hw/arm/virt: Fix PL061 node name and properties
PATCH 1/1 hw/arm/virt: Fix PL061 node name and properties
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.
---
Split off from "[PATCH QEMU v2 2/5] ARM: PL061: Extract pl061_create_fdt()"
(https://lore.kernel.org/r/[email protected]).
--- hw/arm/virt.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
hw/arm/virt.c +3 -1
@@ -818,13 +818,15 @@ static void create_gpio(const VirtMachineState *vms) qdev_get_gpio_in(vms->gic, irq));uint32_t phandle =qemu_fdt_alloc_phandle (vms->fdt); - nodename =g_strdup_printf ("/pl061@%" PRIx64, base); + nodename =g_strdup_printf ("/gpio@%" PRIx64, base);qemu_fdt_add_subnode (vms->fdt, nodename);qemu_fdt_setprop_sized_cells (vms->fdt, nodename,"reg" ,2 , base,2 , size);qemu_fdt_setprop (vms->fdt, nodename,"compatible" , compat,sizeof (compat));qemu_fdt_setprop_cell (vms->fdt, nodename,"#gpio-cells" ,2 );qemu_fdt_setprop (vms->fdt, nodename,"gpio-controller" ,NULL ,0 ); +qemu_fdt_setprop_cell (vms->fdt, nodename,"#interrupt-cells" ,2 ); +qemu_fdt_setprop (vms->fdt, nodename,"interrupt-controller" ,NULL ,0 );qemu_fdt_setprop_cells (vms->fdt, nodename,"interrupts" ,GIC_FDT_IRQ_TYPE_SPI , irq,GIC_FDT_IRQ_FLAGS_LEVEL_HI );
Signature · 1 line
2.17.1
Quoted 4 lines · Geert Uytterhoeven
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.
Where have these properties come from? They must be optional, because in the version of the binding documentation from Linux 5.0 they're not described:
https://elixir.bootlin.com/linux/v5.0/source/Documentation/devicetree/bindings/gpio/pl061-gpio.txt
They seem to have magically appeared in kernel commit 910f38bed9439e765f7e, which purports to only be a change of format from plain text to yaml but has added some extra properties and claimed them to be mandatory.
Since the devicetree spec says that the interrupt-controller property "defines a node as an interrupt controller node" and a GPIO chip isn't an interrupt controller, this seems like some kind of error in the dtb binding. Maybe I'm missing something...
What actually goes wrong if QEMU doesn't specify these properties?
hw/arm/virt.c static void create_gpio(const VirtMachineState *vms) · line 821
Quoted 10 lines
hw/arm/virt.c +1
-1@@ -818,13 +818,15 @@ static void create_gpio(const VirtMachineState *vms) qdev_get_gpio_in(vms->gic, irq));uint32_t phandle =qemu_fdt_alloc_phandle (vms->fdt); - nodename =g_strdup_printf ("/pl061@%" PRIx64, base); + nodename =g_strdup_printf ("/gpio@%" PRIx64, base);
Does the devicetree binding really mandate what the node name is? I thought that finding the right device was doe via the 'compatible' string and the nodename could be whatever the device tree creator wanted.
thanks
-- PMM
Hi Peter,
Quoted 11 lines · Peter Maydell
Quoted 4 lines · Geert Uytterhoeven
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.Where have these properties come from? They must be optional, because in the version of the binding documentation from Linux 5.0 they're not described:
https://elixir.bootlin.com/linux/v5.0/source/Documentation/devicetree/bindings/gpio/pl061-gpio.txt
Many old DT bindings are incomplete.
Quoted 4 lines
They seem to have magically appeared in kernel commit
910f38bed9439e765f7e, which purports to only be a change of format from plain text to yaml but has added some extra properties and claimed them to be mandatory.
The main motivation behind the conversion from plain text to yaml is to do automatic validation of DTS files, based on the bindings. During the conversion process, many issues are detected, and fixed; not only in the DTS files, but also in the bindings (e.g. missing properties, like in this case).
