| Message ID | 20210805211200.491275-4-jeremy.linton@arm.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
| Series | CM4 ACPI PCIe quirk | expand |
In subject, this or similar would match history: PCI/ACPI: Add Broadcom bcm2711 MCFG quirk On Thu, Aug 05, 2021 at 04:12:00PM -0500, Jeremy Linton wrote: > Now that we have a bcm2711 quirk, we need to be able to > detect it when the MCFG is missing. Use a namespace > property as an alternative to the MCFG OEM. Rewrap to use ~75 columns. Mention the DT namespace property here. > Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> > --- > drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ > 1 file changed, 14 insertions(+) > > diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c > index 53cab975f612..7d77fc72c2a4 100644 > --- a/drivers/acpi/pci_mcfg.c > +++ b/drivers/acpi/pci_mcfg.c > @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { > ALTRA_ECAM_QUIRK(1, 13), > ALTRA_ECAM_QUIRK(1, 14), > ALTRA_ECAM_QUIRK(1, 15), > + > + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, > + DEFINE_RES_MEM(0xFD500000, 0xA000) }, > }; > > static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; > @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, > u16 segment = root->segment; > struct resource *bus_range = &root->secondary; > struct mcfg_fixup *f; > + const char *soc; > int i; > > + /* > + * This could be a machine with a PCI/SMC conduit, > + * which means it doens't have MCFG. Get the machineid from > + * the namespace definition instead. s/SMC/SMCCC/ ? Cover letter uses SMCCC (not sure it's relevant anyway) s/doens't/doesn't/ Rewrap comment to use ~80 columns. Seems pretty reasonable that a platform without standard ECAM might not have MCFG, since MCFG basically implies ECAM. Is "linux,pcie-quirk" the right property to look for? It doesn't sound very generic, and it doesn't sound like anything related to ECAM. Is it new? I don't see it in the tree yet. Should it be in Documentation/devicetree/bindings/pci/pci.txt so we don't get a different property name for every new platform? > + */ > + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), > + "linux,pcie-quirk", &soc)) { > + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); > + } > + > for (i = 0, f = mcfg_quirks; i < ARRAY_SIZE(mcfg_quirks); i++, f++) { > if (pci_mcfg_quirk_matches(f, segment, bus_range)) { > if (f->cfgres.start) > -- > 2.31.1 >
Hi, Thanks for looking at this. On 8/6/21 5:12 PM, Bjorn Helgaas wrote: > In subject, this or similar would match history: > > PCI/ACPI: Add Broadcom bcm2711 MCFG quirk > > On Thu, Aug 05, 2021 at 04:12:00PM -0500, Jeremy Linton wrote: >> Now that we have a bcm2711 quirk, we need to be able to >> detect it when the MCFG is missing. Use a namespace >> property as an alternative to the MCFG OEM. > > Rewrap to use ~75 columns. > > Mention the DT namespace property here. > >> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> >> --- >> drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ >> 1 file changed, 14 insertions(+) >> >> diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c >> index 53cab975f612..7d77fc72c2a4 100644 >> --- a/drivers/acpi/pci_mcfg.c >> +++ b/drivers/acpi/pci_mcfg.c >> @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { >> ALTRA_ECAM_QUIRK(1, 13), >> ALTRA_ECAM_QUIRK(1, 14), >> ALTRA_ECAM_QUIRK(1, 15), >> + >> + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, >> + DEFINE_RES_MEM(0xFD500000, 0xA000) }, >> }; >> >> static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; >> @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, >> u16 segment = root->segment; >> struct resource *bus_range = &root->secondary; >> struct mcfg_fixup *f; >> + const char *soc; >> int i; >> >> + /* >> + * This could be a machine with a PCI/SMC conduit, >> + * which means it doens't have MCFG. Get the machineid from >> + * the namespace definition instead. > > s/SMC/SMCCC/ ? Cover letter uses SMCCC (not sure it's relevant anyway) > s/doens't/doesn't/ > > Rewrap comment to use ~80 columns. > > Seems pretty reasonable that a platform without standard ECAM might > not have MCFG, since MCFG basically implies ECAM. Sure, on all the above comments. > > Is "linux,pcie-quirk" the right property to look for? It doesn't > sound very generic, and