Message ID | 12ce9caa682545cd43318c4679530202140117c0.1696542537.git.balaton@eik.bme.hu (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | Add emulation of AmigaOne XE board | expand |
On 05/10/2023 23:13, BALATON Zoltan wrote: > The initial value for BARs were set in reset method for emulating > legacy mode at start but this does not work because PCI code resets > BARs after calling device reset method. This is certainly something I've noticed when testing previous versions of the VIA patches. Perhaps it's worth a separate thread to the PCI devs? > Additionally the values > written to BARs were also wrong. I don't believe this is correct: according to the datasheet the values on reset are the ones given in the current reset code, so even if the reset function is overridden at a later data during PCI bus reset, I would leave these for now since it is a different issue. > Move setting the BARs to a callback on writing the PCI config regsiter > that sets the compatibility mode (which firmwares needing this mode > seem to do) and fix their values to program it to use legacy port > numbers. As noted in a comment, we only do this when the BARs were > unset before, because logs from real machine show this is how real > chip works, even if it contradicts the data sheet which is not very > clear about this. > > Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> > --- > hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- > 1 file changed, 30 insertions(+), 5 deletions(-) > > diff --git a/hw/ide/via.c b/hw/ide/via.c > index fff23803a6..8186190207 100644 > --- a/hw/ide/via.c > +++ b/hw/ide/via.c > @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) > pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | > PCI_STATUS_DEVSEL_MEDIUM); > > - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); > - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); > - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); > - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); > - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ > pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); > > /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ > @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) > pci_set_long(pci_conf + 0xc0, 0x00020001); > } > > +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, > + uint32_t val, int len) > +{ > + pci_default_write_config(pd, addr, val, len); > + /* > + * Only set BARs if they are unset. Logs from real hardware show that > + * writing class_prog to enable compatibility mode after BARs were set > + * (possibly by firmware) it will use ports set by BARs not ISA ports > + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). Can you remind me again where the references are to non-100% native mode? The only thing I can find in Linux is https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". > + * But if 0x8a is written after reset without setting BARs then we want > + * legacy ports (this is done by AmigaOne firmware for example). > + */ What should happen here is that writing 0x8a should *always* switch to legacy mode, so the BARs are unused... > + if (addr == PCI_CLASS_PROG && val == 0x8a && > + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == > + PCI_BASE_ADDRESS_SPACE_IO) { > + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 > + | PCI_BASE_ADDRESS_SPACE_IO); > + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 > + | PCI_BASE_ADDRESS_SPACE_IO); > + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 > + | PCI_BASE_ADDRESS_SPACE_IO); > + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 > + | PCI_BASE_ADDRESS_SPACE_IO); > + /* BMIBA: 20-23h */ > + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 > + | PCI_BASE_ADDRESS_SPACE_IO); > + } > +} ... but what you're doing here is effectively forcing the PCI BARs to the legacy addresses. The reason we know this is because that is why you have the off-by-2 error in BARs 1 and 3. I could live with this approach for now if you're willing to adjust the comments accordingly clarifying that forcing the PCI BARs to the legacy addresses is a hack to be replaced in future with proper legacy ioports. At some point I will revive my PoC series on top of Bernhard's last series that implements this properly. > static void via_ide_realize(PCIDevice *dev, Error **errp) > { > PCIIDEState *d = PCI_IDE(dev); > @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void *data) > /* Reason: only works as function of VIA southbridge */ > dc->user_creatable = false; > > + k->config_write = via_ide_cfg_write; > k->realize = via_ide_realize; > k->exit = via_ide_exitfn; > k->vendor_id = PCI_VENDOR_ID_VIA; ATB, Mark.
On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: > On 05/10/2023 23:13, BALATON Zoltan wrote: > >> The initial value for BARs were set in reset method for emulating >> legacy mode at start but this does not work because PCI code resets >> BARs after calling device reset method. > > This is certainly something I've noticed when testing previous versions of > the VIA patches. Perhaps it's worth a separate thread to the PCI devs? I think I brought up this back then but was told current PCI code won't change and since that could break everything else that makes sense so this is something that we should take as given and accomodate that. >> Additionally the values >> written to BARs were also wrong. > > I don't believe this is correct: according to the datasheet the values on > reset are the ones given in the current reset code, so even if the reset > function is overridden at a later data during PCI bus reset, I would leave > these for now since it is a different issue. Those values are missing the IO space bit for one so they can't be correct as a BAR value no matter what the datasheet says. And since they are ineffective now I think it's best to remove them to avoid confusion. >> Move setting the BARs to a callback on writing the PCI config regsiter >> that sets the compatibility mode (which firmwares needing this mode >> seem to do) and fix their values to program it to use legacy port >> numbers. As noted in a comment, we only do this when the BARs were >> unset before, because logs from real machine show this is how real >> chip works, even if it contradicts the data sheet which is not very >> clear about this. >> >> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >> --- >> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >> 1 file changed, 30 insertions(+), 5 deletions(-) >> >> diff --git a/hw/ide/via.c b/hw/ide/via.c >> index fff23803a6..8186190207 100644 >> --- a/hw/ide/via.c >> +++ b/hw/ide/via.c >> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >> PCI_STATUS_DEVSEL_MEDIUM); >> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: >> 20-23h */ >> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >> pci_set_long(pci_conf + 0xc0, 0x00020001); >> } >> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >> + uint32_t val, int len) >> +{ >> + pci_default_write_config(pd, addr, val, len); >> + /* >> + * Only set BARs if they are unset. Logs from real hardware show that >> + * writing class_prog to enable compatibility mode after BARs were set >> + * (possibly by firmware) it will use ports set by BARs not ISA ports >> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). > > Can you remind me again where the references are to non-100% native mode? The > only thing I can find in Linux is > https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 > but that simply forces a switch to legacy mode, with no mention of "non-100% > native mode". It was discussed somewhere in the via-ide thread we had when this was last touched for pegasos2 in March 2020. Basically the non-100% native mode is when ports are set by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 possible modes: legacy (both ports and IRQs are hard coded to ISA values), native (using BARs and PCI config 0x3c for a single interrupt for both channels, vt82c686 data sheet does not document this but vt8231 has a comment saying native mode only) and non-100% native mode where BARs are effective to set port addresses but IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines only work in non-100% native mode such as pegasos2 and Linux has some quirks for this. Other OSes running on this machine work with what the firmware has set up and can't work with anything else so we need to emulate what those OSes want (which matches real hardware) because Linux can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ 14-15 which is what the firmware also sets up by setting mode to native in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts by setting mode back to legacy but it still uses BARs and this is what it calls non-100% native mode. On amigaone firmware sets legacy mode and AmigaOS does not change this but uses it with legacy ports and IRQs. Linux finds the same and uses legacy mode on amigaone. >> + * But if 0x8a is written after reset without setting BARs then we >> want >> + * legacy ports (this is done by AmigaOne firmware for example). >> + */ > > What should happen here is that writing 0x8a should *always* switch to legacy > mode, so the BARs are unused... Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation being static and only allows adding ports but not removing them later so we can't switch between legacy ISA and PCI here so we use the BARs for legacy ports as well to emulate that. The reason we only do this if BARs are not yet set is because Linux changes this back to legacy mode on pegasos2 but still uses BARs as shown in boot logs from real hardware so it seems BARs take priority over legacy mode on real chip and Linux only uses the mode reg to decide what IRQs to use. On amigaone firmware writes 0x8a right after reset in which case we want legacy mode so this works for both machines. >> + if (addr == PCI_CLASS_PROG && val == 0x8a && >> + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == >> + PCI_BASE_ADDRESS_SPACE_IO) { >> + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 >> + | PCI_BASE_ADDRESS_SPACE_IO); >> + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 >> + | PCI_BASE_ADDRESS_SPACE_IO); >> + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 >> + | PCI_BASE_ADDRESS_SPACE_IO); >> + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 >> + | PCI_BASE_ADDRESS_SPACE_IO); >> + /* BMIBA: 20-23h */ >> + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 >> + | PCI_BASE_ADDRESS_SPACE_IO); >> + } >> +} > > ... but what you're doing here is effectively forcing the PCI BARs to the > legacy addresses. The reason we know this is because that is why you have the > off-by-2 error in BARs 1 and 3. > > I could live with this approach for now if you're willing to adjust the > comments accordingly clarifying that forcing the PCI BARs to the legacy > addresses is a hack to be replaced in future with proper legacy ioports. At > some point I will revive my PoC series on top of Bernhard's last series that > implements this properly. That's fair enoough, I can try to clarify thid more in the comments and commit message. I'll try to come up with something for v2. Regards, BALATON Zoltan >> static void via_ide_realize(PCIDevice *dev, Error **errp) >> { >> PCIIDEState *d = PCI_IDE(dev); >> @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void >> *data) >> /* Reason: only works as function of VIA southbridge */ >> dc->user_creatable = false; >> + k->config_write = via_ide_cfg_write; >> k->realize = via_ide_realize; >> k->exit = via_ide_exitfn; >> k->vendor_id = PCI_VENDOR_ID_VIA; > > > ATB, > > Mark. > > >
