_addr_t addr, +size_t size,enum dma_data_direction dir) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->sync_single_for_device (vdev, addr, size, dir); +} + +static void vmd_sync_sg_for_cpu (struct device *dev,struct scatterlist *sg, +int nents,enum dma_data_direction dir) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->sync_sg_for_cpu (vdev, sg, nents, dir); +} + +static void vmd_sync_sg_for_device (struct device *dev,struct scatterlist *sg, +int nents,enum dma_data_direction dir) +{ +struct device *vdev =dev_to_vmd_dev (dev); + vdev->archdata.dma_ops->sync_sg_for_device (dev, sg, nents, dir); +} + +static int vmd_mapping_error (struct device *dev,dma_addr_t addr) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->mapping_error (vdev, addr); +} + +static int vmd_dma_supported (struct device *dev,u64 mask) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->dma_supported (vdev, mask); +} + +#ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK +u64 vmd_get_required_mask (struct device *dev) +{ +struct device *vdev =dev_to_vmd_dev (dev); +return vdev->archdata.dma_ops->get_required_mask (vdev); +} +#endif +#endif + +static void vmd_teardown_dma_ops (struct vmd_dev *vmd) +{ +#ifdef CONFIG_X86_DEV_DMA_OPS +if (vmd->dev->dev.archdata.dma_ops) +vmd_del_dma_device (vmd); +#endif +} + +static void vmd_setup_dma_ops (struct vmd_dev *vmd) +{ +#ifdef CONFIG_X86_DEV_DMA_OPS +struct dma_map_ops *source = vmd->dev->dev.archdata.dma_ops; +struct dma_map_ops *dest = &vmd->dma_ops; + +if (!source) +return ; +if (source->alloc) + dest->alloc = vmd_alloc; +if (source->free) + dest->free = vmd_free; +if (source->mmap) + dest->mmap = vmd_mmap; +if (source->get_sgtable) + dest->get_sgtable = vmd_get_sgtable; +if (source->map_page) + dest->map_page = vmd_map_page; +if (source->unmap_page) + dest->unmap_page = vmd_unmap_page; +if (source->map_sg) + dest->map_sg = vmd_map_sg; +if (source->unmap_sg) + dest->unmap_sg = vmd_unmap_sg; +if (source->sync_single_for_cpu) + dest->sync_single_for_cpu = vmd_sync_single_for_cpu; +if (source->sync_single_for_device) + dest->sync_single_for_device = vmd_sync_single_for_device; +if (source->sync_sg_for_cpu) + dest->sync_sg_for_cpu = vmd_sync_sg_for_cpu; +if (source->sync_sg_for_device) + dest->sync_sg_for_device = vmd_sync_sg_for_device; +if (source->mapping_error) + dest->mapping_error = vmd_mapping_error; +if (source->dma_supported) + dest->dma_supported = vmd_dma_supported; +#ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK +if (source->get_required_mask) + dest->get_required_mask = vmd_get_required_mask; +#endif +vmd_add_dma_device (vmd); +#endif +} + +/* +* CPU may deadlock if config space is not serialized on some versions of this +* hardware, so all config space access is done under a spinlock. +*/ +static int vmd_pci_read (struct pci_bus *bus,unsigned int devfn,int reg, +int len,u32 *value) +{ +int ret =0 ; +unsigned long flags; +struct vmd_dev *vmd =vmd_from_bus (bus); +char __iomem *addr = vmd->cfgbar + (bus->number <<20 ) + + (devfn <<12 ) + reg; + +spin_lock_irqsave (&vmd->cfg_lock, flags); +switch (len) { +case 1 : + *value =readb (addr); +break ; +case 2 : + *value =readw (addr); +break ; +case 4 : + *value =readl (addr); +break ; +default : + ret = -EINVAL ; +break ; + } +spin_unlock_irqrestore (&vmd->cfg_lock, flags); +return ret; +} + +/* +* VMD h/w converts posted config writes to non-posted. The read-back in this +* function forces the completion so it returns only after the config space was +* written, as expected. +*/ +static int vmd_pci_write (struct pci_bus *bus,unsigned int devfn,int reg, +int len,u32 value) +{ +int ret =0 ; +unsigned long flags; +struct vmd_dev *vmd =vmd_from_bus (bus); +char __iomem *addr = vmd->cfgbar + (bus->number <<20 ) + + (devfn <<12 ) + reg; + +spin_lock_irqsave (&vmd->cfg_lock, flags); +switch (len) { +case 1 : +writeb (value, addr); +readb (addr); +break ; +case 2 : +writew (value, addr); +readw (addr); +break ; +case 4 : +writel (value, addr); +readl (addr); +break ; +default : + ret = -EINVAL ; +break ; + } +spin_unlock_irqrestore (&vmd->cfg_lock, flags); +return ret; +} + +static struct pci_ops vmd_ops = { + .read = vmd_pci_read, + .write = vmd_pci_write, +}; + +static int vmd_find_free_domain (void )
arch/x86/pci/vmd.c line 419
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+{ +int domain =-1 ; +struct pci_bus *bus =NULL ; + +while ((bus =pci_find_next_bus (bus)) !=NULL ) + domain =max_t (int, domain,pci_domain_nr (bus)); +if (domain <0 ) +return -ENODEV ; +return domain +1 ; +}
The PCI core should own domain number allocation. We have a little bit of generic domain support, e.g., pci_get_new_domain_nr(). On x86, I think that is compiled-in, but currently not used -- currently x86 only uses _SEG from ACPI. How would you handle collisions between a domain number assigned here (or via pci_get_new_domain_nr()) and a hot-added host bridge where _SEG uses the same domain?
Quoted 8 lines
arch/x86/pci/vmd.c new file +8
-0+ +static int vmd_enable_domain (struct vmd_dev *vmd) +{ +static const u8 vmd_membars[] = {2 ,4 }; +static const u64 vmd_membar_offsets[] = {0 ,0x2000 }; +int i =0 ; +LIST_HEAD (resources); +struct pci_sysdata *sd = &vmd->sysdata;
arch/x86/pci/vmd.c line 437
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+struct resource_entry *entry; + + sd->domain =vmd_find_free_domain (); +if (sd->domain <0 ) +return sd->domain; + sd->node =pcibus_to_node (vmd->dev->bus); + +pci_add_resource (&resources,NULL ); +pci_add_resource (&resources,NULL ); +pci_add_resource (&resources,NULL );
I don't really like this style of "use pci_add_resource(..., NULL) to preallocate resource structures, then fill them in later" because
a) Nobody else does it that way,
b) We have (transitory) empty resources in the "resources" list, and
c) If there's a failure, e.g., if the kzalloc() in resource_list_create_entry() fails, pci_add_resource() doesn't return an error, and we'll oops with a NULL pointer dereference below.
I'd prefer to just alloc your own resource here, initialize it, then add it.
arch/x86/pci/vmd.c line 445
Quoted 8 lines
arch/x86/pci/vmd.c new file +8
-0+resource_list_for_each_entry (entry, &resources) { +struct resource *source, *resource = entry->res; + +if (!i) { + resource->start =0 ; + resource->end = (resource_size ( + &vmd->dev->resource[0 ]) >>20 ) -1 ; + resource->flags =IORESOURCE_BUS |IORESOURCE_PCI_FIXED ;
I thought BAR0 was CFGBAR. I missed the connection to a bus number aperture.