When running the validation on a device tree passed to the guest (extracted from /sys/firmware/devicetree/base, converted to dts, and
manually fixed up the phandles), the following is reported about the
pl061 node:
virt.dt.yaml: pl061@9030000: {'reg': [[0, 151191552, 0, 4096]],'gpio-controller': True, 'phandle': [[32771]], '#gpio-cells': [[2]], 'clocks': [[32768]], '#interrupt-cells': [[2]], 'compatible':
['arm,pl061', 'arm,primecell'], 'clock-names': ['apb_pclk'], '$nodename': ['pl061@9030000']} is not valid under any of the given schemas
[...]
virt.dt.yaml: pl061@9030000: 'interrupts' is a required property
virt.dt.yaml: pl061@9030000: $nodename:0: 'pl061@9030000' does notmatch '^gpio@[0-9a-f]+$'
virt.dt.yaml: pl061@9030000: 'interrupt-controller' is a required property
Quoted 5 lines
Since the devicetree spec says that the interrupt-controller property "defines a node as an interrupt controller node" and a GPIO chip isn't an interrupt controller, this seems like some kind of error in the dtb binding. Maybe I'm missing something...
PL061 is an interrupt controller, as it can assert its interrupt output based on activity on GPIO input lines.
Quoted 2 lines
What actually goes wrong if QEMU doesn't specify these properties?
It means that other devices that have their interrupt output connected to a PL061 GPIO input won't work, as their driver in the guest OS cannot find the interrupt. Note that arm/virt.c currently doesn't instantiate such devices.
Quoted 7 lines
Quoted 7 lines
hw/arm/virt.c +0
-0@@ -818,13 +818,15 @@ static void create_gpio(const VirtMachineState *vms) qdev_get_gpio_in(vms->gic, irq));
Quoted 8 lines
hw/arm/virt.c static void create_gpio(const VirtMachineState *vms) · line 821
Quoted 3 lines
hw/arm/virt.c +1
-1uint32_t phandle =qemu_fdt_alloc_phandle (vms->fdt); - nodename =g_strdup_printf ("/pl061@%" PRIx64, base); + nodename =g_strdup_printf ("/gpio@%" PRIx64, base);Does the devicetree binding really mandate what the node name is? I thought that finding the right device was doe via the 'compatible' string and the nodename could be whatever the device tree creator wanted.
Matching is indeed done based on compatible value.
For node names, please see
https://github.com/devicetree-org/devicetree-specification/releases/tag/v0.3
"2.2.2 Generic Names Recommendation
The name of a node should be somewhat generic, reflecting the function of the device and not its precise programming model. If appropriate, the name should be one of the following choices:
[...]
- gpio
[...]"
While many new generic names have been added recently, "gpio" was already documented in the Power.orgTM Standard for Embedded Power ArchitectureTM Platform Requirements (ePAPR).
I hope the above explains the rationale behind these change better. Thanks!
Gr{oetje,eeting}s,
Geert
Signature · 5 lines
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected] In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
Quoted 11 lines · Geert Uytterhoeven
Hi Peter,
Quoted 7 lines · Peter Maydell
Quoted 4 lines · Geert Uytterhoeven
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.
Quoted 6 lines
Quoted 4 lines
Where have these properties come from? They must be optional, because in the version of the binding documentation from Linux 5.0 they're not described:
https://elixir.bootlin.com/linux/v5.0/source/Documentation/devicetree/bindings/gpio/pl061-gpio.txtMany old DT bindings are incomplete.
Yeah, but production QEMU is out there in the world based on the old DT binding documentation. So you can't unilaterally make a part of the binding that wasn't documented and that QEMU didn't emit mandatory. It might be preferable for new QEMU to emit it, of course.