it doesn't sound like anything related to > ECAM. Is it new? I don't see it in the tree yet. Should it be in > Documentation/devicetree/bindings/pci/pci.txt so we don't get a > different property name for every new platform? Yes, I made it up. Someone else commented about the "linux," partially because it should be "linux-" to conform with https://github.com/UEFI/DSD-Guide. But also in the same context of it being linux specific. I think that guide is where it should end up, rather than the devicetree bindings. I guess we can request addition to the uefi- but that seems like a mistake this is really (hopefully?) a Linux specific properly as other OS's will simply use the SMC. I think we could request another prefix if we come up with a good one and think it belongs in that guide. > >> + */ >> + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), >> + "linux,pcie-quirk", &soc)) { >> + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); >> + } >> + >> for (i = 0, f = mcfg_quirks; i < ARRAY_SIZE(mcfg_quirks); i++, f++) { >> if (pci_mcfg_quirk_matches(f, segment, bus_range)) { >> if (f->cfgres.start) >> -- >> 2.31.1 >>
On Fri, Aug 6, 2021 at 6:35 PM Jeremy Linton <jeremy.linton@arm.com> wrote: > > Hi, > > Thanks for looking at this. > > On 8/6/21 5:12 PM, Bjorn Helgaas wrote: > > In subject, this or similar would match history: > > > > PCI/ACPI: Add Broadcom bcm2711 MCFG quirk > > > > On Thu, Aug 05, 2021 at 04:12:00PM -0500, Jeremy Linton wrote: > >> Now that we have a bcm2711 quirk, we need to be able to > >> detect it when the MCFG is missing. Use a namespace > >> property as an alternative to the MCFG OEM. > > > > Rewrap to use ~75 columns. > > > > Mention the DT namespace property here. > > > >> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> > >> --- > >> drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ > >> 1 file changed, 14 insertions(+) > >> > >> diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c > >> index 53cab975f612..7d77fc72c2a4 100644 > >> --- a/drivers/acpi/pci_mcfg.c > >> +++ b/drivers/acpi/pci_mcfg.c > >> @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { > >> ALTRA_ECAM_QUIRK(1, 13), > >> ALTRA_ECAM_QUIRK(1, 14), > >> ALTRA_ECAM_QUIRK(1, 15), > >> + > >> + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, > >> + DEFINE_RES_MEM(0xFD500000, 0xA000) }, > >> }; > >> > >> static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; > >> @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, > >> u16 segment = root->segment; > >> struct resource *bus_range = &root->secondary; > >> struct mcfg_fixup *f; > >> + const char *soc; > >> int i; > >> > >> + /* > >> + * This could be a machine with a PCI/SMC conduit, > >> + * which means it doens't have MCFG. Get the machineid from > >> + * the namespace definition instead. > > > > s/SMC/SMCCC/ ? Cover letter uses SMCCC (not sure it's relevant anyway) > > s/doens't/doesn't/ > > > > Rewrap comment to use ~80 columns. > > > > Seems pretty reasonable that a platform without standard ECAM might > > not have MCFG, since MCFG basically implies ECAM. > > > Sure, on all the above comments. > > > > > Is "linux,pcie-quirk" the right property to look for? It doesn't > > sound very generic, and it doesn't sound like anything related to > > ECAM. Is it new? I don't see it in the tree yet. Should it be in > > Documentation/devicetree/bindings/pci/pci.txt so we don't get a > > different property name for every new platform? No, it should not be in pci.txt. Nothing to do with DT and I don't review ACPI bindings (someone should) if I can help it. > Yes, I made it up. Someone else commented about the "linux," partially > because it should be "linux-" to conform with > https://github.com/UEFI/DSD-Guide. But also in the same context of it > being linux specific. I think that guide is where it should end up, > rather than the devicetree bindings. > > I guess we can request addition to the uefi- but that seems like a > mistake this is really (hopefully?) a Linux specific properly as other > OS's will simply use the SMC. I think we could request another prefix if > we come up with a good one and think it belongs in that guide. It's only Linux specific until it's not. What happens when there's a second PCIe quirk here? Say what the quirk is (and not in terms of Linux policy). Don't you know the device here and can imply all this (like implying off of 'compatible' in DT)? Adding properties to address issues creates compatibility issues. Maybe not an issue in this case if the firmware is not stable, but not a good practice to be in. Rob
Hi, Thanks for looking at this. On 8/9/21 10:27 AM, Rob Herring wrote: > On Fri, Aug 6, 2021 at 6:35 PM Jeremy Linton <jeremy.linton@arm.com> wrote: >> >> Hi, >> >> Thanks for looking at this. >> >> On 8/6/21 5:12 PM, Bjorn Helgaas wrote: >>> In subject, this or similar would match history: >>> >>> PCI/ACPI: Add Broadcom bcm2711 MCFG quirk >>> >>> On Thu, Aug 05, 2021 at 04:12:00PM -0500, Jeremy Linton wrote: >>>> Now that we have a bcm2711 quirk, we need to be able to >>>> detect it when the MCFG is missing. Use a namespace >>>> property as an alternative to the MCFG OEM. >>> >>> Rewrap to use ~75 columns. >>> >>> Mention the DT namespace property here. >>> >>>> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> >>>> --- >>>> drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ >>>> 1 file changed, 14 insertions(+) >>>> >>>> diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c >>>> index 53cab975f612..7d77fc72c2a4 100644 >>>> --- a/drivers/acpi/pci_mcfg.c >>>> +++ b/drivers/acpi/pci_mcfg.c >>>> @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { >>>> ALTRA_ECAM_QUIRK(1, 13), >>>> ALTRA_ECAM_QUIRK(1, 14), >>>> ALTRA_ECAM_QUIRK(1, 15), >>>> + >>>> + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, >>>> + DEFINE_RES_MEM(0xFD500000, 0xA000) }, >>>> }; >>>> >>>> static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; >>>> @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, >>>> u16 segment = root->segment; >>>> struct resource *bus_range = &root->secondary; >>>> struct mcfg_fixup *f; >>>> + const char *soc; >>>> int i; >>>> >>>> + /* >>>> + * This could be a machine with a PCI/SMC conduit, >>>> + * which means it doens't have MCFG. Get the machineid from >>>> + * the namespace definition instead. >>> >>> s/SMC/SMCCC/ ? Cover letter uses SMCCC (not sure it's relevant anyway) >>> s/doens't/doesn't/ >>> >>> Rewrap comment to use ~80 columns. >>> >>> Seems pretty reasonable that a platform without standard ECAM might >>> not have MCFG, since MCFG basically implies ECAM. >> >> >> Sure, on all the above comments. >> >>> >>> Is "linux,pcie-quirk" the right property to look for? It doesn't >>> sound very generic, and it doesn't sound like anything related to >>> ECAM. Is it new? I don't see it in the tree yet. Should it be in >>> Documentation/devicetree/bindings/pci/pci.txt so we don't get a >>> different property name for every new platform? > > No, it should not be in pci.txt. Nothing to do with DT and I don't > review ACPI bindings (someone should) if I can help it. That is the intention of the uefi properties registry I referred to earlier. It has a code first model too, which allows us to review it here and then update the document with the property and the intended behavior. > >> Yes, I made it up. Someone else commented about the "linux," partially >> because it should be "linux-" to conform with >> https://github.com/UEFI/DSD-Guide. But also in the same context of it >> being linux specific. I think that guide is where it should end up, >> rather than the devicetree bindings. >> >> I guess we can request addition to the uefi- but that seems like a >> mistake this is really (hopefully?) a Linux specific properly as other >> OS's will simply use the SMC. I think we could request another prefix if >> we come up with a good one and think it belongs in that guide. > > It's only Linux specific until it's not. > > What happens when there's a second PCIe quirk here? Say what the quirk > is (and not in terms of Linux policy). This is really just a stand in for the existing MCFG oem id, its an identifier to key off, nothing else. Maybe a better name is "linux-ecam-quirk-id" or something to that effect? > > Don't you know the device here and can imply all this (like implying > off of 'compatible' in DT)? Adding properties to address issues > creates compatibility issues. Maybe not an issue in this case if the > firmware is not stable, but not a good practice to be in. Yes, and no, I think there was some discussion a few years back about using non standard HID's for these nonstandard implementations, but that was discouraged at the time in favor of using the standard identifiers and and keying off a Soc/platform specific ID to enable a "quirk". Given we are a bit far down that path and the decision was made to continue down it (rather than solving much of it with a new firmware interface) this seems like the straightforward solution. The vast majority of these are SoC specific, and just minor tweaks for alignment/etc. Given its an ACPI/UEFI machine we are already hiding a lot of the platform specific behavior for configuring the bridge/etc that might require DT like properties in the firmware. Thanks again.
Hi Jeremy, On 8/5/21 4:12 PM, Jeremy Linton wrote: > Now that we have a bcm2711 quirk, we need to be able to > detect it when the MCFG is missing. Use a namespace > property as an alternative to the MCFG OEM. > > Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> > --- > drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ > 1 file changed, 14 insertions(+) > > diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c > index 53cab975f612..7d77fc72c2a4 100644 > --- a/drivers/acpi/pci_mcfg.c > +++ b/drivers/acpi/pci_mcfg.c > @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { > ALTRA_ECAM_QUIRK(1, 13), > ALTRA_ECAM_QUIRK(1, 14), > ALTRA_ECAM_QUIRK(1, 15), > + > + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, > + DEFINE_RES_MEM(0xFD500000, 0xA000) }, > }; > > static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; > @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, > u16 segment = root->segment; > struct resource *bus_range = &root->secondary; > struct mcfg_fixup *f; > + const char *soc; > int i; > > + /* > + * This could be a machine with a PCI/SMC conduit, > + * which means it doens't have MCFG. Get the machineid from > + * the namespace definition instead. > + */ > + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), > + "linux,pcie-quirk", &soc)) { > + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); > + } > + Is there any specific reason for not using the firmware agnostic API to get properties? if (!device_property_read_string(root->device, "linux,pcie-quirk", &soc)) { memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); }
Hi, Thanks for looking at this! On 8/10/21 9:31 AM, Shanker R Donthineni wrote: > Hi Jeremy, > > On 8/5/21 4:12 PM, Jeremy Linton wrote: >> Now that we have a bcm2711 quirk, we need to be able to >> detect it when the MCFG is missing. Use a namespace >> property as an alternative to the MCFG OEM. >> >> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> >> --- >> drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ >> 1 file changed, 14 insertions(+) >> >> diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c >> index 53cab975f612..7d77fc72c2a4 100644 >> --- a/drivers/acpi/pci_mcfg.c >> +++ b/drivers/acpi/pci_mcfg.c >> @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { >> ALTRA_ECAM_QUIRK(1, 13), >> ALTRA_ECAM_QUIRK(1, 14), >> ALTRA_ECAM_QUIRK(1, 15), >> + >> + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, >> + DEFINE_RES_MEM(0xFD500000, 0xA000) }, >> }; >> >> static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; >> @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, >> u16 segment = root->segment; >> struct resource *bus_range = &root->secondary; >> struct mcfg_fixup *f; >> + const char *soc; >> int i; >> >> + /* >> + * This could be a machine with a PCI/SMC conduit, >> + * which means it doens't have MCFG. Get the machineid from >> + * the namespace definition instead. >> + */ >> + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), >> + "linux,pcie-quirk", &soc)) { >> + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); >> + } >> + > > Is there any specific reason for not using the firmware agnostic API to get properties? > > > if (!device_property_read_string(root->device, "linux,pcie-quirk", &soc)) { > memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); > } > > IIRC it was because the "device" here isn't a struct device, rather a struct acpi_device. I think this is the normal way in this situation since we are directly picking up the fwnode rather than finding a generic node and then backtracking to get the fwnode.
On 8/10/21 9:47 AM, Jeremy Linton wrote: >>> diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c >>> index 53cab975f612..7d77fc72c2a4 100644 >>> --- a/drivers/acpi/pci_mcfg.c >>> +++ b/drivers/acpi/pci_mcfg.c >>> @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { >>> ALTRA_ECAM_QUIRK(1, 13), >>> ALTRA_ECAM_QUIRK(1, 14), >>> ALTRA_ECAM_QUIRK(1, 15), >>> + >>> + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, >>> + DEFINE_RES_MEM(0xFD500000, 0xA000) }, >>> }; >>> >>> static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; >>> @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, >>> u16 segment = root->segment; >>> struct resource *bus_range = &root->secondary; >>> struct mcfg_fixup *f; >>> + const char *soc; >>> int i; >>> >>> + /* >>> + * This could be a machine with a PCI/SMC conduit, >>> + * which means it doens't have MCFG. Get the machineid from >>> + * the namespace definition instead. >>> + */ >>> + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), >>> + "linux,pcie-quirk", &soc)) { >>> + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); >>> + } >>> + >> >> Is there any specific reason for not using the firmware agnostic API to get properties? >> >> >> if (!device_property_read_string(root->device, "linux,pcie-quirk", &soc)) { >> memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); >> } >> >> > > IIRC it was because the "device" here isn't a struct device, rather a > struct acpi_device. I think this is the normal way in this situation > since we are directly picking up the fwnode rather than finding a > generic node and then backtracking to get the fwnode. Yes, you are right.
diff --git a/drivers/acpi/pci_mcfg.c b/drivers/acpi/pci_mcfg.c index 53cab975f612..7d77fc72c2a4 100644 --- a/drivers/acpi/pci_mcfg.c +++ b/drivers/acpi/pci_mcfg.c @@ -169,6 +169,9 @@ static struct mcfg_fixup mcfg_quirks[] = { ALTRA_ECAM_QUIRK(1, 13), ALTRA_ECAM_QUIRK(1, 14), ALTRA_ECAM_QUIRK(1, 15), + + { "bcm2711", "", 0, 0, MCFG_BUS_ANY, &bcm2711_pcie_ops, + DEFINE_RES_MEM(0xFD500000, 0xA000) }, }; static char mcfg_oem_id[ACPI_OEM_ID_SIZE]; @@ -198,8 +201,19 @@ static void pci_mcfg_apply_quirks(struct acpi_pci_root *root, u16 segment = root->segment; struct resource *bus_range = &root->secondary; struct mcfg_fixup *f; + const char *soc; int i; + /* + * This could be a machine with a PCI/SMC conduit, + * which means it doens't have MCFG. Get the machineid from + * the namespace definition instead. + */ + if (!fwnode_property_read_string(acpi_fwnode_handle(root->device), + "linux,pcie-quirk", &soc)) { + memcpy(mcfg_oem_id, soc, ACPI_OEM_ID_SIZE); + } + for (i = 0, f = mcfg_quirks; i < ARRAY_SIZE(mcfg_quirks); i++, f++) { if (pci_mcfg_quirk_matches(f, segment, bus_range)) { if (f->cfgres.start)
Now that we have a bcm2711 quirk, we need to be able to detect it when the MCFG is missing. Use a namespace property as an alternative to the MCFG OEM. Signed-off-by: Jeremy Linton <jeremy.linton@arm.com> --- drivers/acpi/pci_mcfg.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)