Quoted 7 lines
arch/x86/pci/vmd.c new file +7
-0+ }else { + source = &vmd->dev->resource[vmd_membars[i -1 ]]; + resource->start = source->start + + vmd_membar_offsets[i -1 ]; + resource->end = source->end; + resource->flags = source->flags & ~IORESOURCE_SIZEALIGN ; + resource->parent = source;
arch/x86/pci/vmd.c line 461
Quoted 10 lines
arch/x86/pci/vmd.c new file +9
-0+if (!upper_32_bits (resource->end)) + resource->flags &= ~IORESOURCE_MEM_64 ; + } + i++; + } + + vmd->irq_domain =arch_create_vmd_msi_irq_domain (vmd->dev, &vmd_domain_ops); +if (!vmd->irq_domain) +return -ENODEV ;
This failure path leaks resource_entry structures, doesn't it?
Quoted 92 lines
arch/x86/pci/vmd.c new file +92
-0+ + vmd->bus =pci_create_root_bus (NULL ,0 , &vmd_ops, sd, &resources); +if (!vmd->bus) { +pci_free_resource_list (&resources); +return -ENODEV ; + } +dev_set_msi_domain (&vmd->bus->dev, vmd->irq_domain); +vmd_setup_dma_ops (vmd); +pci_scan_child_bus (vmd->bus); +pci_bus_size_bridges (vmd->bus); +pci_bus_assign_resources (vmd->bus); +pci_bus_add_devices (vmd->bus); + +WARN (sysfs_create_link (&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,"domain" ), +"Can't create symlink to domain\n" ); +return 0 ; +} + +static void vmd_flow_handler (struct irq_desc *desc) +{ +struct irq_chip *chip =irq_desc_get_chip (desc); +struct vmd_irq_list *irqs =irq_desc_get_handler_data (desc); +struct vmd_irq *vmdirq; + +chained_irq_enter (chip, desc); +spin_lock (&irqs->irq_lock); +list_for_each_entry (vmdirq, &irqs->irq_list, node) +generic_handle_irq (vmdirq->virq); +spin_unlock (&irqs->irq_lock); +chained_irq_exit (chip, desc); +} + +static int vmd_probe (struct pci_dev *dev,const struct pci_device_id *id) +{ +struct vmd_dev *vmd; +int i, err; + + vmd =devm_kzalloc (&dev->dev,sizeof (*vmd),GFP_KERNEL ); +if (!vmd) +return -ENOMEM ; + + err =pcim_enable_device (dev); +if (err <0 ) +return err; + + vmd->cfgbar =pcim_iomap (dev,0 ,0 ); +if (!vmd->cfgbar) +return -ENOMEM ; + +pci_set_master (dev); +if (dma_set_mask_and_coherent (&dev->dev,DMA_BIT_MASK (64 )) && +dma_set_mask_and_coherent (&dev->dev,DMA_BIT_MASK (32 ))) +return -ENODEV ; + + vmd->dev = dev; + vmd->msix_count =pci_msix_vec_count (dev); +if (!vmd->msix_count) +return -ENODEV ; + + vmd->irqs =devm_kcalloc (&dev->dev, vmd->msix_count,sizeof (*vmd->irqs), +GFP_KERNEL ); +if (!vmd->irqs) +return -ENOMEM ; + + vmd->msix_entries =devm_kcalloc (&dev->dev, vmd->msix_count, +sizeof (*vmd->msix_entries),GFP_KERNEL ); +if (!vmd->msix_entries) +return -ENOMEM ; +for (i =0 ; i < vmd->msix_count; i++) + vmd->msix_entries[i].entry = i; + + vmd->msix_count =pci_enable_msix_range (vmd->dev, vmd->msix_entries,1 , + vmd->msix_count); +if (vmd->msix_count <0 ) +return vmd->msix_count; + +for (i =0 ; i < vmd->msix_count; i++) { +INIT_LIST_HEAD (&vmd->irqs[i].irq_list); +spin_lock_init (&vmd->irqs[i].irq_lock); + vmd->irqs[i].vmd_vector = vmd->msix_entries[i].vector; + vmd->irqs[i].index = i; + +irq_set_chained_handler_and_data (vmd->msix_entries[i].vector, + vmd_flow_handler, &vmd->irqs[i]); + } +spin_lock_init (&vmd->cfg_lock); + err =vmd_enable_domain (vmd); +if (err) +return err; + +pci_set_drvdata (dev, vmd); +return 0 ;
arch/x86/pci/vmd.c line 563
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+} + +static void vmd_remove (struct pci_dev *dev) +{ +struct pci_bus *bus; +struct pci_dev *child, *tmp; +struct vmd_dev *vmd =pci_get_drvdata (dev); + +if (!vmd) +return ;
How is it possible to get here with "vmd == NULL"? If it's impossible, remove the test.