Quoted 11 lines
When running the validation on a device tree passed to the guest (extracted from
/sys/firmware/devicetree/base, converted to dts, and
manually fixed up the phandles), the following is reported about the
pl061 node:virt.dt.yaml: pl061@9030000: {'reg': [[0, 151191552, 0, 4096]],'gpio-controller': True, 'phandle': [[32771]], '#gpio-cells': [[2]], 'clocks': [[32768]], '#interrupt-cells': [[2]], 'compatible':
['arm,pl061', 'arm,primecell'], 'clock-names': ['apb_pclk'], '$nodename': ['pl061@9030000']} is not valid under any of the given schemas
Quoted 6 lines
[...] virt.dt.yaml: pl061@9030000: 'interrupts' is a required property virt.dt.yaml: pl061@9030000: $nodename:0: 'pl061@9030000' does notmatch '^gpio@[0-9a-f]+$'
virt.dt.yaml: pl061@9030000: 'interrupt-controller' is a required property
This is just saying "the yaml says these things are mandatory". You could equally get rid of them by marking them optional in the yaml, right?
Also, complaining about the nodename seems like a bug in the
validation: it is not a mandatory part of the spec, just a recommendation.
Quoted 8 lines
Quoted 5 lines
Since the devicetree spec says that the interrupt-controller property "defines a node as an interrupt controller node" and a GPIO chip isn't an interrupt controller, this seems like some kind of error in the dtb binding. Maybe I'm missing something...
PL061 is an interrupt controller, as it can assert its interrupt output based on activity on GPIO input lines.
By that logic the PL011 UART is an interrupt controller, because it can assert its interrupt output based on activity on the serial port input lines.
A GPIO controller is not an interrupt controller inherently. Maybe you can use it in a system design as an interrupt controller if you want to, and in that system's dtb perhaps it would make sense to label it as one, but the virt board's PL061 is in no way an interrupt controller -- it's just a GPIO controller.
Quoted 7 lines
Quoted 2 lines
What actually goes wrong if QEMU doesn't specify these properties?
It means that other devices that have their interrupt output connected to a PL061 GPIO input won't work, as their driver in the guest OS cannot find the interrupt. Note that
arm/virt.ccurrently doesn't instantiate such devices.
OK. But why would we want to run an interrupt line through the GPIO controller when we have a perfectly good interrupt controller in the system already?
It might be reasonable to add the properties now to avoid setting a bear trap for ourselves in future; on the other hand if running interrupt lines through the GPIO controller doesn't work then it acts as a nudge to stop people adding devices that are wired up in a weird way.
Quoted 16 lines
Quoted 7 lines
Quoted 7 lines
hw/arm/virt.c +0
-0@@ -818,13 +818,15 @@ static void create_gpio(const VirtMachineState *vms) qdev_get_gpio_in(vms->gic, irq));Quoted 8 lines
hw/arm/virt.c static void create_gpio(const VirtMachineState *vms) · line 821
Quoted 3 lines
hw/arm/virt.c +1
-1uint32_t phandle =qemu_fdt_alloc_phandle (vms->fdt); - nodename =g_strdup_printf ("/pl061@%" PRIx64, base); + nodename =g_strdup_printf ("/gpio@%" PRIx64, base);Does the devicetree binding really mandate what the node name is? I thought that finding the right device was doe via the 'compatible' string and the nodename could be whatever the device tree creator wanted.
Quoted 1 line
Matching is indeed done based on compatible value.
OK, then we don't need to change the node name here. Lots of the other devices on the virt board have node names that happen to use the device name rather than being more generic.
thanks
-- PMM
Quoted 9 lines · Peter Maydell
Quoted 8 lines · Geert Uytterhoeven
Quoted 7 lines · Peter Maydell
Quoted 4 lines · Geert Uytterhoeven
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.
[...]
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Since the devicetree spec says that the interrupt-controller property "defines a node as an interrupt controller node" and a GPIO chip isn't an interrupt controller, this seems like some kind of error in the dtb binding. Maybe I'm missing something...
