| Message ID | 20220701161624.2844305-2-pierre.gondois@arm.com (mailing list archive) |
|---|---|
| State | Changes Requested, archived |
| Headers | show |
| Series | Make _PRS and _SRS methods optional | expand |
On Fri, Jul 01, 2022 at 06:16:23PM +0200, Pierre Gondois wrote: > From: Pierre Gondois <Pierre.Gondois@arm.com> > > In ACPI 6.4, s6.2.13 "_PRT (PCI Routing Table)", PCI legacy > interrupts can be described though a link device (first model). > From s6.2.16 "_SRS (Set Resource Settings)": > "This optional control method [...]" > > Make it optional to have a _SRS method for link devices. I think it would be helpful to outline the reason for wanting these changes in the commit log. Otherwise we don't know the benefit and it's harder to justify making the change since it's not an obvious cleanup. IIRC from [1] there *is* a good reason: you need to use Interrupt Link devices so you can specify "level triggered, active high". Without an Interrupt Link, you would get the default "level triggered, active low" setting, which apparently isn't compatible with GICv2. I assume this fixes a device that previously didn't work correctly, but I don't see the details of that in the bugzilla. I'm a little confused about this. Isn't GICv2 widely used already? How are things working now? Or are there just a lot of broken devices? [1] https://lore.kernel.org/r/e2ae06ba-de8f-2cae-60fa-fe9a215d779b@arm.com > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=215560 > Signed-off-by: Pierre Gondois <pierre.gondois@arm.com> > --- > drivers/acpi/pci_link.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/acpi/pci_link.c b/drivers/acpi/pci_link.c > index 58647051c948..129e3e7e80ee 100644 > --- a/drivers/acpi/pci_link.c > +++ b/drivers/acpi/pci_link.c > @@ -288,6 +288,13 @@ static int acpi_pci_link_set(struct acpi_pci_link *link, int irq) > if (!irq) > return -EINVAL; > > + if (!acpi_has_method(handle, METHOD_NAME__SRS)) { > + if (link->irq.active == irq) > + return 0; > + acpi_handle_err(handle, "Unable to set IRQ %d: No _SRS.\n", irq); > + return -ENODEV; > + } > + > resource = kzalloc(sizeof(*resource) + 1, irqs_disabled() ? GFP_ATOMIC: GFP_KERNEL); > if (!resource) > return -ENOMEM; > -- > 2.25.1 >
On Fri, Jul 1, 2022 at 6:17 PM Pierre Gondois <pierre.gondois@arm.com> wrote: > > From: Pierre Gondois <Pierre.Gondois@arm.com> > > In ACPI 6.4, s6.2.13 "_PRT (PCI Routing Table)", PCI legacy > interrupts can be described though a link device (first model). > From s6.2.16 "_SRS (Set Resource Settings)": > "This optional control method [...]" > > Make it optional to have a _SRS method for link devices. Note that if _DIS is present, _SRS is necessary to enable the link and acpi_pci_link_add() evaluates _DIS for all links. So you need to check both, not just one. Moreover, it doesn't make much sense to provide _PRS without _SRS and arguably _PRS is needed if _SRS is present, so this needs to be taken into account too. AFAICS, the only valid configuration in which _SRS and _PRS are not present is when _DIS is not present too, so only _CRS is present and the IRQ listed by it is actually in use. However, in that case it is hardly necessary to add a device object for the PCI link device at all. > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=215560 > Signed-off-by: Pierre Gondois <pierre.gondois@arm.com> > --- > drivers/acpi/pci_link.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/acpi/pci_link.c b/drivers/acpi/pci_link.c > index 58647051c948..129e3e7e80ee 100644 > --- a/drivers/acpi/pci_link.c > +++ b/drivers/acpi/pci_link.c > @@ -288,6 +288,13 @@ static int acpi_pci_link_set(struct acpi_pci_link *link, int irq) > if (!irq) > return -EINVAL; > > + if (!acpi_has_method(handle, METHOD_NAME__SRS)) { > + if (link->irq.active == irq) > + return 0; > + acpi_handle_err(handle, "Unable to set IRQ %d: No _SRS.\n", irq); > + return -ENODEV; > + } > + > resource = kzalloc(sizeof(*resource) + 1, irqs_disabled() ? GFP_ATOMIC: GFP_KERNEL); > if (!resource) > return -ENOMEM; > -- > 2.25.1 >
On 7/5/22 19:29, Bjorn Helgaas wrote: > On Fri, Jul 01, 2022 at 06:16:23PM +0200, Pierre Gondois wrote: >> From: Pierre Gondois <Pierre.Gondois@arm.com> >> >> In ACPI 6.4, s6.2.13 "_PRT (PCI Routing Table)", PCI legacy >> interrupts can be described though a link device (first model). >> From s6.2.16 "_SRS (Set Resource Settings)": >> "This optional control method [...]" >> >> Make it optional to have a _SRS method for link devices. > > I think it would be helpful to outline the reason for wanting these > changes in the commit log. Otherwise we don't know the benefit and > it's harder to justify making the change since it's not an obvious > cleanup. > > IIRC from [1] there *is* a good reason: you need to use Interrupt Link > devices so you can specify "level triggered, active high". > > Without an Interrupt Link, you would get the default "level triggered, > active low" setting, which apparently isn't compatible with GICv2. > > I assume this fixes a device that previously didn't work correctly, > but I don't see the details of that in the bugzilla. I'm a little > confused about this. Isn't GICv2 widely used already? How are things > working now? Or are there just a lot of broken devices? It was unsure which of the 2 models described in ACPI 6.4, s6.2.13 "_PRT (PCI Routing Table)" would be used for UEFI for kvmtool. Remainder: The first model allows to accurately describe interrupts: level/edge triggered and active high/low. Interrupts are also configurable with _CRS/_PRS/_SRS/_DIS methods The second model allows to describe hardwired interrupts, and are by default level triggered, active low. The kernel is aware that GivV2 interrupts are active high, so there was actually no need to accurately describe them. Thus the second model was used. While experimenting, we temporarily had a configuration using the first model, and only had a _CRS method (no _PRS/_SRS), which triggered some warnings. So these patches are not fixes for existing platforms, but merely to make _PRS/_SRS methods optional. In [1] I said I would submit patches to change that. If you think this is not necessary as the configuration is non-existing, I am perfectly fine to drop the patches. Also as Rafael noted, the _DIS method should also be taken into consideration if _PRS/_SRS are made optional. Regards, Pierre > > [1] https://lore.kernel.org/r/e2ae06ba-de8f-2cae-60fa-fe9a215d779b@arm.com > >> Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=215560 >> Signed-off-by: Pierre Gondois <pierre.gondois@arm.com> >> --- >> drivers/acpi/pci_link.c | 7 +++++++ >> 1 file changed, 7 insertions(+) >> >> diff --git a/drivers/acpi/pci_link.c b/drivers/acpi/pci_link.c >> index 58647051c948..129e3e7e80ee 100644 >> --- a/drivers/acpi/pci_link.c >> +++ b/drivers/acpi/pci_link.c >> @@ -288,6 +288,13 @@ static int acpi_pci_link_set(struct acpi_pci_link *link, int irq) >> if (!irq) >> return -EINVAL; >> >> + if (!acpi_has_method(handle, METHOD_NAME__SRS)) { >> + if (link->irq.active == irq) >> + return 0; >> + acpi_handle_err(handle, "Unable to set IRQ %d: No _SRS.\n", irq); >> + return -ENODEV; >> + } >> + >> resource = kzalloc(sizeof(*resource) + 1, irqs_disabled() ? GFP_ATOMIC: GFP_KERNEL); >> if (!resource) >> return -ENOMEM; >> -- >> 2.25.1 >>
On Wed, Jul 06, 2022 at 11:52:56AM +0200, Pierre Gondois wrote: > On 7/5/22 19:29, Bjorn Helgaas wrote: > > On Fri, Jul 01, 2022 at 06:16:23PM +0200, Pierre Gondois wrote: > > > From: Pierre Gondois <Pierre.Gondois@arm.com> > > > > > > In ACPI 6.4, s6.2.13 "_PRT (PCI Routing Table)", PCI legacy > > > interrupts can be described though a link device (first model). > > > From s6.2.16 "_SRS (Set Resource Settings)": > > > "This optional control method [...]" > > > > > > Make it optional to have a _SRS method for link devices. > > > > I think it would be helpful to outline the reason for wanting these > > changes in the commit log. Otherwise we don't know the benefit and > > it's harder to justify making the change since it's not an obvious > > cleanup. > > > > IIRC from [1] there *is* a good reason: you need to use Interrupt Link > > devices so you can specify "level triggered, active high". > > > > Without an Interrupt Link, you would get the default "level triggered, > > active low" setting, which apparently isn't compatible with GICv2. > > > > I assume this fixes a device that previously didn't work correctly, > > but I don't see the details of that in the bugzilla. I'm a little > > confused about this. Isn't GICv2 widely used already? How are things > > working now? Or are there just a lot of broken devices? > > It was unsure which of the 2 models described in ACPI 6.4, s6.2.13 > "_PRT (PCI Routing Table)" would be used for UEFI for kvmtool. > > Remainder: > The first model allows to accurately describe interrupts: level/edge > triggered and active high/low. Interrupts are also configurable with > _CRS/_PRS/_SRS/_DIS methods > The second model allows to describe hardwired interrupts, and are > by default level triggered, active low. > > The kernel is aware that GivV2 interrupts are active high, so there > was actually no need to accurately describe them. Thus the second > model was used. > While experimenting, we temporarily had a configuration using > the first model, and only had a _CRS method (no _PRS/_SRS), which > triggered some warnings. OK, thanks. So it sounds like there is some existing kernel code that special-cases GICv2 interrupts to make them level/high, and that code would not have been necessary if _PRS/_SRS had been optional from the beginning. I don't think we could ever *remove* that code because there's firmware in the field that relies on it, and that firmware will never be updated. > So these patches are not fixes for existing platforms, but merely > to make _PRS/_SRS methods optional. > > In [1] I said I would submit patches to change that. If you think > this is not necessary as the configuration is non-existing, I am > perfectly fine to drop the patches. > > Also as Rafael noted, the _DIS method should also be taken into > consideration if _PRS/_SRS are made optional. But that said, I'm not opposed to making _PRS/_SRS optional if that makes legal and reasonable _PRT descriptions work, and if all the considerations Rafael mentioned are taken care of. Bjorn
diff --git a/drivers/acpi/pci_link.c b/drivers/acpi/pci_link.c index 58647051c948..129e3e7e80ee 100644 --- a/drivers/acpi/pci_link.c +++ b/drivers/acpi/pci_link.c @@ -288,6 +288,13 @@ static int acpi_pci_link_set(struct acpi_pci_link *link, int irq) if (!irq) return -EINVAL; + if (!acpi_has_method(handle, METHOD_NAME__SRS)) { + if (link->irq.active == irq) + return 0; + acpi_handle_err(handle, "Unable to set IRQ %d: No _SRS.\n", irq); + return -ENODEV; + } + resource = kzalloc(sizeof(*resource) + 1, irqs_disabled() ? GFP_ATOMIC: GFP_KERNEL); if (!resource) return -ENOMEM;