Quoted 17 lines
arch/x86/pci/vmd.c new file +17
-0+sysfs_remove_link (&vmd->dev->dev.kobj,"domain" ); +pci_set_drvdata (dev,NULL ); + bus = vmd->bus; +list_for_each_entry_safe (child, tmp, &bus->devices, bus_list) +pci_stop_and_remove_bus_device (child); +pci_remove_bus (bus); +vmd_teardown_dma_ops (vmd); +} + +static const struct pci_device_id vmd_ids[] = { + {PCI_DEVICE (PCI_VENDOR_ID_INTEL ,0x201d ),}, + {0 ,} +}; +MODULE_DEVICE_TABLE (pci, vmd_ids); + +struct pci_driver vmd_drv = { + .name ="vmd" ,
arch/x86/pci/vmd.c line 590
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+ .id_table = vmd_ids, + .probe = vmd_probe, + .remove = vmd_remove, +}; + +static int __init vmd_init (void ) +{ +return pci_register_driver (&vmd_drv); +} +module_init (vmd_init);
module_pci_driver(vmd_drv)?
Quoted 48 lines
arch/x86/pci/vmd.c new file +4
-0+ +MODULE_AUTHOR ("Intel Corporation" ); +MODULE_LICENSE ("GPL v2" ); +MODULE_VERSION ("0.1" );kernel/irq/chip.c +1
-0@@ -1062,3 +1062,4 @@ int irq_chip_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)return 0 ; } +EXPORT_SYMBOL_GPL (irq_chip_compose_msi_msg);kernel/irq/irqdomain.c +3
-0@@ -910,6 +910,7 @@ struct irq_data *irq_domain_get_irq_data(struct irq_domain *domain,return NULL ; } +EXPORT_SYMBOL_GPL (irq_domain_get_irq_data);/** * irq_domain_set_hwirq_and_chip - Set hwirq and irqchip of @virq at @domain @@ -934,6 +935,7 @@ int irq_domain_set_hwirq_and_chip(struct irq_domain *domain, unsigned int virq,return 0 ; } +EXPORT_SYMBOL_GPL (irq_domain_set_hwirq_and_chip);/** * irq_domain_set_info - Set the complete data for a @virq in @domain @@ -1240,6 +1242,7 @@ struct irq_data *irq_domain_get_irq_data(struct irq_domain *domain,return (irq_data && irq_data->domain == domain) ? irq_data :NULL ; } +EXPORT_SYMBOL_GPL (irq_domain_get_irq_data);/** * irq_domain_set_info - Set the complete data for a @virq in @domain Signature · 6 lines
1.7.10.4 -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Bjorn,
Thanks for the feedback. Much of the issues you mentioned look pretty straight forward to resolve, and will fix of for the next revision.
I have some immediate follow up comments to two issues you brought up:
Quoted 10 lines · Bjorn Helgaas
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+static int vmd_find_free_domain (void ) +{ +int domain =-1 ; +struct pci_bus *bus =NULL ; + +while ((bus =pci_find_next_bus (bus)) !=NULL ) + domain =max_t (int, domain,pci_domain_nr (bus)); +if (domain <0 ) +return -ENODEV ; +return domain +1 ;
Quoted 8 lines
Quoted 1 line
+}
The PCI core should own domain number allocation. We have a little bit of generic domain support, e.g.,
pci_get_new_domain_nr(). On x86, I think that is compiled-in, but currently not used -- currently x86 only uses_SEGfrom ACPI. How would you handle collisions between a domain number assigned here (or viapci_get_new_domain_nr()) and a hot-added host bridge where_SEGuses the same domain?
Thank you for bringing this up as I hadn't thought much about it and may have misunderstood the meaning of _SEG. AIUI, it is used to identify a logical grouping. The OS does not need to use the same identifier for the domain, so we there's no need to collide if we can assign the domain of a a new _SEG to the next available domain_nr.
On pci_get_new_domain_nr(), can we make this use something like an ida instead of an atomic? I think we'd like to put domains back into the available pool if someone unbinds it. I imagine someone will test unbinding/rebinding these devices in a loop for a while, and will file a bug after the atomic increments to 64k. :)
Quoted 11 lines
arch/x86/pci/vmd.c line 445
Quoted 8 lines
arch/x86/pci/vmd.c new file +8
-0+resource_list_for_each_entry (entry, &resources) { +struct resource *source, *resource = entry->res; + +if (!i) { + resource->start =0 ; + resource->end = (resource_size ( + &vmd->dev->resource[0 ]) >>20 ) -1 ; + resource->flags =IORESOURCE_BUS |IORESOURCE_PCI_FIXED ;I thought BAR0 was CFGBAR. I missed the connection to a bus number aperture.
Right, BAR0 is the CFGBAR and is the device's aperture to access its domain's config space. Based on your comment, I assume that's not a bus resource, though it seems to work with the posted code. :)
Quoted 14 lines · Keith Busch
Quoted 4 lines · Bjorn Helgaas
Quoted 4 lines
arch/x86/pci/vmd.c new file +4
-0+resource_list_for_each_entry (entry, &resources) { +struct resource *source, *resource = entry->res; + +if (!i) {Quoted 8 lines
Quoted 5 lines
arch/x86/pci/vmd.c new file +4
-0+ resource->start =0 ; + resource->end = (resource_size ( + &vmd->dev->resource[0 ]) >>20 ) -1 ; + resource->flags = IORESOURCE_BUS |
IORESOURCE_PCI_FIXED;I thought BAR0 was CFGBAR. I missed the connection to a bus number aperture.
Quoted 2 lines
Right, BAR0 is the CFGBAR and is the device's aperture to access its domain's config space.
It's a new day, I'll try a new explanation on what this is about. The size of the CFGBAR determines how many bus numbers can be reached through the device's config space aperture. We are not setting the bus resource to BAR0; just determining the bus resource ending based on BAR0's size. We expect the bar to be 256M to access config space for 256 buses:
8 functions * 32 devices * 256 buses * 4k config space per function = 256M
If the BAR wasn't provided 256M for any reason, we reduce the number of bus resources this domain can provide.
Quoted 23 lines · Keith Busch
Quoted 13 lines · Keith Busch
Quoted 11 lines · Bjorn Helgaas
arch/x86/pci/vmd.c line 445
Quoted 8 lines
arch/x86/pci/vmd.c new file +8
-0+resource_list_for_each_entry (entry, &resources) { +struct resource *source, *resource = entry->res; + +if (!i) { + resource->start =0 ; + resource->end = (resource_size ( + &vmd->dev->resource[0 ]) >>20 ) -1 ; + resource->flags =IORESOURCE_BUS |IORESOURCE_PCI_FIXED ;I thought BAR0 was CFGBAR. I missed the connection to a bus number aperture.
Quoted 2 lines
Right, BAR0 is the CFGBAR and is the device's aperture to access its domain's config space.
It's a new day, I'll try a new explanation on what this is about. The size of the CFGBAR determines how many bus numbers can be reached through the device's config space aperture. We are not setting the bus resource to BAR0; just determining the bus resource ending based on BAR0's size. We expect the bar to be 256M to access config space for 256 buses:
Quoted 4 lines
8 functions * 32 devices * 256 buses * 4k config space per function = 256M
If the BAR wasn't provided 256M for any reason, we reduce the number of bus resources this domain can provide.
Yep, that makes sense. A comment here about the fact that BAR0 maps ECAM space for the new domain might help.
Thanks!
Bjorn
Quoted 27 lines · Keith Busch
Hi Bjorn,
Thanks for the feedback. Much of the issues you mentioned look pretty straight forward to resolve, and will fix of for the next revision.
I have some immediate follow up comments to two issues you brought up:
Quoted 10 lines · Bjorn Helgaas
Quoted 10 lines
arch/x86/pci/vmd.c new file +10
-0+static int vmd_find_free_domain (void ) +{ +int domain =-1 ; +struct pci_bus *bus =NULL ; + +while ((bus =pci_find_next_bus (bus)) !=NULL ) + domain =max_t (int, domain,pci_domain_nr (bus)); +if (domain <0 ) +return -ENODEV ; +return domain +1 ;Quoted 8 lines
Quoted 1 line
+}
The PCI core should own domain number allocation. We have a little bit of generic domain support, e.g.,
pci_get_new_domain_nr(). On x86, I think that is compiled-in, but currently not used -- currently x86 only uses_SEGfrom ACPI. How would you handle collisions between a domain number assigned here (or viapci_get_new_domain_nr()) and a hot-added host bridge where_SEGuses the same domain?
Quoted 5 lines
Thank you for bringing this up as I hadn't thought much about it and may have misunderstood the meaning of
_SEG. AIUI, it is used to identify a logical grouping. The OS does not need to use the same identifier for the domain, so we there's no need to collide if we can assign the domain of a a new_SEGto the next availabledomain_nr.
Yes, I guess it would be possible to decouple _SEG and Linux PCI domain numbers. It's *convenient* to have them the same, so dmesg and lspci output matches _SEG directly, but I guess you could argue that's not essential.
I think we'd have to maintain a mapping from domain back to _SEG to deal with firmware interfaces that expect _SEG, e.g., ia64 PAL calls.
Quoted 5 lines
On
pci_get_new_domain_nr(), can we make this use something like an ida instead of an atomic? I think we'd like to put domains back into the available pool if someone unbinds it. I imagine someone will test unbinding/rebinding these devices in a loop for a while, and will file a bug after the atomic increments to 64k. :)
I assume you're talking about the "ID to pointer translation service" in lib/idr.c. That looks like it would make sense to me.
If we add such a mapping, it would be nice if we could attempt to allocate a domain equal to _SEG, and only fall back to a different domain if that fails. Then the mapping would be invisible in the common case of all host bridges being present at boot.
Bjorn
Quoted 7 lines · Bjorn Helgaas
Quoted 5 lines · Keith Busch
Thank you for bringing this up as I hadn't thought much about it and may have misunderstood the meaning of
_SEG. AIUI, it is used to identify a logical grouping. The OS does not need to use the same identifier for the domain, so we there's no need to collide if we can assign the domain of a a new_SEGto the next availabledomain_nr.
Quoted 7 lines
Yes, I guess it would be possible to decouple
_SEGand Linux PCI domain numbers. It's *convenient* to have them the same, so dmesg and lspci output matches_SEGdirectly, but I guess you could argue that's not essential.I think we'd have to maintain a mapping from domain back to
_SEGto deal with firmware interfaces that expect_SEG, e.g., ia64 PAL calls.
It looks like there are lots of assumptions in the kernel that segment and domain are the same thing. I don't have the necessary h/w to test any changes here to verify the mappings are handled correctly, so I'm apprehensive to start changing this much code that I can't test.
Given that domain_nr is a 32-bit integer, APCI's _SEG is only 16 bits, and the pci domain is purely a software construct, do you see any problem if we start these 'special' domains at 0x10000? I've tested this in the driver and lspci + setpci with the single line update in pciutils' lib/pci.h, and it all seems to work just fine.
[+cc Martin]
Quoted 13 lines · Keith Busch
Quoted 12 lines · Bjorn Helgaas
Quoted 5 lines · Keith Busch
Thank you for bringing this up as I hadn't thought much about it and may have misunderstood the meaning of
_SEG. AIUI, it is used to identify a logical grouping. The OS does not need to use the same identifier for the domain, so we there's no need to collide if we can assign the domain of a a new_SEGto the next availabledomain_nr.Yes, I guess it would be possible to decouple
_SEGand Linux PCI domain numbers. It's *convenient* to have them the same, so dmesg and lspci output matches_SEGdirectly, but I guess you could argue that's not essential.
Quoted 7 lines
Quoted 2 lines
I think we'd have to maintain a mapping from domain back to
_SEGto deal with firmware interfaces that expect_SEG, e.g., ia64 PAL calls.It looks like there are lots of assumptions in the kernel that segment and domain are the same thing. I don't have the necessary h/w to test any changes here to verify the mappings are handled correctly, so I'm apprehensive to start changing this much code that I can't test.
Yes, this could be a real can of worms.
Quoted 5 lines
Given that
domain_nris a 32-bit integer, APCI's_SEGis only 16 bits, and the pci domain is purely a software construct, do you see any problem if we start these 'special' domains at0x10000? I've tested this in the driver and lspci + setpci with the single line update in pciutils'lib/pci.h, and it all seems to work just fine.
If you make the kernel start using domain numbers that don't fit in 16 bits, what happens if you run the old lspci on a new kernel?
Bjorn
Quoted 8 lines · Bjorn Helgaas
Quoted 5 lines
Given that
domain_nris a 32-bit integer, APCI's_SEGis only 16 bits, and the pci domain is purely a software construct, do you see any problem if we start these 'special' domains at0x10000? I've tested this in the driver and lspci + setpci with the single line update in pciutils'lib/pci.h, and it all seems to work just fine.If you make the kernel start using domain numbers that don't fit in 16 bits, what happens if you run the old lspci on a new kernel?
If VMD is enabled and enumerated by the driver with this proposal, older versions of pciutils will not work; lspci will exit with error status for being unable to find the sysfs entry for devices in that domain since the upper bits are truncated in the down cast.
Assuming we can change pciutils to use 'int' for domain, and have a driver for this new device merged in the near future, I'm not too worried about users and distros with older utils. I don't think this a device someone could turn on by accident, so their shouldn't be any fallout if I properly communicate the dependencies to those who do want to enable it.
Hi!
Quoted 5 lines
Assuming we can change pciutils to use 'int' for domain, and have a driver for this new device merged in the near future, I'm not too worried about users and distros with older utils. I don't think this a device someone could turn on by accident, so their shouldn't be any fallout if I properly communicate the dependencies to those who do want to enable it.
I would prefer using an explicit 32-bit type instead of an int, but otherwise I am fine with the change.
We probably should keep the same naming convention of sysfs entries on buses whose domain number fits in 16 bits and use 8-digit domain number only in cases where it is really needed.
Martin
Quoted 7 lines · Bjorn Helgaas
arch/x86/pci/vmd.c line 590
Quoted 5 lines
arch/x86/pci/vmd.c new file +5
-0+static int __init vmd_init (void ) +{ +return pci_register_driver (&vmd_drv); +} +module_init (vmd_init);
module_pci_driver(vmd_drv)?
We actually only have a module_init in this driver, and purposely left out module_exit. We don't want to be able to unload this because we can't reference count this module for modules depending on it.
'rmmod <module>' returns busy if you're actively using a device that module is driving, but the end devices VMD provides are not driven by the VMD driver; it'd be kind of like being able to unload the pci core driver.