Quoted 6 lines
Quoted 2 lines
PL061 is an interrupt controller, as it can assert its interrupt output based on activity on GPIO input lines.
By that logic the PL011 UART is an interrupt controller, because it can assert its interrupt output based on activity on the serial port input lines.
Yes :)
Quoted 10 lines
A GPIO controller is not an interrupt controller inherently. Maybe you can use it in a system design as an interrupt controller if you want to, and in that system's dtb perhaps it would make sense to label it as one, but the virt board's PL061 is in no way an interrupt controller -- it's just a GPIO controller.
Quoted 3 lines
Quoted 2 lines
What actually goes wrong if QEMU doesn't specify these properties?
Quoted 8 lines
Quoted 4 lines
It means that other devices that have their interrupt output connected to a PL061 GPIO input won't work, as their driver in the guest OS cannot find the interrupt. Note that
arm/virt.ccurrently doesn't instantiate such devices.OK. But why would we want to run an interrupt line through the GPIO controller when we have a perfectly good interrupt controller in the system already?
This is sometimes done on embedded devices when all the INTC lines are already wired. You'd use extra lines on free peripherals. Usually the peripheral offer a limited GPIO mode as passthru interrupt, else you use nasty hacks...
Quoted 6 lines
It might be reasonable to add the properties now to avoid setting a bear trap for ourselves in future; on the other hand if running interrupt lines through the GPIO controller doesn't work then it acts as a nudge to stop people adding devices that are wired up in a weird way.
[...]
Hi Peter,
CC Rob, who made these properties mandatory in the pl061 DT bindings.
Quoted 14 lines · Peter Maydell
Quoted 13 lines · Geert Uytterhoeven
Quoted 11 lines · Peter Maydell
Quoted 4 lines · Geert Uytterhoeven
Make the created node comply with the PL061 Device Tree bindings:
- Use generic node name "gpio" instead of "pl061",
- Add missing "#interrupt-cells" and "interrupt-controller" properties.Where have these properties come from? They must be optional, because in the version of the binding documentation from Linux 5.0 they're not described:
https://elixir.bootlin.com/linux/v5.0/source/Documentation/devicetree/bindings/gpio/pl061-gpio.txt
Quoted 7 lines
Quoted 1 line
Many old DT bindings are incomplete.
Yeah, but production QEMU is out there in the world based on the old DT binding documentation. So you can't unilaterally make a part of the binding that wasn't documented and that QEMU didn't emit mandatory. It might be preferable for new QEMU to emit it, of course.
Indeed, else QEMU will differ (more) from real hardware. Most PL061 GPIO controllers in DTS files in upstream Linux are marked as interrupt controllers (a few haven't been updated yet).
Quoted 18 lines
Quoted 11 lines
When running the validation on a device tree passed to the guest (extracted from
/sys/firmware/devicetree/base, converted to dts, and
manually fixed up the phandles), the following is reported about the
pl061 node:virt.dt.yaml: pl061@9030000: {'reg': [[0, 151191552, 0, 4096]],'gpio-controller': True, 'phandle': [[32771]], '#gpio-cells': [[2]], 'clocks': [[32768]], '#interrupt-cells': [[2]], 'compatible':
['arm,pl061', 'arm,primecell'], 'clock-names': ['apb_pclk'], '$nodename': ['pl061@9030000']} is not valid under any of the given schemasQuoted 6 lines
[...] virt.dt.yaml: pl061@9030000: 'interrupts' is a required property virt.dt.yaml: pl061@9030000: $nodename:0: 'pl061@9030000' does notmatch '^gpio@[0-9a-f]+$'
virt.dt.yaml: pl061@9030000: 'interrupt-controller' is a required property
Quoted 7 lines
This is just saying "the yaml says these things are mandatory". You could equally get rid of them by marking them optional in the yaml, right?