Am 8. Oktober 2023 11:08:58 UTC schrieb BALATON Zoltan <balaton@eik.bme.hu>: >On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >> On 05/10/2023 23:13, BALATON Zoltan wrote: >> >>> The initial value for BARs were set in reset method for emulating >>> legacy mode at start but this does not work because PCI code resets >>> BARs after calling device reset method. >> >> This is certainly something I've noticed when testing previous versions of the VIA patches. Perhaps it's worth a separate thread to the PCI devs? > >I think I brought up this back then but was told current PCI code won't change and since that could break everything else that makes sense so this is something that we should take as given and accomodate that. Why not play safe like: 1. add a class property such as `reset_bar_addrs[PCI_NUM_REGIONS]` 2. set all elements to zero in `pci_device_class_init()` 3. respect `reset_bar_addrs` in `pci_reset_regions()` 4. assign the proper reset addresses of TYPE_VIA_IDE in `via_ide_class_init()` That would pretty obviously preserve the behavior of existing device models while allowing TYPE_VIA_IDE to be reset properly. It would also perform the main part of the workaround in the code that exhibits the limitation, so the code could potentially be simplified at some point without impacting all PCI device models. Best regards, Bernhard > >>> Additionally the values >>> written to BARs were also wrong. >> >> I don't believe this is correct: according to the datasheet the values on reset are the ones given in the current reset code, so even if the reset function is overridden at a later data during PCI bus reset, I would leave these for now since it is a different issue. > >Those values are missing the IO space bit for one so they can't be correct as a BAR value no matter what the datasheet says. And since they are ineffective now I think it's best to remove them to avoid confusion. > >>> Move setting the BARs to a callback on writing the PCI config regsiter >>> that sets the compatibility mode (which firmwares needing this mode >>> seem to do) and fix their values to program it to use legacy port >>> numbers. As noted in a comment, we only do this when the BARs were >>> unset before, because logs from real machine show this is how real >>> chip works, even if it contradicts the data sheet which is not very >>> clear about this. >>> >>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>> --- >>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>> 1 file changed, 30 insertions(+), 5 deletions(-) >>> >>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>> index fff23803a6..8186190207 100644 >>> --- a/hw/ide/via.c >>> +++ b/hw/ide/via.c >>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>> PCI_STATUS_DEVSEL_MEDIUM); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ >>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>> } >>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>> + uint32_t val, int len) >>> +{ >>> + pci_default_write_config(pd, addr, val, len); >>> + /* >>> + * Only set BARs if they are unset. Logs from real hardware show that >>> + * writing class_prog to enable compatibility mode after BARs were set >>> + * (possibly by firmware) it will use ports set by BARs not ISA ports >>> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). >> >> Can you remind me again where the references are to non-100% native mode? The only thing I can find in Linux is https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". > >It was discussed somewhere in the via-ide thread we had when this was last touched for pegasos2 in March 2020. Basically the non-100% native mode is when ports are set by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 possible modes: legacy (both ports and IRQs are hard coded to ISA values), native (using BARs and PCI config 0x3c for a single interrupt for both channels, vt82c686 data sheet does not document this but vt8231 has a comment saying native mode only) and non-100% native mode where BARs are effective to set port addresses but IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines only work in non-100% native mode such as pegasos2 and Linux has some quirks for this. Other OSes running on this machine work with what the firmware has set up and can't work with anything else so we need to emulate what those OSes want (which matches real hardware) because Linux can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ 14-15 which is what the firmware also sets up by setting mode to native in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts by setting mode back to legacy but it still uses BARs and this is what it calls non-100% native mode. On amigaone firmware sets legacy mode and AmigaOS does not change this but uses it with legacy ports and IRQs. Linux finds the same and uses legacy mode on amigaone. > >>> + * But if 0x8a is written after reset without setting BARs then we want >>> + * legacy ports (this is done by AmigaOne firmware for example). >>> + */ >> >> What should happen here is that writing 0x8a should *always* switch to legacy mode, so the BARs are unused... > >Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation being static and only allows adding ports but not removing them later so we can't switch between legacy ISA and PCI here so we use the BARs for legacy ports as well to emulate that. The reason we only do this if BARs are not yet set is because Linux changes this back to legacy mode on pegasos2 but still uses BARs as shown in boot logs from real hardware so it seems BARs take priority over legacy mode on real chip and Linux only uses the mode reg to decide what IRQs to use. On amigaone firmware writes 0x8a right after reset in which case we want legacy mode so this works for both machines. > >>> + if (addr == PCI_CLASS_PROG && val == 0x8a && >>> + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == >>> + PCI_BASE_ADDRESS_SPACE_IO) { >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + /* BMIBA: 20-23h */ >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + } >>> +} >> >> ... but what you're doing here is effectively forcing the PCI BARs to the legacy addresses. The reason we know this is because that is why you have the off-by-2 error in BARs 1 and 3. >> >> I could live with this approach for now if you're willing to adjust the comments accordingly clarifying that forcing the PCI BARs to the legacy addresses is a hack to be replaced in future with proper legacy ioports. At some point I will revive my PoC series on top of Bernhard's last series that implements this properly. > >That's fair enoough, I can try to clarify thid more in the comments and commit message. I'll try to come up with something for v2. > >Regards, >BALATON Zoltan > >>> static void via_ide_realize(PCIDevice *dev, Error **errp) >>> { >>> PCIIDEState *d = PCI_IDE(dev); >>> @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void *data) >>> /* Reason: only works as function of VIA southbridge */ >>> dc->user_creatable = false; >>> + k->config_write = via_ide_cfg_write; >>> k->realize = via_ide_realize; >>> k->exit = via_ide_exitfn; >>> k->vendor_id = PCI_VENDOR_ID_VIA; >> >> >> ATB, >> >> Mark. >> >> >>
Am 8. Oktober 2023 11:08:58 UTC schrieb BALATON Zoltan <balaton@eik.bme.hu>: >On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >> On 05/10/2023 23:13, BALATON Zoltan wrote: >> >>> The initial value for BARs were set in reset method for emulating >>> legacy mode at start but this does not work because PCI code resets >>> BARs after calling device reset method. >> >> This is certainly something I've noticed when testing previous versions of the VIA patches. Perhaps it's worth a separate thread to the PCI devs? > >I think I brought up this back then but was told current PCI code won't change and since that could break everything else that makes sense so this is something that we should take as given and accomodate that. > >>> Additionally the values >>> written to BARs were also wrong. >> >> I don't believe this is correct: according to the datasheet the values on reset are the ones given in the current reset code, so even if the reset function is overridden at a later data during PCI bus reset, I would leave these for now since it is a different issue. > >Those values are missing the IO space bit for one so they can't be correct as a BAR value no matter what the datasheet says. And since they are ineffective now I think it's best to remove them to avoid confusion. > >>> Move setting the BARs to a callback on writing the PCI config regsiter >>> that sets the compatibility mode (which firmwares needing this mode >>> seem to do) and fix their values to program it to use legacy port >>> numbers. As noted in a comment, we only do this when the BARs were >>> unset before, because logs from real machine show this is how real >>> chip works, even if it contradicts the data sheet which is not very >>> clear about this. >>> >>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>> --- >>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>> 1 file changed, 30 insertions(+), 5 deletions(-) >>> >>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>> index fff23803a6..8186190207 100644 >>> --- a/hw/ide/via.c >>> +++ b/hw/ide/via.c >>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>> PCI_STATUS_DEVSEL_MEDIUM); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ >>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>> } >>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>> + uint32_t val, int len) >>> +{ >>> + pci_default_write_config(pd, addr, val, len); >>> + /* >>> + * Only set BARs if they are unset. Logs from real hardware show that >>> + * writing class_prog to enable compatibility mode after BARs were set >>> + * (possibly by firmware) it will use ports set by BARs not ISA ports >>> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). >> >> Can you remind me again where the references are to non-100% native mode? The only thing I can find in Linux is https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". > >It was discussed somewhere in the via-ide thread we had when this was last touched for pegasos2 in March 2020. Basically the non-100% native mode is when ports are set by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 possible modes: legacy (both ports and IRQs are hard coded to ISA values), native (using BARs and PCI config 0x3c for a single interrupt for both channels, vt82c686 data sheet does not document this but vt8231 has a comment saying native mode only) and non-100% native mode where BARs are effective to set port addresses but IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines only work in non-100% native mode such as pegasos2 and Linux has some quirks for this. Other OSes running on this machine work with what the firmware has set up and can't work with anything else so we need to emulate what those OSes want (which matches real hardware) because Linux can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ 14-15 which is what the firmware also sets up by setting mode to native in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts by setting mode back to legacy but it still uses BARs and this is what it calls non-100% native mode. On amigaone firmware sets legacy mode and AmigaOS does not change this but uses it with legacy ports and IRQs. Linux finds the same and uses legacy mode on amigaone. > >>> + * But if 0x8a is written after reset without setting BARs then we want >>> + * legacy ports (this is done by AmigaOne firmware for example). >>> + */ >> >> What should happen here is that writing 0x8a should *always* switch to legacy mode, so the BARs are unused... > >Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation being static and only allows adding ports but not removing them later so we can't switch between legacy ISA and PCI here so we use the BARs for legacy ports as well to emulate that. The reason we only do this if BARs are not yet set is because Linux changes this back to legacy mode on pegasos2 but still uses BARs Just curious: How can you tell the difference in real hardware whether "raw" IO ports vs. BARs mapped to IO space are used? > as shown in boot logs from real hardware Could you provide links to such logs? That would be very helpful to have -- even in the code -- for documentation. Best regards, Bernhard > so it seems BARs take priority over legacy mode on real chip and Linux only uses the mode reg to decide what IRQs to use. On amigaone firmware writes 0x8a right after reset in which case we want legacy mode so this works for both machines. > >>> + if (addr == PCI_CLASS_PROG && val == 0x8a && >>> + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == >>> + PCI_BASE_ADDRESS_SPACE_IO) { >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + /* BMIBA: 20-23h */ >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + } >>> +} >> >> ... but what you're doing here is effectively forcing the PCI BARs to the legacy addresses. The reason we know this is because that is why you have the off-by-2 error in BARs 1 and 3. >> >> I could live with this approach for now if you're willing to adjust the comments accordingly clarifying that forcing the PCI BARs to the legacy addresses is a hack to be replaced in future with proper legacy ioports. At some point I will revive my PoC series on top of Bernhard's last series that implements this properly. > >That's fair enoough, I can try to clarify thid more in the comments and commit message. I'll try to come up with something for v2. > >Regards, >BALATON Zoltan > >>> static void via_ide_realize(PCIDevice *dev, Error **errp) >>> { >>> PCIIDEState *d = PCI_IDE(dev); >>> @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void *data) >>> /* Reason: only works as function of VIA southbridge */ >>> dc->user_creatable = false; >>> + k->config_write = via_ide_cfg_write; >>> k->realize = via_ide_realize; >>> k->exit = via_ide_exitfn; >>> k->vendor_id = PCI_VENDOR_ID_VIA; >> >> >> ATB, >> >> Mark. >> >> >>
On Mon, 9 Oct 2023, Bernhard Beschow wrote: > Am 8. Oktober 2023 11:08:58 UTC schrieb BALATON Zoltan <balaton@eik.bme.hu>: >> On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >>> On 05/10/2023 23:13, BALATON Zoltan wrote: >>> >>>> The initial value for BARs were set in reset method for emulating >>>> legacy mode at start but this does not work because PCI code resets >>>> BARs after calling device reset method. >>> >>> This is certainly something I've noticed when testing previous versions of the VIA patches. Perhaps it's worth a separate thread to the PCI devs? >> >> I think I brought up this back then but was told current PCI code won't change and since that could break everything else that makes sense so this is something that we should take as given and accomodate that. > > Why not play safe like: > 1. add a class property such as `reset_bar_addrs[PCI_NUM_REGIONS]` > 2. set all elements to zero in `pci_device_class_init()` > 3. respect `reset_bar_addrs` in `pci_reset_regions()` > 4. assign the proper reset addresses of TYPE_VIA_IDE in `via_ide_class_init()` > > That would pretty obviously preserve the behavior of existing device > models while allowing TYPE_VIA_IDE to be reset properly. It would also > perform the main part of the workaround in the code that exhibits the > limitation, so the code could potentially be simplified at some point > without impacting all PCI device models. That's a lot of complication for setting some values that will be overwritten first thing after reset. Either the guest sets native mode and writes the BARs itself so the reset values don't matter or it sets legacy mode when the workaround has to set legacy ports (that strangely don't match the reset values of the BARs) so sticking to those reset values makes no sense to me as it does not help to run any guest just makes the code more complex and harder to understand for no reason. So the patch with clarified comment as Mark asked should do for now. I'll send that in v2 later. Regards, BALATON Zoltan
On Mon, 9 Oct 2023, Bernhard Beschow wrote: > Am 8. Oktober 2023 11:08:58 UTC schrieb BALATON Zoltan <balaton@eik.bme.hu>: >> On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >>> On 05/10/2023 23:13, BALATON Zoltan wrote: >>> >>>> The initial value for BARs were set in reset method for emulating >>>> legacy mode at start but this does not work because PCI code resets >>>> BARs after calling device reset method. >>> >>> This is certainly something I've noticed when testing previous versions of the VIA patches. Perhaps it's worth a separate thread to the PCI devs? >> >> I think I brought up this back then but was told current PCI code won't change and since that could break everything else that makes sense so this is something that we should take as given and accomodate that. >> >>>> Additionally the values >>>> written to BARs were also wrong. >>> >>> I don't believe this is correct: according to the datasheet the values on reset are the ones given in the current reset code, so even if the reset function is overridden at a later data during PCI bus reset, I would leave these for now since it is a different issue. >> >> Those values are missing the IO space bit for one so they can't be correct as a BAR value no matter what the datasheet says. And since they are ineffective now I think it's best to remove them to avoid confusion. >> >>>> Move setting the BARs to a callback on writing the PCI config regsiter >>>> that sets the compatibility mode (which firmwares needing this mode >>>> seem to do) and fix their values to program it to use legacy port >>>> numbers. As noted in a comment, we only do this when the BARs were >>>> unset before, because logs from real machine show this is how real >>>> chip works, even if it contradicts the data sheet which is not very >>>> clear about this. >>>> >>>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>>> --- >>>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>>> 1 file changed, 30 insertions(+), 5 deletions(-) >>>> >>>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>>> index fff23803a6..8186190207 100644 >>>> --- a/hw/ide/via.c >>>> +++ b/hw/ide/via.c >>>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>>> PCI_STATUS_DEVSEL_MEDIUM); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ >>>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >>>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>>> } >>>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>>> + uint32_t val, int len) >>>> +{ >>>> + pci_default_write_config(pd, addr, val, len); >>>> + /* >>>> + * Only set BARs if they are unset. Logs from real hardware show that >>>> + * writing class_prog to enable compatibility mode after BARs were set >>>> + * (possibly by firmware) it will use ports set by BARs not ISA ports >>>> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). >>> >>> Can you remind me again where the references are to non-100% native mode? The only thing I can find in Linux is https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". >> >> It was discussed somewhere in the via-ide thread we had when this was last touched for pegasos2 in March 2020. Basically the non-100% native mode is when ports are set by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 possible modes: legacy (both ports and IRQs are hard coded to ISA values), native (using BARs and PCI config 0x3c for a single interrupt for both channels, vt82c686 data sheet does not document this but vt8231 has a comment saying native mode only) and non-100% native mode where BARs are effective to set port addresses but IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines only work in non-100% native mode such as pegasos2 and Linux has some quirks for this. Other OSes running on this machine work with what the firmware has set up and can't work with anything else so we need to emulate what those OSes want (which matches real hardware) because Linux can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ 14-1 5 which is what the firmware also sets up by setting mode to native in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts by setting mode back to legacy but it still uses BARs and this is what it calls non-100% native mode. On amigaone firmware sets legacy mode and AmigaOS does not change this but uses it with legacy ports and IRQs. Linux finds the same and uses legacy mode on amigaone. >> >>>> + * But if 0x8a is written after reset without setting BARs then we want >>>> + * legacy ports (this is done by AmigaOne firmware for example). >>>> + */ >>> >>> What should happen here is that writing 0x8a should *always* switch to legacy mode, so the BARs are unused... >> >> Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation being static and only allows adding ports but not removing them later so we can't switch between legacy ISA and PCI here so we use the BARs for legacy ports as well to emulate that. The reason we only do this if BARs are not yet set is because Linux changes this back to legacy mode on pegasos2 but still uses BARs > > Just curious: How can you tell the difference in real hardware whether > "raw" IO ports vs. BARs mapped to IO space are used? Not sure what you mean but if you see ports not matching legacy ports then you can be fairly sure they are BARs. The chip may internally implement legacy mode by just fixing BARs at legacy values (the same as this patch does) but the data sheet does not tell how it works so we don't know that. >> as shown in boot logs from real hardware > > Could you provide links to such logs? That would be very helpful to have > -- even in the code -- for documentation. I think if you don't believe me you can just try the emulated pegasos2 with its original firmware and see how it configures the IDE device then boot Linux with that and see how it detects that. The relevant lines in the Linux dmesg are: [ 0.066337] pci 0000:00:0c.1: Fixing VIA IDE, force legacy mode on [ 0.514351] pata_via 0000:00:0c.1: version 0.3.4 [ 0.536431] scsi0 : pata_via [ 0.557917] scsi1 : pata_via [ 0.558249] ata1: PATA max UDMA/100 cmd 0x1000 ctl 0x100c bmdma 0x1020 irq 14 [ 0.558272] ata2: PATA max UDMA/100 cmd 0x1010 ctl 0x101c bmdma 0x1028 irq 15 These are from real machine but emulated one shold do the same with original firmware. The comments near the IDE fixup in Linux source is misleading as it says the controller only works in legacy mode but what it should say is that it uses legacy interrupts even in native mode as seen from the above logs. The ports are still as mapped by BARs but IRQs aren't the 9 that the firmware writes to 0x3c PCI config reg for all devices but 14-15 as in legacy mode. This is the non-100% native mode that I think older Linux versions printed in this case and what's MorphOS and AmigaOS expect so happen on pegasos2 irrespective of what's in the device tree. Also the forcing legacy mode quirk in Linux may only edit the device tree so the driver uses legacy interrupts, I'm not sure, it's been a while I've looked at this and it was already discussed to death back when I've added the pegasos2 machine so I don't want to get back to that again. This patch is tested to allow amigaone to work while not breaking pegasos2 so I'd go with this for now. Regards, BALATON Zoltan
On 08/10/2023 12:08, BALATON Zoltan wrote: > On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >> On 05/10/2023 23:13, BALATON Zoltan wrote: >> >>> The initial value for BARs were set in reset method for emulating >>> legacy mode at start but this does not work because PCI code resets >>> BARs after calling device reset method. >> >> This is certainly something I've noticed when testing previous versions of the VIA >> patches. Perhaps it's worth a separate thread to the PCI devs? > > I think I brought up this back then but was told current PCI code won't change and > since that could break everything else that makes sense so this is something that we > should take as given and accomodate that. I don't remember the details of that thread, but that's not too much of an issue here as the values won't be used. >>> Additionally the values >>> written to BARs were also wrong. >> >> I don't believe this is correct: according to the datasheet the values on reset are >> the ones given in the current reset code, so even if the reset function is >> overridden at a later data during PCI bus reset, I would leave these for now since >> it is a different issue. > > Those values are missing the IO space bit for one so they can't be correct as a BAR > value no matter what the datasheet says. And since they are ineffective now I think > it's best to remove them to avoid confusion. Maybe, or perhaps just fix up the missing IO space bit and add a comment pointing out these are the defaults, but currently they are erased on PCI bus reset? I have found it useful to have the values around to save having to reference the datasheet. >>> Move setting the BARs to a callback on writing the PCI config regsiter >>> that sets the compatibility mode (which firmwares needing this mode >>> seem to do) and fix their values to program it to use legacy port >>> numbers. As noted in a comment, we only do this when the BARs were >>> unset before, because logs from real machine show this is how real >>> chip works, even if it contradicts the data sheet which is not very >>> clear about this. >>> >>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>> --- >>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>> 1 file changed, 30 insertions(+), 5 deletions(-) >>> >>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>> index fff23803a6..8186190207 100644 >>> --- a/hw/ide/via.c >>> +++ b/hw/ide/via.c >>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>> PCI_STATUS_DEVSEL_MEDIUM); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ >>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>> } >>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>> + uint32_t val, int len) >>> +{ >>> + pci_default_write_config(pd, addr, val, len); >>> + /* >>> + * Only set BARs if they are unset. Logs from real hardware show that >>> + * writing class_prog to enable compatibility mode after BARs were set >>> + * (possibly by firmware) it will use ports set by BARs not ISA ports >>> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). >> >> Can you remind me again where the references are to non-100% native mode? The only >> thing I can find in Linux is >> https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". > > It was discussed somewhere in the via-ide thread we had when this was last touched > for pegasos2 in March 2020. Basically the non-100% native mode is when ports are set > by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 possible > modes: legacy (both ports and IRQs are hard coded to ISA values), native (using BARs > and PCI config 0x3c for a single interrupt for both channels, vt82c686 data sheet > does not document this but vt8231 has a comment saying native mode only) and non-100% > native mode where BARs are effective to set port addresses but IRQs don't respect > 0x3c but use 14-15 as in legacy mode. Some machines only work in non-100% native mode > such as pegasos2 and Linux has some quirks for this. Other OSes running on this > machine work with what the firmware has set up and can't work with anything else so > we need to emulate what those OSes want (which matches real hardware) because Linux > can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ 14-15 which is > what the firmware also sets up by setting mode to native in the PCI config of the IDE > func yet IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts by > setting mode back to legacy but it still uses BARs and this is what it calls non-100% > native mode. On amigaone firmware sets legacy mode and AmigaOS does not change this > but uses it with legacy ports and IRQs. Linux finds the same and uses legacy mode on > amigaone. Thanks for summarising: there have been a number of threads and changes over the years, so it is easy to lose track of where things are. From the above then everything except MorphOS that explicitly sets legacy/native mode should just work? Just to double check I don't think we ever managed to get the PCI configuration information from real hardware to confirm how the hardware thinks it is set? Is it possible to dump the PCI config space of the PCI-ISA bridge and the VIA IDE from a real Pegasos2 booted into Smart Firmware? You can get the information using the Forth below: : dump-single-config ( addr ) dup 100 + swap do i u. " : " type i " config-b@" $call-parent u. cr loop ; " /pci@80000000/isa@C" open-dev to my-self 800 c * 100 0 * + dump-single-config cr cr 800 c * 100 1 * + dump-single-config cr cr With the corresponding MorphOS debug log that will help to confirm the routing being used. Also it would be useful to get the output of any MorphOS program that can dump the current kernel IRQ routing configuration. The reason for wanting to double-check this now is because my work with the Mac q800 machine showed that both legacy and modern IRQ routing worked, so it would be good to confirm exactly what IRQ is currently active on real hardware. >>> + * But if 0x8a is written after reset without setting BARs then we want >>> + * legacy ports (this is done by AmigaOne firmware for example). >>> + */ >> >> What should happen here is that writing 0x8a should *always* switch to legacy mode, >> so the BARs are unused... > > Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation > being static and only allows adding ports but not removing them later Fortunately this bug has been fixed, so it should now be possible using portio_list_*() functions: see https://gitlab.com/qemu-project/qemu/-/commit/690705ca0b0f1ed24a34ccd14c9866fbe47c69a6. With that (and a bit of refactoring to allow the sharing of the IDE ioport list) I was able to switch between compatibility and native modes at will in my PoC. Once that is working then it doesn't matter if the default BARs aren't set correctly. > so we can't > switch between legacy ISA and PCI here so we use the BARs for legacy ports as well to > emulate that. The reason we only do this if BARs are not yet set is because Linux > changes this back to legacy mode on pegasos2 but still uses BARs as shown in boot > logs from real hardware so it seems BARs take priority over legacy mode on real chip > and Linux only uses the mode reg to decide what IRQs to use. On amigaone firmware > writes 0x8a right after reset in which case we want legacy mode so this works for > both machines. > >>> + if (addr == PCI_CLASS_PROG && val == 0x8a && >>> + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == >>> + PCI_BASE_ADDRESS_SPACE_IO) { >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + /* BMIBA: 20-23h */ >>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 >>> + | PCI_BASE_ADDRESS_SPACE_IO); >>> + } >>> +} >> >> ... but what you're doing here is effectively forcing the PCI BARs to the legacy >> addresses. The reason we know this is because that is why you have the off-by-2 >> error in BARs 1 and 3. >> >> I could live with this approach for now if you're willing to adjust the comments >> accordingly clarifying that forcing the PCI BARs to the legacy addresses is a hack >> to be replaced in future with proper legacy ioports. At some point I will revive my >> PoC series on top of Bernhard's last series that implements this properly. > > That's fair enoough, I can try to clarify thid more in the comments and commit > message. I'll try to come up with something for v2. > > Regards, > BALATON Zoltan > >>> static void via_ide_realize(PCIDevice *dev, Error **errp) >>> { >>> PCIIDEState *d = PCI_IDE(dev); >>> @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void *data) >>> /* Reason: only works as function of VIA southbridge */ >>> dc->user_creatable = false; >>> + k->config_write = via_ide_cfg_write; >>> k->realize = via_ide_realize; >>> k->exit = via_ide_exitfn; >>> k->vendor_id = PCI_VENDOR_ID_VIA; ATB, Mark.
On Mon, 9 Oct 2023, Mark Cave-Ayland wrote: > On 08/10/2023 12:08, BALATON Zoltan wrote: > >> On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >>> On 05/10/2023 23:13, BALATON Zoltan wrote: >>> >>>> The initial value for BARs were set in reset method for emulating >>>> legacy mode at start but this does not work because PCI code resets >>>> BARs after calling device reset method. >>> >>> This is certainly something I've noticed when testing previous versions of >>> the VIA patches. Perhaps it's worth a separate thread to the PCI devs? >> >> I think I brought up this back then but was told current PCI code won't >> change and since that could break everything else that makes sense so this >> is something that we should take as given and accomodate that. > > I don't remember the details of that thread, but that's not too much of an > issue here as the values won't be used. > >>>> Additionally the values >>>> written to BARs were also wrong. >>> >>> I don't believe this is correct: according to the datasheet the values on >>> reset are the ones given in the current reset code, so even if the reset >>> function is overridden at a later data during PCI bus reset, I would leave >>> these for now since it is a different issue. >> >> Those values are missing the IO space bit for one so they can't be correct >> as a BAR value no matter what the datasheet says. And since they are >> ineffective now I think it's best to remove them to avoid confusion. > > Maybe, or perhaps just fix up the missing IO space bit and add a comment > pointing out these are the defaults, but currently they are erased on PCI bus > reset? I have found it useful to have the values around to save having to > reference the datasheet. The data sheet does not list the io space bits so fixing that would lead to values not matching data sheet any more. Also the defaults in the data sheet don't make much sense even with io space but as some of them match legacy ports while others don't. I can either drop this hunk leaving the current values there or add a FIXME comment saying they are ineffective but because they are overwritten (either by PCI code now or firmware/guest later) I think it's best to remove them any maybe only bring them back if we find they would be needed for any guest and what would be the correct default valuss here. I don't trust the data sheet on that and getting it from real hardware is also not really possible because the firmware could have overwritten them by the time you can get them. So I don't think keeping these here would help anybody, just cause confusion. >>>> Move setting the BARs to a callback on writing the PCI config regsiter >>>> that sets the compatibility mode (which firmwares needing this mode >>>> seem to do) and fix their values to program it to use legacy port >>>> numbers. As noted in a comment, we only do this when the BARs were >>>> unset before, because logs from real machine show this is how real >>>> chip works, even if it contradicts the data sheet which is not very >>>> clear about this. >>>> >>>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>>> --- >>>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>>> 1 file changed, 30 insertions(+), 5 deletions(-) >>>> >>>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>>> index fff23803a6..8186190207 100644 >>>> --- a/hw/ide/via.c >>>> +++ b/hw/ide/via.c >>>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>>> PCI_STATUS_DEVSEL_MEDIUM); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: >>>> 20-23h */ >>>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO >>>> Configuration*/ >>>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>>> } >>>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>>> + uint32_t val, int len) >>>> +{ >>>> + pci_default_write_config(pd, addr, val, len); >>>> + /* >>>> + * Only set BARs if they are unset. Logs from real hardware show >>>> that >>>> + * writing class_prog to enable compatibility mode after BARs were >>>> set >>>> + * (possibly by firmware) it will use ports set by BARs not ISA >>>> ports >>>> + * (e.g. pegasos2 Linux does this and calls it non-100% native >>>> mode). >>> >>> Can you remind me again where the references are to non-100% native mode? >>> The only thing I can find in Linux is >>> https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 >>> but that simply forces a switch to legacy mode, with no mention of >>> "non-100% native mode". >> >> It was discussed somewhere in the via-ide thread we had when this was last >> touched for pegasos2 in March 2020. Basically the non-100% native mode is >> when ports are set by BARs but IRQs are still hard coded to 14-15. Linux >> can work with all 3 possible modes: legacy (both ports and IRQs are hard >> coded to ISA values), native (using BARs and PCI config 0x3c for a single >> interrupt for both channels, vt82c686 data sheet does not document this but >> vt8231 has a comment saying native mode only) and non-100% native mode >> where BARs are effective to set port addresses but IRQs don't respect 0x3c >> but use 14-15 as in legacy mode. Some machines only work in non-100% native >> mode such as pegasos2 and Linux has some quirks for this. Other OSes >> running on this machine work with what the firmware has set up and can't >> work with anything else so we need to emulate what those OSes want (which >> matches real hardware) because Linux can usually cope anyway. On pegasso2 >> MorphOS uses BARs but expects IRQ 14-15 which is what the firmware also >> sets up by setting mode to native in the PCI config of the IDE func yet >> IRQs are fixed at 14-15. Linux forces its driver to use legacy interrupts >> by setting mode back to legacy but it still uses BARs and this is what it >> calls non-100% native mode. On amigaone firmware sets legacy mode and >> AmigaOS does not change this but uses it with legacy ports and IRQs. Linux >> finds the same and uses legacy mode on amigaone. > > Thanks for summarising: there have been a number of threads and changes over > the years, so it is easy to lose track of where things are. From the above > then everything except MorphOS that explicitly sets legacy/native mode should > just work? No, everything but Linux (i.e. MorphOS and AmigaOS) only work with behaviour matching real hardware which is BARs + legacy interrupts. Linux may work with other settings but it also has fix ups to detect and work with the non-100% native mode as on real hardware. (I think this non-100% native mode was used hy the older Linus via ata driver before it was ported to libata and became pata_via as I remember seeing these in older logs but not in newer ones any more which just list IRQs and port numbers from which you can tell it's neither the legacy nor the native mode described in the data sheet but a mix between the two.) > Just to double check I don't think we ever managed to get the PCI > configuration information from real hardware to confirm how the hardware > thinks it is set? Is it possible to dump the PCI config space of the PCI-ISA > bridge and the VIA IDE from a real Pegasos2 booted into Smart Firmware? You > can get the information using the Forth below: > > > : dump-single-config ( addr ) > dup 100 + swap do i u. " : " type i " config-b@" $call-parent u. cr loop > ; > > " /pci@80000000/isa@C" open-dev to my-self > 800 c * 100 0 * + dump-single-config cr cr > 800 c * 100 1 * + dump-single-config cr cr I don't have dumps from firmware but got an lspci output from Linux from real pegasos2 which says: 0000:00:0c.1 IDE interface: VIA Technologies, Inc. VT82C586A/B/VT82C686/A/B/VT823x/A/C PIPC Bus Master IDE (rev 06) (prog-if 8a [Master SecP PriP]) Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- DisINTx- Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx- Latency: 0 Interrupt: pin ? routed to IRQ 14 Region 0: [virtual] I/O ports at 1000 [size=8] Region 1: [virtual] I/O ports at 100c [size=4] Region 2: [virtual] I/O ports at 1010 [size=8] Region 3: [virtual] I/O ports at 101c [size=4] Region 4: I/O ports at 1020 [size=16] Capabilities: [c0] Power Management version 2 Flags: PMEClk- DSI- D1- D2- AuxCurrent=0mA PME(D0-,D1-,D2-,D3hot-,D3cold-) Status: D0 NoSoftRst- PME-Enable- DSel=0 DScale=0 PME- Kernel driver in use: pata_via This says prog_if is 0x8a (probably set by Linux fixup as firmware sets it to native) but still uses BARs for ports and legacy interrupts although only says 14 above but in dmesg it listed interruts for the channels as quoted in reply to Bernhard. Also got /proc infos from same machine that says: /proc/ioports: 00001000-00001007 : 0000:00:0c.1 /proc/ioports: 00001000-00001007 : pata_via /proc/ioports: 0000100c-0000100f : 0000:00:0c.1 /proc/ioports: 0000100c-0000100f : pata_via /proc/ioports: 00001010-00001017 : 0000:00:0c.1 /proc/ioports: 00001010-00001017 : pata_via /proc/ioports: 0000101c-0000101f : 0000:00:0c.1 /proc/ioports: 0000101c-0000101f : pata_via /proc/ioports: 00001020-0000102f : 0000:00:0c.1 /proc/ioports: 00001020-0000102f : pata_via /proc/interrupts: 14: 1847 i8259 14 Level pata_via /proc/interrupts: 15: 1210 i8259 15 Level pata_via So what other proof you still need to believe this is how it works on real machine? > With the corresponding MorphOS debug log that will help to confirm the > routing being used. Also it would be useful to get the output of any MorphOS > program that can dump the current kernel IRQ routing configuration. MorphOS does not give much debug logs but it works currenntly as well as AmigaOS and Linux so what we have now is good enough and this patch does not break that. > The reason for wanting to double-check this now is because my work with the > Mac q800 machine showed that both legacy and modern IRQ routing worked, so it > would be good to confirm exactly what IRQ is currently active on real > hardware. What does q800 has to do with VT8231 and pegaos2? These are totally different hardware and software designed by separate people so other than you were looking at both there's no connection to consider between these. >>>> + * But if 0x8a is written after reset without setting BARs then we >>>> want >>>> + * legacy ports (this is done by AmigaOne firmware for example). >>>> + */ >>> >>> What should happen here is that writing 0x8a should *always* switch to >>> legacy mode, so the BARs are unused... >> >> Yes, but as we've found before that can't be emulated in QEMU due to ISA >> emulation being static and only allows adding ports but not removing them >> later > > Fortunately this bug has been fixed, so it should now be possible using > portio_list_*() functions: see > https://gitlab.com/qemu-project/qemu/-/commit/690705ca0b0f1ed24a34ccd14c9866fbe47c69a6. > With that (and a bit of refactoring to allow the sharing of the IDE ioport > list) I was able to switch between compatibility and native modes at will in > my PoC. Once that is working then it doesn't matter if the default BARs > aren't set correctly. Great but maybe that's not needed because firmware and guests usually configure this once at boot then use it with that setting without ever switching it again so this simple patch allows that to work without breaking what's there and tested already so I'd like this to be merged first with the amigaone machine then if you want to rewrite it later to emulate the chip more closely then we have at least two test cases with pegasos2 and amigaone to verify evertyhing still works correctly. But since this patch allows all guests to boot it may be a waste of time to try to patch this emulation further to add functionality that won't be used by anything. But if it keeps working like this patch does and still boots the guests this one allows I don't care that much about that. Regards, BALATON Zoltan >> so we can't switch between legacy ISA and PCI here so we use the BARs for >> legacy ports as well to emulate that. The reason we only do this if BARs >> are not yet set is because Linux changes this back to legacy mode on >> pegasos2 but still uses BARs as shown in boot logs from real hardware so it >> seems BARs take priority over legacy mode on real chip and Linux only uses >> the mode reg to decide what IRQs to use. On amigaone firmware writes 0x8a >> right after reset in which case we want legacy mode so this works for both >> machines. >> >>>> + if (addr == PCI_CLASS_PROG && val == 0x8a && >>>> + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == >>>> + PCI_BASE_ADDRESS_SPACE_IO) { >>>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 >>>> + | PCI_BASE_ADDRESS_SPACE_IO); >>>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 >>>> + | PCI_BASE_ADDRESS_SPACE_IO); >>>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 >>>> + | PCI_BASE_ADDRESS_SPACE_IO); >>>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 >>>> + | PCI_BASE_ADDRESS_SPACE_IO); >>>> + /* BMIBA: 20-23h */ >>>> + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 >>>> + | PCI_BASE_ADDRESS_SPACE_IO); >>>> + } >>>> +} >>> >>> ... but what you're doing here is effectively forcing the PCI BARs to the >>> legacy addresses. The reason we know this is because that is why you have >>> the off-by-2 error in BARs 1 and 3. >>> >>> I could live with this approach for now if you're willing to adjust the >>> comments accordingly clarifying that forcing the PCI BARs to the legacy >>> addresses is a hack to be replaced in future with proper legacy ioports. >>> At some point I will revive my PoC series on top of Bernhard's last series >>> that implements this properly. >> >> That's fair enoough, I can try to clarify thid more in the comments and >> commit message. I'll try to come up with something for v2. >> >> Regards, >> BALATON Zoltan >> >>>> static void via_ide_realize(PCIDevice *dev, Error **errp) >>>> { >>>> PCIIDEState *d = PCI_IDE(dev); >>>> @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, >>>> void *data) >>>> /* Reason: only works as function of VIA southbridge */ >>>> dc->user_creatable = false; >>>> + k->config_write = via_ide_cfg_write; >>>> k->realize = via_ide_realize; >>>> k->exit = via_ide_exitfn; >>>> k->vendor_id = PCI_VENDOR_ID_VIA; > > > ATB, > > Mark. > > >
On 09/10/2023 23:23, BALATON Zoltan wrote: > On Mon, 9 Oct 2023, Mark Cave-Ayland wrote: >> On 08/10/2023 12:08, BALATON Zoltan wrote: >> >>> On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >>>> On 05/10/2023 23:13, BALATON Zoltan wrote: >>>> >>>>> The initial value for BARs were set in reset method for emulating >>>>> legacy mode at start but this does not work because PCI code resets >>>>> BARs after calling device reset method. >>>> >>>> This is certainly something I've noticed when testing previous versions of the >>>> VIA patches. Perhaps it's worth a separate thread to the PCI devs? >>> >>> I think I brought up this back then but was told current PCI code won't change and >>> since that could break everything else that makes sense so this is something that >>> we should take as given and accomodate that. >> >> I don't remember the details of that thread, but that's not too much of an issue >> here as the values won't be used. >> >>>>> Additionally the values >>>>> written to BARs were also wrong. >>>> >>>> I don't believe this is correct: according to the datasheet the values on reset >>>> are the ones given in the current reset code, so even if the reset function is >>>> overridden at a later data during PCI bus reset, I would leave these for now >>>> since it is a different issue. >>> >>> Those values are missing the IO space bit for one so they can't be correct as a >>> BAR value no matter what the datasheet says. And since they are ineffective now I >>> think it's best to remove them to avoid confusion. >> >> Maybe, or perhaps just fix up the missing IO space bit and add a comment pointing >> out these are the defaults, but currently they are erased on PCI bus reset? I have >> found it useful to have the values around to save having to reference the datasheet. > > The data sheet does not list the io space bits so fixing that would lead to values > not matching data sheet any more. Also the defaults in the data sheet don't make much > sense even with io space but as some of them match legacy ports while others don't. I > can either drop this hunk leaving the current values there or add a FIXME comment > saying they are ineffective but because they are overwritten (either by PCI code now > or firmware/guest later) I think it's best to remove them any maybe only bring them > back if we find they would be needed for any guest and what would be the correct > default valuss here. I don't trust the data sheet on that and getting it from real > hardware is also not really possible because the firmware could have overwritten them > by the time you can get them. So I don't think keeping these here would help anybody, > just cause confusion. We can check the values on real hardware given the Forth in my previous reply which will then tell us the correct values for once and for all. My guess is that since the address is a separate field to the BAR type in the datasheet, the IO bit was simply missed. >>>>> Move setting the BARs to a callback on writing the PCI config regsiter >>>>> that sets the compatibility mode (which firmwares needing this mode >>>>> seem to do) and fix their values to program it to use legacy port >>>>> numbers. As noted in a comment, we only do this when the BARs were >>>>> unset before, because logs from real machine show this is how real >>>>> chip works, even if it contradicts the data sheet which is not very >>>>> clear about this. >>>>> >>>>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>>>> --- >>>>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>>>> 1 file changed, 30 insertions(+), 5 deletions(-) >>>>> >>>>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>>>> index fff23803a6..8186190207 100644 >>>>> --- a/hw/ide/via.c >>>>> +++ b/hw/ide/via.c >>>>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>>>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>>>> PCI_STATUS_DEVSEL_MEDIUM); >>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ >>>>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>>>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ >>>>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>>>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>>>> } >>>>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>>>> + uint32_t val, int len) >>>>> +{ >>>>> + pci_default_write_config(pd, addr, val, len); >>>>> + /* >>>>> + * Only set BARs if they are unset. Logs from real hardware show that >>>>> + * writing class_prog to enable compatibility mode after BARs were set >>>>> + * (possibly by firmware) it will use ports set by BARs not ISA ports >>>>> + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). >>>> >>>> Can you remind me again where the references are to non-100% native mode? The >>>> only thing I can find in Linux is >>>> https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 but that simply forces a switch to legacy mode, with no mention of "non-100% native mode". >>> >>> It was discussed somewhere in the via-ide thread we had when this was last touched >>> for pegasos2 in March 2020. Basically the non-100% native mode is when ports are >>> set by BARs but IRQs are still hard coded to 14-15. Linux can work with all 3 >>> possible modes: legacy (both ports and IRQs are hard coded to ISA values), native >>> (using BARs and PCI config 0x3c for a single interrupt for both channels, vt82c686 >>> data sheet does not document this but vt8231 has a comment saying native mode >>> only) and non-100% native mode where BARs are effective to set port addresses but >>> IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines only work >>> in non-100% native mode such as pegasos2 and Linux has some quirks for this. Other >>> OSes running on this machine work with what the firmware has set up and can't work >>> with anything else so we need to emulate what those OSes want (which matches real >>> hardware) because Linux can usually cope anyway. On pegasso2 MorphOS uses BARs but >>> expects IRQ 14-15 which is what the firmware also sets up by setting mode to >>> native in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux forces >>> its driver to use legacy interrupts by setting mode back to legacy but it still >>> uses BARs and this is what it calls non-100% native mode. On amigaone firmware >>> sets legacy mode and AmigaOS does not change this but uses it with legacy ports >>> and IRQs. Linux finds the same and uses legacy mode on amigaone. >> >> Thanks for summarising: there have been a number of threads and changes over the >> years, so it is easy to lose track of where things are. From the above then >> everything except MorphOS that explicitly sets legacy/native mode should just work? > > No, everything but Linux (i.e. MorphOS and AmigaOS) only work with behaviour matching > real hardware which is BARs + legacy interrupts. Linux may work with other settings > but it also has fix ups to detect and work with the non-100% native mode as on real > hardware. (I think this non-100% native mode was used hy the older Linus via ata > driver before it was ported to libata and became pata_via as I remember seeing these > in older logs but not in newer ones any more which just list IRQs and port numbers > from which you can tell it's neither the legacy nor the native mode described in the > data sheet but a mix between the two.) > >> Just to double check I don't think we ever managed to get the PCI configuration >> information from real hardware to confirm how the hardware thinks it is set? Is it >> possible to dump the PCI config space of the PCI-ISA bridge and the VIA IDE from a >> real Pegasos2 booted into Smart Firmware? You can get the information using the >> Forth below: >> >> >> : dump-single-config ( addr ) >> dup 100 + swap do i u. " : " type i " config-b@" $call-parent u. cr loop >> ; >> >> " /pci@80000000/isa@C" open-dev to my-self >> 800 c * 100 0 * + dump-single-config cr cr >> 800 c * 100 1 * + dump-single-config cr cr > > I don't have dumps from firmware but got an lspci output from Linux from real > pegasos2 which says: Ah no that won't work because Linux has already made changes to PCI configuration space. What we need is the value from the firmware to confirm exactly how the hardware has been configured before the kernel has loaded (as you mentioned above). Who can we ask to get this information? > 0000:00:0c.1 IDE interface: VIA Technologies, Inc. > VT82C586A/B/VT82C686/A/B/VT823x/A/C PIPC Bus Master IDE (rev 06) (prog-if 8a [Master > SecP PriP]) > Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- > Stepping- SERR- FastB2B- DisINTx- > Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort- <TAbort- > <MAbort- >SERR- <PERR- INTx- > Latency: 0 > Interrupt: pin ? routed to IRQ 14 Now this is interesting - I wonder what "?" means here? Perhaps the switch to compatibility mode has set the PCI interrupt pin to 0? Oh and what kernel version is this? > Region 0: [virtual] I/O ports at 1000 [size=8] > Region 1: [virtual] I/O ports at 100c [size=4] > Region 2: [virtual] I/O ports at 1010 [size=8] > Region 3: [virtual] I/O ports at 101c [size=4] > Region 4: I/O ports at 1020 [size=16] > Capabilities: [c0] Power Management version 2 > Flags: PMEClk- DSI- D1- D2- AuxCurrent=0mA > PME(D0-,D1-,D2-,D3hot-,D3cold-) > Status: D0 NoSoftRst- PME-Enable- DSel=0 DScale=0 PME- > Kernel driver in use: pata_via > > This says prog_if is 0x8a (probably set by Linux fixup as firmware sets it to native) > but still uses BARs for ports and legacy interrupts although only says 14 above but > in dmesg it listed interruts for the channels as quoted in reply to Bernhard. Also > got /proc infos from same machine that says: > > /proc/ioports: 00001000-00001007 : 0000:00:0c.1 > /proc/ioports: 00001000-00001007 : pata_via > /proc/ioports: 0000100c-0000100f : 0000:00:0c.1 > /proc/ioports: 0000100c-0000100f : pata_via > /proc/ioports: 00001010-00001017 : 0000:00:0c.1 > /proc/ioports: 00001010-00001017 : pata_via > /proc/ioports: 0000101c-0000101f : 0000:00:0c.1 > /proc/ioports: 0000101c-0000101f : pata_via > /proc/ioports: 00001020-0000102f : 0000:00:0c.1 > /proc/ioports: 00001020-0000102f : pata_via Interesting. The PCI-IDE specification states that the BAR values are ignored in compatibility mode, so it could be that they contain values left over from the firmware they are not being actively decoded on the bus. They certainly look similar to the values set by the firmware. In fact it would be possible to write some Forth to switch to compatibility mode in the firmware and then try and access the BARs to confirm whether or not the BARs are really being decoded. > /proc/interrupts: 14: 1847 i8259 14 Level pata_via > /proc/interrupts: 15: 1210 i8259 15 Level pata_via > > So what other proof you still need to believe this is how it works on real machine? Again this would have been overwritten by the switch to compatibility mode. >> With the corresponding MorphOS debug log that will help to confirm the routing >> being used. Also it would be useful to get the output of any MorphOS program that >> can dump the current kernel IRQ routing configuration. > > MorphOS does not give much debug logs but it works currenntly as well as AmigaOS and > Linux so what we have now is good enough and this patch does not break that. It looks like you can get PCI debug logs, at least that seems work here on QEMU. They would be very useful to understand the exact behaviour. >> The reason for wanting to double-check this now is because my work with the Mac >> q800 machine showed that both legacy and modern IRQ routing worked, so it would be >> good to confirm exactly what IRQ is currently active on real hardware. > > What does q800 has to do with VT8231 and pegaos2? These are totally different > hardware and software designed by separate people so other than you were looking at > both there's no connection to consider between these. The point I was making was that there may be multiple IRQ routing possibilities for the IDE controller, in which case it makes sense to be absolutely sure which one is being used on real hardware. >>>>> + * But if 0x8a is written after reset without setting BARs then we want >>>>> + * legacy ports (this is done by AmigaOne firmware for example). >>>>> + */ >>>> >>>> What should happen here is that writing 0x8a should *always* switch to legacy >>>> mode, so the BARs are unused... >>> >>> Yes, but as we've found before that can't be emulated in QEMU due to ISA emulation >>> being static and only allows adding ports but not removing them later >> >> Fortunately this bug has been fixed, so it should now be possible using >> portio_list_*() functions: see >> https://gitlab.com/qemu-project/qemu/-/commit/690705ca0b0f1ed24a34ccd14c9866fbe47c69a6. With that (and a bit of refactoring to allow the sharing of the IDE ioport list) I was able to switch between compatibility and native modes at will in my PoC. Once that is working then it doesn't matter if the default BARs aren't set correctly. > > Great but maybe that's not needed because firmware and guests usually configure this > once at boot then use it with that setting without ever switching it again so this > simple patch allows that to work without breaking what's there and tested already so > I'd like this to be merged first with the amigaone machine then if you want to > rewrite it later to emulate the chip more closely then we have at least two test > cases with pegasos2 and amigaone to verify evertyhing still works correctly. But > since this patch allows all guests to boot it may be a waste of time to try to patch > this emulation further to add functionality that won't be used by anything. But if it > keeps working like this patch does and still boots the guests this one allows I don't > care that much about that. The ultimate aim is to combine mine and Bernhard's work so that the compatibility mode/PCI native mode switch is handle automatically by the PCI IDE core, so it's something that would not just be restricted to the VIA. ATB, Mark.
On Mon, 9 Oct 2023, Mark Cave-Ayland wrote: > On 09/10/2023 23:23, BALATON Zoltan wrote: >> On Mon, 9 Oct 2023, Mark Cave-Ayland wrote: >>> On 08/10/2023 12:08, BALATON Zoltan wrote: >>> >>>> On Sun, 8 Oct 2023, Mark Cave-Ayland wrote: >>>>> On 05/10/2023 23:13, BALATON Zoltan wrote: >>>>> >>>>>> The initial value for BARs were set in reset method for emulating >>>>>> legacy mode at start but this does not work because PCI code resets >>>>>> BARs after calling device reset method. >>>>> >>>>> This is certainly something I've noticed when testing previous versions >>>>> of the VIA patches. Perhaps it's worth a separate thread to the PCI >>>>> devs? >>>> >>>> I think I brought up this back then but was told current PCI code won't >>>> change and since that could break everything else that makes sense so >>>> this is something that we should take as given and accomodate that. >>> >>> I don't remember the details of that thread, but that's not too much of an >>> issue here as the values won't be used. >>> >>>>>> Additionally the values >>>>>> written to BARs were also wrong. >>>>> >>>>> I don't believe this is correct: according to the datasheet the values >>>>> on reset are the ones given in the current reset code, so even if the >>>>> reset function is overridden at a later data during PCI bus reset, I >>>>> would leave these for now since it is a different issue. >>>> >>>> Those values are missing the IO space bit for one so they can't be >>>> correct as a BAR value no matter what the datasheet says. And since they >>>> are ineffective now I think it's best to remove them to avoid confusion. >>> >>> Maybe, or perhaps just fix up the missing IO space bit and add a comment >>> pointing out these are the defaults, but currently they are erased on PCI >>> bus reset? I have found it useful to have the values around to save having >>> to reference the datasheet. >> >> The data sheet does not list the io space bits so fixing that would lead to >> values not matching data sheet any more. Also the defaults in the data >> sheet don't make much sense even with io space but as some of them match >> legacy ports while others don't. I can either drop this hunk leaving the >> current values there or add a FIXME comment saying they are ineffective but >> because they are overwritten (either by PCI code now or firmware/guest >> later) I think it's best to remove them any maybe only bring them back if >> we find they would be needed for any guest and what would be the correct >> default valuss here. I don't trust the data sheet on that and getting it >> from real hardware is also not really possible because the firmware could >> have overwritten them by the time you can get them. So I don't think >> keeping these here would help anybody, just cause confusion. > > We can check the values on real hardware given the Forth in my previous reply > which will then tell us the correct values for once and for all. My guess is > that since the address is a separate field to the BAR type in the datasheet, > the IO bit was simply missed. I only have device tree dump from real pegasos2 but it has BARs so you can see here: package 0xFC5CE70 /pci/ide: parent=0xFC58E08 children=0xFC5D418 link=0xFC5D948 dict=0xFC5CEA0: open func[0x11F4] close func[0x11F5] dma-alloc func[0x11F6] dma-free func[0x11F7] decode-unit func[0x11F8] encode-unit func[0x11F9] selftest func[0x11FA] props=0xFC5CEB8: vendor-id 0x1106 (4358) device-id 0x571 (1393) revision-id 0x6 (6) class-code 0x1018F (65935) subsystem-id 0x0 (0) subsystem-vendor-id 0x0 (0) .vendor-name "VIA" .part-number "VT82C586/596/686" .description "PCI IDE Controller" interrupts 0x1 (1) devsel-speed 0x1 (1) fast-back-to-back min-grant 0x0 (0) max-latency 0x0 (0) name "ide" reg C,1:0 iC,1,10,0:8 iC,1,14,0:4 iC,1,18,0:8 iC,1,1C,0:4 iC,1,20,0:10 device_type "spi" assigned-addresses iC,1,10,FE001000:8 iC,1,14,FE00100C:4 iC,1,18,FE001010:8 iC,1,1C,FE00101C:4 iC,1,20,FE001020:10 That means by the time you could run your Forth the firmware already set the IDE device up se we get the same values as in Linux. If you want to see how the firmware progrems it you can find out with the emulated pegasos2 using -bios pegasos2.rom and enabling pci traces, but we can't find out what are the default values on real machine before this without replacing firmware that likely nobody dares to try. So just let it go and accept that whatever defaults are there, they don't matter because the firmware on pegasos2 overwrites them and the firmware on amigaone sets legacy mode so it should use legacy ports there and if anything switches to native mode later then it will also program BARs so it really does not matter what the data sheet says, what real hardware does and weather we put any default valiues there in QEMU or not (especially with PCI code resetting it to 0 now anyway). So I think the way to go is to get rid of these lines and don't bother with it until we find something that depends on defaults but that's very unlilkely as even in PCs the BIOS would program it or if you boot an OS then that will either use legacy mode or it programs the device as a PCI deviec and sets BARs. I can't imagine anything just setting native mode and then not change BARs and on real hardware it can't do that before firmware that most likely has already overwritten defaults according to BIOS setup where you can usually select native or legacy mode so what we need to match here is how it is after firmware. >>>>>> Move setting the BARs to a callback on writing the PCI config regsiter >>>>>> that sets the compatibility mode (which firmwares needing this mode >>>>>> seem to do) and fix their values to program it to use legacy port >>>>>> numbers. As noted in a comment, we only do this when the BARs were >>>>>> unset before, because logs from real machine show this is how real >>>>>> chip works, even if it contradicts the data sheet which is not very >>>>>> clear about this. >>>>>> >>>>>> Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> >>>>>> --- >>>>>> hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- >>>>>> 1 file changed, 30 insertions(+), 5 deletions(-) >>>>>> >>>>>> diff --git a/hw/ide/via.c b/hw/ide/via.c >>>>>> index fff23803a6..8186190207 100644 >>>>>> --- a/hw/ide/via.c >>>>>> +++ b/hw/ide/via.c >>>>>> @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) >>>>>> pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | >>>>>> PCI_STATUS_DEVSEL_MEDIUM); >>>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); >>>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); >>>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); >>>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); >>>>>> - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: >>>>>> 20-23h */ >>>>>> pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); >>>>>> /* IDE chip enable, IDE configuration 1/2, IDE FIFO >>>>>> Configuration*/ >>>>>> @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) >>>>>> pci_set_long(pci_conf + 0xc0, 0x00020001); >>>>>> } >>>>>> +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, >>>>>> + uint32_t val, int len) >>>>>> +{ >>>>>> + pci_default_write_config(pd, addr, val, len); >>>>>> + /* >>>>>> + * Only set BARs if they are unset. Logs from real hardware show >>>>>> that >>>>>> + * writing class_prog to enable compatibility mode after BARs were >>>>>> set >>>>>> + * (possibly by firmware) it will use ports set by BARs not ISA >>>>>> ports >>>>>> + * (e.g. pegasos2 Linux does this and calls it non-100% native >>>>>> mode). >>>>> >>>>> Can you remind me again where the references are to non-100% native >>>>> mode? The only thing I can find in Linux is >>>>> https://github.com/torvalds/linux/blob/master/arch/powerpc/platforms/chrp/pci.c#L360 >>>>> but that simply forces a switch to legacy mode, with no mention of >>>>> "non-100% native mode". >>>> >>>> It was discussed somewhere in the via-ide thread we had when this was >>>> last touched for pegasos2 in March 2020. Basically the non-100% native >>>> mode is when ports are set by BARs but IRQs are still hard coded to >>>> 14-15. Linux can work with all 3 possible modes: legacy (both ports and >>>> IRQs are hard coded to ISA values), native (using BARs and PCI config >>>> 0x3c for a single interrupt for both channels, vt82c686 data sheet does >>>> not document this but vt8231 has a comment saying native mode only) and >>>> non-100% native mode where BARs are effective to set port addresses but >>>> IRQs don't respect 0x3c but use 14-15 as in legacy mode. Some machines >>>> only work in non-100% native mode such as pegasos2 and Linux has some >>>> quirks for this. Other OSes running on this machine work with what the >>>> firmware has set up and can't work with anything else so we need to >>>> emulate what those OSes want (which matches real hardware) because Linux >>>> can usually cope anyway. On pegasso2 MorphOS uses BARs but expects IRQ >>>> 14-15 which is what the firmware also sets up by setting mode to native >>>> in the PCI config of the IDE func yet IRQs are fixed at 14-15. Linux >>>> forces its driver to use legacy interrupts by setting mode back to legacy >>>> but it still uses BARs and this is what it calls non-100% native mode. On >>>> amigaone firmware sets legacy mode and AmigaOS does not change this but >>>> uses it with legacy ports and IRQs. Linux finds the same and uses legacy >>>> mode on amigaone. >>> >>> Thanks for summarising: there have been a number of threads and changes >>> over the years, so it is easy to lose track of where things are. From the >>> above then everything except MorphOS that explicitly sets legacy/native >>> mode should just work? >> >> No, everything but Linux (i.e. MorphOS and AmigaOS) only work with >> behaviour matching real hardware which is BARs + legacy interrupts. Linux >> may work with other settings but it also has fix ups to detect and work >> with the non-100% native mode as on real hardware. (I think this non-100% >> native mode was used hy the older Linus via ata driver before it was ported >> to libata and became pata_via as I remember seeing these in older logs but >> not in newer ones any more which just list IRQs and port numbers from which >> you can tell it's neither the legacy nor the native mode described in the >> data sheet but a mix between the two.) >> >>> Just to double check I don't think we ever managed to get the PCI >>> configuration information from real hardware to confirm how the hardware >>> thinks it is set? Is it possible to dump the PCI config space of the >>> PCI-ISA bridge and the VIA IDE from a real Pegasos2 booted into Smart >>> Firmware? You can get the information using the Forth below: >>> >>> >>> : dump-single-config ( addr ) >>> dup 100 + swap do i u. " : " type i " config-b@" $call-parent u. cr loop >>> ; >>> >>> " /pci@80000000/isa@C" open-dev to my-self >>> 800 c * 100 0 * + dump-single-config cr cr >>> 800 c * 100 1 * + dump-single-config cr cr >> >> I don't have dumps from firmware but got an lspci output from Linux from >> real pegasos2 which says: > > Ah no that won't work because Linux has already made changes to PCI > configuration space. What we need is the value from the firmware to confirm > exactly how the hardware has been configured before the kernel has loaded (as > you mentioned above). Who can we ask to get this information? > >> 0000:00:0c.1 IDE interface: VIA Technologies, Inc. >> VT82C586A/B/VT82C686/A/B/VT823x/A/C PIPC Bus Master IDE (rev 06) (prog-if >> 8a [Master SecP PriP]) >> Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- >> ParErr- Stepping- SERR- FastB2B- DisINTx- >> Status: Cap+ 66MHz- UDF- FastB2B+ ParErr- DEVSEL=medium >TAbort- >> <TAbort- <MAbort- >SERR- <PERR- INTx- >> Latency: 0 >> Interrupt: pin ? routed to IRQ 14 > > Now this is interesting - I wonder what "?" means here? Perhaps the switch to > compatibility mode has set the PCI interrupt pin to 0? Oh and what kernel > version is this? I think: [ 0.000000] Linux version 3.16.0-6-powerpc (debian-kernel@lists.debian.org) (gcc version 4.8.4 (Debian 4.8.4-1) ) #1 Debian 3.16.56-1+deb8u1 (2018-05-08) but I got logs in different files so not entirely sure. >> Region 0: [virtual] I/O ports at 1000 [size=8] >> Region 1: [virtual] I/O ports at 100c [size=4] >> Region 2: [virtual] I/O ports at 1010 [size=8] >> Region 3: [virtual] I/O ports at 101c [size=4] >> Region 4: I/O ports at 1020 [size=16] >> Capabilities: [c0] Power Management version 2 >> Flags: PMEClk- DSI- D1- D2- AuxCurrent=0mA >> PME(D0-,D1-,D2-,D3hot-,D3cold-) >> Status: D0 NoSoftRst- PME-Enable- DSel=0 DScale=0 PME- >> Kernel driver in use: pata_via >> >> This says prog_if is 0x8a (probably set by Linux fixup as firmware sets it >> to native) but still uses BARs for ports and legacy interrupts although >> only says 14 above but in dmesg it listed interruts for the channels as >> quoted in reply to Bernhard. Also got /proc infos from same machine that >> says: >> >> /proc/ioports: 00001000-00001007 : 0000:00:0c.1 >> /proc/ioports: 00001000-00001007 : pata_via >> /proc/ioports: 0000100c-0000100f : 0000:00:0c.1 >> /proc/ioports: 0000100c-0000100f : pata_via >> /proc/ioports: 00001010-00001017 : 0000:00:0c.1 >> /proc/ioports: 00001010-00001017 : pata_via >> /proc/ioports: 0000101c-0000101f : 0000:00:0c.1 >> /proc/ioports: 0000101c-0000101f : pata_via >> /proc/ioports: 00001020-0000102f : 0000:00:0c.1 >> /proc/ioports: 00001020-0000102f : pata_via > > Interesting. The PCI-IDE specification states that the BAR values are ignored > in compatibility mode, so it could be that they contain values left over from > the firmware they are not being actively decoded on the bus. They certainly > look similar to the values set by the firmware. These are the values set by the firmware as you can see from the device tree dump above. > In fact it would be possible to write some Forth to switch to compatibility > mode in the firmware and then try and access the BARs to confirm whether or > not the BARs are really being decoded. So Linux does that on boot by writing 0x8a to prog_if when forcing legacy mode in fix up and then still uses BARs so I think that confirms it works like that even in legacy mode unless the BARs are unset which is what happens on the amigaone that only writes 0x8a even though that should be the default (again you can check this with the emulated machine). >> /proc/interrupts: 14: 1847 i8259 14 Level pata_via >> /proc/interrupts: 15: 1210 i8259 15 Level pata_via >> >> So what other proof you still need to believe this is how it works on real >> machine? > > Again this would have been overwritten by the switch to compatibility mode. But I think AmigaOS and MorphOS do not switch to legacy mode and still use BARs and legacy IRQs as set by the firmware. These now boot so our device model is matching real hardware well enough. >>> With the corresponding MorphOS debug log that will help to confirm the >>> routing being used. Also it would be useful to get the output of any >>> MorphOS program that can dump the current kernel IRQ routing >>> configuration. >> >> MorphOS does not give much debug logs but it works currenntly as well as >> AmigaOS and Linux so what we have now is good enough and this patch does >> not break that. > > It looks like you can get PCI debug logs, at least that seems work here on > QEMU. They would be very useful to understand the exact behaviour. > >>> The reason for wanting to double-check this now is because my work with >>> the Mac q800 machine showed that both legacy and modern IRQ routing >>> worked, so it would be good to confirm exactly what IRQ is currently >>> active on real hardware. >> >> What does q800 has to do with VT8231 and pegaos2? These are totally >> different hardware and software designed by separate people so other than >> you were looking at both there's no connection to consider between these. > > The point I was making was that there may be multiple IRQ routing > possibilities for the IDE controller, in which case it makes sense to be > absolutely sure which one is being used on real hardware. > >>>>>> + * But if 0x8a is written after reset without setting BARs then we >>>>>> want >>>>>> + * legacy ports (this is done by AmigaOne firmware for example). >>>>>> + */ >>>>> >>>>> What should happen here is that writing 0x8a should *always* switch to >>>>> legacy mode, so the BARs are unused... >>>> >>>> Yes, but as we've found before that can't be emulated in QEMU due to ISA >>>> emulation being static and only allows adding ports but not removing them >>>> later >>> >>> Fortunately this bug has been fixed, so it should now be possible using >>> portio_list_*() functions: see >>> https://gitlab.com/qemu-project/qemu/-/commit/690705ca0b0f1ed24a34ccd14c9866fbe47c69a6. >>> With that (and a bit of refactoring to allow the sharing of the IDE ioport >>> list) I was able to switch between compatibility and native modes at will >>> in my PoC. Once that is working then it doesn't matter if the default BARs >>> aren't set correctly. >> >> Great but maybe that's not needed because firmware and guests usually >> configure this once at boot then use it with that setting without ever >> switching it again so this simple patch allows that to work without >> breaking what's there and tested already so I'd like this to be merged >> first with the amigaone machine then if you want to rewrite it later to >> emulate the chip more closely then we have at least two test cases with >> pegasos2 and amigaone to verify evertyhing still works correctly. But since >> this patch allows all guests to boot it may be a waste of time to try to >> patch this emulation further to add functionality that won't be used by >> anything. But if it keeps working like this patch does and still boots the >> guests this one allows I don't care that much about that. > > The ultimate aim is to combine mine and Bernhard's work so that the > compatibility mode/PCI native mode switch is handle automatically by the PCI > IDE core, so it's something that would not just be restricted to the VIA. OK but that should be done separately and not intermixed/blocking this series as this model can later be changed but I'd like to get amigaone working now so a later goal should not prevent that now. I'm not sure your goal can be achieved though because these different south bridges may have some differences so not sure they can be handled the same way. I'd expect Intel PIIX models to be similar and VIA chips be similar so these can share code within the same family but just as Intel ac97 is different than VIA ac97 their IDE parts may also not behave the same so these may need to be modelled separately. We can try to find that out as long as it does not block people using these machines now. Regards, BALATON Zoltan > ATB, > > Mark.
diff --git a/hw/ide/via.c b/hw/ide/via.c index fff23803a6..8186190207 100644 --- a/hw/ide/via.c +++ b/hw/ide/via.c @@ -132,11 +132,6 @@ static void via_ide_reset(DeviceState *dev) pci_set_word(pci_conf + PCI_STATUS, PCI_STATUS_FAST_BACK | PCI_STATUS_DEVSEL_MEDIUM); - pci_set_long(pci_conf + PCI_BASE_ADDRESS_0, 0x000001f0); - pci_set_long(pci_conf + PCI_BASE_ADDRESS_1, 0x000003f4); - pci_set_long(pci_conf + PCI_BASE_ADDRESS_2, 0x00000170); - pci_set_long(pci_conf + PCI_BASE_ADDRESS_3, 0x00000374); - pci_set_long(pci_conf + PCI_BASE_ADDRESS_4, 0x0000cc01); /* BMIBA: 20-23h */ pci_set_long(pci_conf + PCI_INTERRUPT_LINE, 0x0000010e); /* IDE chip enable, IDE configuration 1/2, IDE FIFO Configuration*/ @@ -159,6 +154,35 @@ static void via_ide_reset(DeviceState *dev) pci_set_long(pci_conf + 0xc0, 0x00020001); } +static void via_ide_cfg_write(PCIDevice *pd, uint32_t addr, + uint32_t val, int len) +{ + pci_default_write_config(pd, addr, val, len); + /* + * Only set BARs if they are unset. Logs from real hardware show that + * writing class_prog to enable compatibility mode after BARs were set + * (possibly by firmware) it will use ports set by BARs not ISA ports + * (e.g. pegasos2 Linux does this and calls it non-100% native mode). + * But if 0x8a is written after reset without setting BARs then we want + * legacy ports (this is done by AmigaOne firmware for example). + */ + if (addr == PCI_CLASS_PROG && val == 0x8a && + pci_get_long(pd->config + PCI_BASE_ADDRESS_0) == + PCI_BASE_ADDRESS_SPACE_IO) { + pci_set_long(pd->config + PCI_BASE_ADDRESS_0, 0x1f0 + | PCI_BASE_ADDRESS_SPACE_IO); + pci_set_long(pd->config + PCI_BASE_ADDRESS_1, 0x3f6 + | PCI_BASE_ADDRESS_SPACE_IO); + pci_set_long(pd->config + PCI_BASE_ADDRESS_2, 0x170 + | PCI_BASE_ADDRESS_SPACE_IO); + pci_set_long(pd->config + PCI_BASE_ADDRESS_3, 0x376 + | PCI_BASE_ADDRESS_SPACE_IO); + /* BMIBA: 20-23h */ + pci_set_long(pd->config + PCI_BASE_ADDRESS_4, 0xcc00 + | PCI_BASE_ADDRESS_SPACE_IO); + } +} + static void via_ide_realize(PCIDevice *dev, Error **errp) { PCIIDEState *d = PCI_IDE(dev); @@ -221,6 +245,7 @@ static void via_ide_class_init(ObjectClass *klass, void *data) /* Reason: only works as function of VIA southbridge */ dc->user_creatable = false; + k->config_write = via_ide_cfg_write; k->realize = via_ide_realize; k->exit = via_ide_exitfn; k->vendor_id = PCI_VENDOR_ID_VIA;
The initial value for BARs were set in reset method for emulating legacy mode at start but this does not work because PCI code resets BARs after calling device reset method. Additionally the values written to BARs were also wrong. Move setting the BARs to a callback on writing the PCI config regsiter that sets the compatibility mode (which firmwares needing this mode seem to do) and fix their values to program it to use legacy port numbers. As noted in a comment, we only do this when the BARs were unset before, because logs from real machine show this is how real chip works, even if it contradicts the data sheet which is not very clear about this. Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> --- hw/ide/via.c | 35 ++++++++++++++++++++++++++++++----- 1 file changed, 30 insertions(+), 5 deletions(-)