Also, complaining about the nodename seems like a bug in the
validation: it is not a mandatory part of the spec, just a recommendation.
I'll defer that one to Rob, too.
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Since the devicetree spec says that the interrupt-controller property "defines a node as an interrupt controller node" and a GPIO chip isn't an interrupt controller, this seems like some kind of error in the dtb binding. Maybe I'm missing something...
Quoted 6 lines
Quoted 2 lines
PL061 is an interrupt controller, as it can assert its interrupt output based on activity on GPIO input lines.
By that logic the PL011 UART is an interrupt controller, because it can assert its interrupt output based on activity on the serial port input lines.
Doh, bad wording on my side :-)
Quoted 6 lines
A GPIO controller is not an interrupt controller inherently. Maybe you can use it in a system design as an interrupt controller if you want to, and in that system's dtb perhaps it would make sense to label it as one, but the virt board's PL061 is in no way an interrupt controller -- it's just a GPIO controller.
Actually it is: the virt board has a gpio-keys node that ties the Poweroff key to a PL061 GPIO. The Linux gpio-keys driver relies on interrupts to detect key events (for IRQ incapable GPIOs you have to use "gpio-keys-polled" instead). This can easily be verified by looking at /proc/interrupts on the guest:
48: 0 9030000.gpio 3 Edge GPIO Key Poweroff
(Unfortunately I don't know how to trigger a virtual key press in the
guest)
Quoted 7 lines
Quoted 7 lines
Quoted 2 lines
What actually goes wrong if QEMU doesn't specify these properties?
It means that other devices that have their interrupt output connected to a PL061 GPIO input won't work, as their driver in the guest OS cannot find the interrupt. Note that
arm/virt.ccurrently doesn't instantiate such devices.
Seems I was wrong: arm/virt does have the Poweroff key, and Linux does find the interrupt.
Quoted 3 lines
OK. But why would we want to run an interrupt line through the GPIO controller when we have a perfectly good interrupt controller in the system already?
Because the GIC (usually?) does not have its interrupt lines brought outside the SoC, but defers external interrupt handling to a specialized external IRQ handling block, or to a GPIO controller with interrupt capabilities (or to PCI MSI).
If you have an external (non-PCI) device that needs to trigger an interrupt (e.g. an I2C sensor), its to be tied to the specialized block, or to a GPIO controller.
Quoted 5 lines
It might be reasonable to add the properties now to avoid setting a bear trap for ourselves in future; on the other hand if running interrupt lines through the GPIO controller doesn't work then it acts as a nudge to stop people adding devices that are wired up in a weird way.
There are plenty of working examples for this.
Quoted 16 lines
Quoted 16 lines
Quoted 15 lines
hw/arm/virt.c static void create_gpio(const VirtMachineState *vms) · line 821
Quoted 10 lines
hw/arm/virt.c +1
-1@@ -818,13 +818,15 @@ static void create_gpio(const VirtMachineState *vms) qdev_get_gpio_in(vms->gic, irq));uint32_t phandle =qemu_fdt_alloc_phandle (vms->fdt); - nodename =g_strdup_printf ("/pl061@%" PRIx64, base); + nodename =g_strdup_printf ("/gpio@%" PRIx64, base);Does the devicetree binding really mandate what the node name is? I thought that finding the right device was doe via the 'compatible' string and the nodename could be whatever the device tree creator wanted.
Quoted 5 lines
Quoted 1 line
Matching is indeed done based on compatible value.
OK, then we don't need to change the node name here. Lots of the other devices on the virt board have node names that happen to use the device name rather than being more generic.
Obviously these can (and IMHO should) be fixed, too, together with all other validation issues.
Thanks!
Rob: full thread at
https://lore.kernel.org/qemu-devel/[email protected]/
Gr{oetje,eeting}s,
Geert
Signature · 5 lines
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected] In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds