| Message ID | 20250212112413.37553-2-philmd@linaro.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | hw/microblaze: Allow running cross-endian vCPUs | expand |
On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: > Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. > Endianness can be BIG, LITTLE or unspecified (default). > > Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> > --- > qapi/common.json | 16 ++++++++++++++++ > include/hw/qdev-properties-system.h | 7 +++++++ > hw/core/qdev-properties-system.c | 11 +++++++++++ > 3 files changed, 34 insertions(+) > > diff --git a/qapi/common.json b/qapi/common.json > index 6ffc7a37890..217feaaf683 100644 > --- a/qapi/common.json > +++ b/qapi/common.json > @@ -212,3 +212,19 @@ > ## > { 'struct': 'HumanReadableText', > 'data': { 'human-readable-text': 'str' } } > + > +## > +# @EndianMode: > +# > +# An enumeration of three options: little, big, and unspecified > +# > +# @little: Little endianness > +# > +# @big: Big endianness > +# > +# @unspecified: Endianness not specified > +# > +# Since: 10.0 > +## > +{ 'enum': 'EndianMode', > + 'data': [ 'little', 'big', 'unspecified' ] } Should 'unspecified' come first? ... so that it gets the value 0, i.e. when someone forgets to properly initialize a related variable, the chances are higher that it ends up as "unspecified" than as "little" ? Apart from that: Reviewed-by: Thomas Huth <thuth@redhat.com>
On 12/2/25 12:37, Thomas Huth wrote: > On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >> Endianness can be BIG, LITTLE or unspecified (default). >> >> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >> --- >> qapi/common.json | 16 ++++++++++++++++ >> include/hw/qdev-properties-system.h | 7 +++++++ >> hw/core/qdev-properties-system.c | 11 +++++++++++ >> 3 files changed, 34 insertions(+) >> >> diff --git a/qapi/common.json b/qapi/common.json >> index 6ffc7a37890..217feaaf683 100644 >> --- a/qapi/common.json >> +++ b/qapi/common.json >> @@ -212,3 +212,19 @@ >> ## >> { 'struct': 'HumanReadableText', >> 'data': { 'human-readable-text': 'str' } } >> + >> +## >> +# @EndianMode: >> +# >> +# An enumeration of three options: little, big, and unspecified >> +# >> +# @little: Little endianness >> +# >> +# @big: Big endianness >> +# >> +# @unspecified: Endianness not specified >> +# >> +# Since: 10.0 >> +## >> +{ 'enum': 'EndianMode', >> + 'data': [ 'little', 'big', 'unspecified' ] } > > Should 'unspecified' come first? ... so that it gets the value 0, i.e. > when someone forgets to properly initialize a related variable, the > chances are higher that it ends up as "unspecified" than as "little" ? Hmm but then in this series the dual-endianness regions are defined as: +static const MemoryRegionOps pic_ops[2] = { + [0 ... 1] = { + .read = pic_read, + .write = pic_write, + .endianness = DEVICE_BIG_ENDIAN, + .impl = { + .min_access_size = 4, + .max_access_size = 4, + }, + .valid = { + /* + * All XPS INTC registers are accessed through the PLB interface. + * The base address for these registers is provided by the + * configuration parameter, C_BASEADDR. Each register is 32 bits + * although some bits may be unused and is accessed on a 4-byte + * boundary offset from the base address. + */ + .min_access_size = 4, + .max_access_size = 4, + }, }, - .valid = { - .min_access_size = 4, - .max_access_size = 4 - } + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, }; We could declare the array using the confusing __MAX definition (at the price of wasting the ENDIAN_MODE_UNSPECIFIED entry) as: static const MemoryRegionOps pic_ops[ENDIAN_MODE__MAX - 1] { } WDYT? > Apart from that: > Reviewed-by: Thomas Huth <thuth@redhat.com> > Thanks!
On 12/2/25 12:43, Philippe Mathieu-Daudé wrote: > On 12/2/25 12:37, Thomas Huth wrote: >> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>> Endianness can be BIG, LITTLE or unspecified (default). >>> >>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>> --- >>> qapi/common.json | 16 ++++++++++++++++ >>> include/hw/qdev-properties-system.h | 7 +++++++ >>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>> 3 files changed, 34 insertions(+) >>> >>> diff --git a/qapi/common.json b/qapi/common.json >>> index 6ffc7a37890..217feaaf683 100644 >>> --- a/qapi/common.json >>> +++ b/qapi/common.json >>> @@ -212,3 +212,19 @@ >>> ## >>> { 'struct': 'HumanReadableText', >>> 'data': { 'human-readable-text': 'str' } } >>> + >>> +## >>> +# @EndianMode: >>> +# >>> +# An enumeration of three options: little, big, and unspecified >>> +# >>> +# @little: Little endianness >>> +# >>> +# @big: Big endianness >>> +# >>> +# @unspecified: Endianness not specified >>> +# >>> +# Since: 10.0 >>> +## >>> +{ 'enum': 'EndianMode', >>> + 'data': [ 'little', 'big', 'unspecified' ] } >> >> Should 'unspecified' come first? ... so that it gets the value 0, i.e. >> when someone forgets to properly initialize a related variable, the >> chances are higher that it ends up as "unspecified" than as "little" ? BTW I'm not sure QAPI guaranty enums are following an order (at least, as in C, I wouldn't rely on that assumption).
On Wed, Feb 12, 2025 at 01:02:18PM +0100, Philippe Mathieu-Daudé wrote: > On 12/2/25 12:43, Philippe Mathieu-Daudé wrote: > > On 12/2/25 12:37, Thomas Huth wrote: > > > On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: > > > > Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. > > > > Endianness can be BIG, LITTLE or unspecified (default). > > > > > > > > Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> > > > > --- > > > > qapi/common.json | 16 ++++++++++++++++ > > > > include/hw/qdev-properties-system.h | 7 +++++++ > > > > hw/core/qdev-properties-system.c | 11 +++++++++++ > > > > 3 files changed, 34 insertions(+) > > > > > > > > diff --git a/qapi/common.json b/qapi/common.json > > > > index 6ffc7a37890..217feaaf683 100644 > > > > --- a/qapi/common.json > > > > +++ b/qapi/common.json > > > > @@ -212,3 +212,19 @@ > > > > ## > > > > { 'struct': 'HumanReadableText', > > > > 'data': { 'human-readable-text': 'str' } } > > > > + > > > > +## > > > > +# @EndianMode: > > > > +# > > > > +# An enumeration of three options: little, big, and unspecified > > > > +# > > > > +# @little: Little endianness > > > > +# > > > > +# @big: Big endianness > > > > +# > > > > +# @unspecified: Endianness not specified > > > > +# > > > > +# Since: 10.0 > > > > +## > > > > +{ 'enum': 'EndianMode', > > > > + 'data': [ 'little', 'big', 'unspecified' ] } > > > > > > Should 'unspecified' come first? ... so that it gets the value 0, > > > i.e. when someone forgets to properly initialize a related variable, > > > the chances are higher that it ends up as "unspecified" than as > > > "little" ? > > BTW I'm not sure QAPI guaranty enums are following an order > (at least, as in C, I wouldn't rely on that assumption). If we don't document a guaranteed order IMHO we should, mostly just for the sake for guaranteeing exactly what will be the 0 value . It is pretty common to want a particular enum constant to be special default for the 0 value. It allows enums to be retrofitted into existing code, with confidence that any code forgetting to initialize a variable/field will get the special default. Missed initialization is relatively common as a C bug, and we use -ftrivial-auto-var-init=zero to give well defined (usually safe) semantics. With regards, Daniel
On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: > On 12/2/25 12:37, Thomas Huth wrote: >> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>> Endianness can be BIG, LITTLE or unspecified (default). >>> >>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>> --- >>> qapi/common.json | 16 ++++++++++++++++ >>> include/hw/qdev-properties-system.h | 7 +++++++ >>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>> 3 files changed, 34 insertions(+) >>> >>> diff --git a/qapi/common.json b/qapi/common.json >>> index 6ffc7a37890..217feaaf683 100644 >>> --- a/qapi/common.json >>> +++ b/qapi/common.json >>> @@ -212,3 +212,19 @@ >>> ## >>> { 'struct': 'HumanReadableText', >>> 'data': { 'human-readable-text': 'str' } } >>> + >>> +## >>> +# @EndianMode: >>> +# >>> +# An enumeration of three options: little, big, and unspecified >>> +# >>> +# @little: Little endianness >>> +# >>> +# @big: Big endianness >>> +# >>> +# @unspecified: Endianness not specified >>> +# >>> +# Since: 10.0 >>> +## >>> +{ 'enum': 'EndianMode', >>> + 'data': [ 'little', 'big', 'unspecified' ] } >> >> Should 'unspecified' come first? ... so that it gets the value 0, i.e. when >> someone forgets to properly initialize a related variable, the chances are >> higher that it ends up as "unspecified" than as "little" ? > > Hmm but then in this series the dual-endianness regions are defined as: > > +static const MemoryRegionOps pic_ops[2] = { > + [0 ... 1] = { This is already confusing as you'd have to know that 0 and 1 here means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants here as well might be clearer). It's easy to miss what this does so maybe repeating the definitions for each case would be longer but less confusing and then it does not matter what the values are. Or what uses the ops.endianness now should look at the property introduced by this patch to avoid having to propagate it like below and drop the ops.endianness? Or it should move to the memory region and set when the ops are assigned? Regards, BALATON Zoltan > + .read = pic_read, > + .write = pic_write, > + .endianness = DEVICE_BIG_ENDIAN, > + .impl = { > + .min_access_size = 4, > + .max_access_size = 4, > + }, > + .valid = { > + /* > + * All XPS INTC registers are accessed through the PLB > interface. > + * The base address for these registers is provided by the > + * configuration parameter, C_BASEADDR. Each register is 32 bits > + * although some bits may be unused and is accessed on a 4-byte > + * boundary offset from the base address. > + */ > + .min_access_size = 4, > + .max_access_size = 4, > + }, > }, > - .valid = { > - .min_access_size = 4, > - .max_access_size = 4 > - } > + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, > + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, > }; > > We could declare the array using the confusing __MAX definition > (at the price of wasting the ENDIAN_MODE_UNSPECIFIED entry) as: > > static const MemoryRegionOps pic_ops[ENDIAN_MODE__MAX - 1] { } > > WDYT? > >> Apart from that: >> Reviewed-by: Thomas Huth <thuth@redhat.com> >> > > Thanks! > >
On 12/2/25 13:56, BALATON Zoltan wrote: > On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >> On 12/2/25 12:37, Thomas Huth wrote: >>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>> Endianness can be BIG, LITTLE or unspecified (default). >>>> >>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>> --- >>>> qapi/common.json | 16 ++++++++++++++++ >>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>> 3 files changed, 34 insertions(+) >>>> >>>> diff --git a/qapi/common.json b/qapi/common.json >>>> index 6ffc7a37890..217feaaf683 100644 >>>> --- a/qapi/common.json >>>> +++ b/qapi/common.json >>>> @@ -212,3 +212,19 @@ >>>> ## >>>> { 'struct': 'HumanReadableText', >>>> 'data': { 'human-readable-text': 'str' } } >>>> + >>>> +## >>>> +# @EndianMode: >>>> +# >>>> +# An enumeration of three options: little, big, and unspecified >>>> +# >>>> +# @little: Little endianness >>>> +# >>>> +# @big: Big endianness >>>> +# >>>> +# @unspecified: Endianness not specified >>>> +# >>>> +# Since: 10.0 >>>> +## >>>> +{ 'enum': 'EndianMode', >>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>> >>> Should 'unspecified' come first? ... so that it gets the value 0, >>> i.e. when someone forgets to properly initialize a related variable, >>> the chances are higher that it ends up as "unspecified" than as >>> "little" ? >> >> Hmm but then in this series the dual-endianness regions are defined as: >> >> +static const MemoryRegionOps pic_ops[2] = { >> + [0 ... 1] = { > > This is already confusing as you'd have to know that 0 and 1 here means > ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants here as > well might be clearer). It's easy to miss what this does so maybe > repeating the definitions for each case would be longer but less > confusing and then it does not matter what the values are. > > Or what uses the ops.endianness now should look at the property > introduced by this patch to avoid having to propagate it like below and > drop the ops.endianness? Or it should move to the memory region and set > when the ops are assigned? I'm not understanding well what you ask, but maybe the answer is in v7 :)
On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: > On 12/2/25 13:56, BALATON Zoltan wrote: >> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>> On 12/2/25 12:37, Thomas Huth wrote: >>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>> >>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>> --- >>>>> qapi/common.json | 16 ++++++++++++++++ >>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>> 3 files changed, 34 insertions(+) >>>>> >>>>> diff --git a/qapi/common.json b/qapi/common.json >>>>> index 6ffc7a37890..217feaaf683 100644 >>>>> --- a/qapi/common.json >>>>> +++ b/qapi/common.json >>>>> @@ -212,3 +212,19 @@ >>>>> ## >>>>> { 'struct': 'HumanReadableText', >>>>> 'data': { 'human-readable-text': 'str' } } >>>>> + >>>>> +## >>>>> +# @EndianMode: >>>>> +# >>>>> +# An enumeration of three options: little, big, and unspecified >>>>> +# >>>>> +# @little: Little endianness >>>>> +# >>>>> +# @big: Big endianness >>>>> +# >>>>> +# @unspecified: Endianness not specified >>>>> +# >>>>> +# Since: 10.0 >>>>> +## >>>>> +{ 'enum': 'EndianMode', >>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>> >>>> Should 'unspecified' come first? ... so that it gets the value 0, >>>> i.e. when someone forgets to properly initialize a related variable, >>>> the chances are higher that it ends up as "unspecified" than as >>>> "little" ? >>> >>> Hmm but then in this series the dual-endianness regions are defined as: >>> >>> +static const MemoryRegionOps pic_ops[2] = { >>> + [0 ... 1] = { >> >> This is already confusing as you'd have to know that 0 and 1 here >> means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants >> here as well might be clearer). It's easy to miss what this does so At this point 0 / 1 only mean "from the index #0 included to the index #1 included", 0 being the first one and 1 the last one. >> maybe repeating the definitions for each case would be longer but less >> confusing and then it does not matter what the values are. This is what I tried to do with: + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, }; but in v7 we are back of picking an arbitrary value. >> Or what uses the ops.endianness now should look at the property >> introduced by this patch to avoid having to propagate it like below >> and drop the ops.endianness? Or it should move to the memory region >> and set when the ops are assigned? > > I'm not understanding well what you ask, but maybe the answer is in v7 :)
On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: > On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >> On 12/2/25 13:56, BALATON Zoltan wrote: >>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>> >>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>> --- >>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>> 3 files changed, 34 insertions(+) >>>>>> >>>>>> diff --git a/qapi/common.json b/qapi/common.json >>>>>> index 6ffc7a37890..217feaaf683 100644 >>>>>> --- a/qapi/common.json >>>>>> +++ b/qapi/common.json >>>>>> @@ -212,3 +212,19 @@ >>>>>> ## >>>>>> { 'struct': 'HumanReadableText', >>>>>> 'data': { 'human-readable-text': 'str' } } >>>>>> + >>>>>> +## >>>>>> +# @EndianMode: >>>>>> +# >>>>>> +# An enumeration of three options: little, big, and unspecified >>>>>> +# >>>>>> +# @little: Little endianness >>>>>> +# >>>>>> +# @big: Big endianness >>>>>> +# >>>>>> +# @unspecified: Endianness not specified >>>>>> +# >>>>>> +# Since: 10.0 >>>>>> +## >>>>>> +{ 'enum': 'EndianMode', >>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>> >>>>> Should 'unspecified' come first? ... so that it gets the value 0, i.e. >>>>> when someone forgets to properly initialize a related variable, the >>>>> chances are higher that it ends up as "unspecified" than as "little" ? >>>> >>>> Hmm but then in this series the dual-endianness regions are defined as: >>>> >>>> +static const MemoryRegionOps pic_ops[2] = { >>>> + [0 ... 1] = { >>> >>> This is already confusing as you'd have to know that 0 and 1 here means >>> ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants here as well >>> might be clearer). It's easy to miss what this does so > > At this point 0 / 1 only mean "from the index #0 included to the index > #1 included", 0 being the first one and 1 the last one. > >>> maybe repeating the definitions for each case would be longer but less >>> confusing and then it does not matter what the values are. > > This is what I tried to do with: > > + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, > + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, > }; > > but in v7 we are back of picking an arbitrary value. > >>> Or what uses the ops.endianness now should look at the property introduced >>> by this patch to avoid having to propagate it like below and drop the >>> ops.endianness? Or it should move to the memory region and set when the >>> ops are assigned? >> >> I'm not understanding well what you ask, but maybe the answer is in v7 :) I'm not sure I understand it well either. I think what I was asking about is the same as what Thomas asked if this could be avoided to make it necessary to allocate two separate ops for this. Looks like from now on this ops struct should really loose the endianness value and this should be assigned when the ops is added to the io region because that's where it decides which endianness is it based on the property added in this series. But I don't know if that could be done or would need deeper changes as what later uses this ops struct might not have access to the property and if we have a single ops struct it may need to be copied to set different endianness intstead of just referencing it. So I'm not sure there's a better way but I think this change makes an already cryptic boiler plate even more confusing for people less knowledgeable about QEMU and C programming so it makes even harder to write devices. But as long as it's just a few devices that need to work with different endianness then it might be OK. But wasn't that what NATIVE_ENDIAN was meant for? What can't that be kept then? Regards, BALATON Zoltan
On 12/2/25 17:23, BALATON Zoltan wrote: > On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >> On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >>> On 12/2/25 13:56, BALATON Zoltan wrote: >>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>>> >>>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>>> --- >>>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>>> 3 files changed, 34 insertions(+) >>>>>>> +{ 'enum': 'EndianMode', >>>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>>> >>>>>> Should 'unspecified' come first? ... so that it gets the value 0, >>>>>> i.e. when someone forgets to properly initialize a related >>>>>> variable, the chances are higher that it ends up as "unspecified" >>>>>> than as "little" ? >>>>> >>>>> Hmm but then in this series the dual-endianness regions are defined >>>>> as: >>>>> >>>>> +static const MemoryRegionOps pic_ops[2] = { >>>>> + [0 ... 1] = { >>>> >>>> This is already confusing as you'd have to know that 0 and 1 here >>>> means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants >>>> here as well might be clearer). It's easy to miss what this does so >> >> At this point 0 / 1 only mean "from the index #0 included to the index >> #1 included", 0 being the first one and 1 the last one. >> >>>> maybe repeating the definitions for each case would be longer but >>>> less confusing and then it does not matter what the values are. >> >> This is what I tried to do with: >> >> + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, >> + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, >> }; >> >> but in v7 we are back of picking an arbitrary value. >> >>>> Or what uses the ops.endianness now should look at the property >>>> introduced by this patch to avoid having to propagate it like below >>>> and drop the ops.endianness? Or it should move to the memory region >>>> and set when the ops are assigned? >>> >>> I'm not understanding well what you ask, but maybe the answer is in >>> v7 :) > > I'm not sure I understand it well either. I think what I was asking > about is the same as what Thomas asked if this could be avoided to make > it necessary to allocate two separate ops for this. Looks like from now > on this ops struct should really loose the endianness value and this > should be assigned when the ops is added to the io region because that's > where it decides which endianness is it based on the property added in > this series. But I don't know if that could be done or would need deeper > changes as what later uses this ops struct might not have access to the > property and if we have a single ops struct it may need to be copied to > set different endianness intstead of just referencing it. So I'm not > sure there's a better way but I think this change makes an already > cryptic boiler plate even more confusing for people less knowledgeable > about QEMU and C programming so it makes even harder to write devices. > But as long as it's just a few devices that need to work with different > endianness then it might be OK. But wasn't that what NATIVE_ENDIAN was > meant for? What can't that be kept then? Moving toward a single binary able to run heterogeneous machines, we can't rely on a particular target endianness, so we need to remove DEVICE_NATIVE_ENDIAN. The endianness is a property a device / machine, not of the binary.
On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: > On 12/2/25 17:23, BALATON Zoltan wrote: >> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>> On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >>>> On 12/2/25 13:56, BALATON Zoltan wrote: >>>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>>>> >>>>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>>>> --- >>>>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>>>> 3 files changed, 34 insertions(+) > > >>>>>>>> +{ 'enum': 'EndianMode', >>>>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>>>> >>>>>>> Should 'unspecified' come first? ... so that it gets the value 0, i.e. >>>>>>> when someone forgets to properly initialize a related variable, the >>>>>>> chances are higher that it ends up as "unspecified" than as "little" ? >>>>>> >>>>>> Hmm but then in this series the dual-endianness regions are defined as: >>>>>> >>>>>> +static const MemoryRegionOps pic_ops[2] = { >>>>>> + [0 ... 1] = { >>>>> >>>>> This is already confusing as you'd have to know that 0 and 1 here means >>>>> ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants here as >>>>> well might be clearer). It's easy to miss what this does so >>> >>> At this point 0 / 1 only mean "from the index #0 included to the index >>> #1 included", 0 being the first one and 1 the last one. >>> >>>>> maybe repeating the definitions for each case would be longer but less >>>>> confusing and then it does not matter what the values are. >>> >>> This is what I tried to do with: >>> >>> + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, >>> + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, >>> }; >>> >>> but in v7 we are back of picking an arbitrary value. >>> >>>>> Or what uses the ops.endianness now should look at the property >>>>> introduced by this patch to avoid having to propagate it like below and >>>>> drop the ops.endianness? Or it should move to the memory region and set >>>>> when the ops are assigned? >>>> >>>> I'm not understanding well what you ask, but maybe the answer is in v7 :) >> >> I'm not sure I understand it well either. I think what I was asking about >> is the same as what Thomas asked if this could be avoided to make it >> necessary to allocate two separate ops for this. Looks like from now on >> this ops struct should really loose the endianness value and this should be >> assigned when the ops is added to the io region because that's where it >> decides which endianness is it based on the property added in this series. >> But I don't know if that could be done or would need deeper changes as what >> later uses this ops struct might not have access to the property and if we >> have a single ops struct it may need to be copied to set different >> endianness intstead of just referencing it. So I'm not sure there's a >> better way but I think this change makes an already cryptic boiler plate >> even more confusing for people less knowledgeable about QEMU and C >> programming so it makes even harder to write devices. But as long as it's >> just a few devices that need to work with different endianness then it >> might be OK. But wasn't that what NATIVE_ENDIAN was meant for? What can't >> that be kept then? > > Moving toward a single binary able to run heterogeneous machines, we > can't rely on a particular target endianness, so we need to remove > DEVICE_NATIVE_ENDIAN. The endianness is a property a device / machine, > not of the binary. So then can the behaviour of NATIVE_ENDIAN be changed to look at the machine endianness instead of replacing it with a constant? Or would that be too much overhead? If always looking up the endianness is not wanted could the ops declaration keep NATIVE_ENDIAN but when the ops is added to the memory region can the register function replace it with the appropriate constant (if needed copying the struct)? Or is there something that prevents doing that? Regards, BALATON Zoltan
On 12/02/2025 23.34, BALATON Zoltan wrote: > On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >> On 12/2/25 17:23, BALATON Zoltan wrote: >>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>> On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >>>>> On 12/2/25 13:56, BALATON Zoltan wrote: >>>>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>>>>> >>>>>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>>>>> --- >>>>>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>>>>> 3 files changed, 34 insertions(+) >> >> >>>>>>>>> +{ 'enum': 'EndianMode', >>>>>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>>>>> >>>>>>>> Should 'unspecified' come first? ... so that it gets the value 0, >>>>>>>> i.e. when someone forgets to properly initialize a related variable, >>>>>>>> the chances are higher that it ends up as "unspecified" than as >>>>>>>> "little" ? >>>>>>> >>>>>>> Hmm but then in this series the dual-endianness regions are defined as: >>>>>>> >>>>>>> +static const MemoryRegionOps pic_ops[2] = { >>>>>>> + [0 ... 1] = { >>>>>> >>>>>> This is already confusing as you'd have to know that 0 and 1 here >>>>>> means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants >>>>>> here as well might be clearer). It's easy to miss what this does so >>>> >>>> At this point 0 / 1 only mean "from the index #0 included to the index >>>> #1 included", 0 being the first one and 1 the last one. >>>> >>>>>> maybe repeating the definitions for each case would be longer but less >>>>>> confusing and then it does not matter what the values are. >>>> >>>> This is what I tried to do with: >>>> >>>> + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, >>>> + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, >>>> }; >>>> >>>> but in v7 we are back of picking an arbitrary value. >>>> >>>>>> Or what uses the ops.endianness now should look at the property >>>>>> introduced by this patch to avoid having to propagate it like below >>>>>> and drop the ops.endianness? Or it should move to the memory region >>>>>> and set when the ops are assigned? >>>>> >>>>> I'm not understanding well what you ask, but maybe the answer is in v7 :) >>> >>> I'm not sure I understand it well either. I think what I was asking about >>> is the same as what Thomas asked if this could be avoided to make it >>> necessary to allocate two separate ops for this. Looks like from now on >>> this ops struct should really loose the endianness value and this should >>> be assigned when the ops is added to the io region because that's where >>> it decides which endianness is it based on the property added in this >>> series. But I don't know if that could be done or would need deeper >>> changes as what later uses this ops struct might not have access to the >>> property and if we have a single ops struct it may need to be copied to >>> set different endianness intstead of just referencing it. So I'm not sure >>> there's a better way but I think this change makes an already cryptic >>> boiler plate even more confusing for people less knowledgeable about QEMU >>> and C programming so it makes even harder to write devices. But as long >>> as it's just a few devices that need to work with different endianness >>> then it might be OK. But wasn't that what NATIVE_ENDIAN was meant for? >>> What can't that be kept then? >> >> Moving toward a single binary able to run heterogeneous machines, we >> can't rely on a particular target endianness, so we need to remove >> DEVICE_NATIVE_ENDIAN. The endianness is a property a device / machine, >> not of the binary. > > So then can the behaviour of NATIVE_ENDIAN be changed to look at the machine > endianness instead of replacing it with a constant? No, that does not work. First, the machine knows about its devices, but a device should not know about the wiring of the global machine (just like in real life). Second, imagine a board with e.g. a big endian main CPU and a little endian service processor - how should a device know the right endianness here? Thomas > Or would that be too > much overhead? If always looking up the endianness is not wanted could the > ops declaration keep NATIVE_ENDIAN IMHO we should get rid of NATIVE_ENDIAN completely since there is no "native" endian in multi-CPU boards. Thomas
On Thu, 13 Feb 2025, Thomas Huth wrote: > On 12/02/2025 23.34, BALATON Zoltan wrote: >> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>> On 12/2/25 17:23, BALATON Zoltan wrote: >>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>> On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >>>>>> On 12/2/25 13:56, BALATON Zoltan wrote: >>>>>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>>>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>>>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. >>>>>>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>>>>>> >>>>>>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>>>>>> --- >>>>>>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>>>>>> 3 files changed, 34 insertions(+) >>> >>> >>>>>>>>>> +{ 'enum': 'EndianMode', >>>>>>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>>>>>> >>>>>>>>> Should 'unspecified' come first? ... so that it gets the value 0, >>>>>>>>> i.e. when someone forgets to properly initialize a related variable, >>>>>>>>> the chances are higher that it ends up as "unspecified" than as >>>>>>>>> "little" ? >>>>>>>> >>>>>>>> Hmm but then in this series the dual-endianness regions are defined >>>>>>>> as: >>>>>>>> >>>>>>>> +static const MemoryRegionOps pic_ops[2] = { >>>>>>>> + [0 ... 1] = { >>>>>>> >>>>>>> This is already confusing as you'd have to know that 0 and 1 here >>>>>>> means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those constants >>>>>>> here as well might be clearer). It's easy to miss what this does so >>>>> >>>>> At this point 0 / 1 only mean "from the index #0 included to the index >>>>> #1 included", 0 being the first one and 1 the last one. >>>>> >>>>>>> maybe repeating the definitions for each case would be longer but less >>>>>>> confusing and then it does not matter what the values are. >>>>> >>>>> This is what I tried to do with: >>>>> >>>>> + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, >>>>> + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, >>>>> }; >>>>> >>>>> but in v7 we are back of picking an arbitrary value. >>>>> >>>>>>> Or what uses the ops.endianness now should look at the property >>>>>>> introduced by this patch to avoid having to propagate it like below >>>>>>> and drop the ops.endianness? Or it should move to the memory region >>>>>>> and set when the ops are assigned? >>>>>> >>>>>> I'm not understanding well what you ask, but maybe the answer is in v7 >>>>>> :) >>>> >>>> I'm not sure I understand it well either. I think what I was asking about >>>> is the same as what Thomas asked if this could be avoided to make it >>>> necessary to allocate two separate ops for this. Looks like from now on >>>> this ops struct should really loose the endianness value and this should >>>> be assigned when the ops is added to the io region because that's where >>>> it decides which endianness is it based on the property added in this >>>> series. But I don't know if that could be done or would need deeper >>>> changes as what later uses this ops struct might not have access to the >>>> property and if we have a single ops struct it may need to be copied to >>>> set different endianness intstead of just referencing it. So I'm not sure >>>> there's a better way but I think this change makes an already cryptic >>>> boiler plate even more confusing for people less knowledgeable about QEMU >>>> and C programming so it makes even harder to write devices. But as long >>>> as it's just a few devices that need to work with different endianness >>>> then it might be OK. But wasn't that what NATIVE_ENDIAN was meant for? >>>> What can't that be kept then? >>> >>> Moving toward a single binary able to run heterogeneous machines, we >>> can't rely on a particular target endianness, so we need to remove >>> DEVICE_NATIVE_ENDIAN. The endianness is a property a device / machine, >>> not of the binary. >> >> So then can the behaviour of NATIVE_ENDIAN be changed to look at the >> machine endianness instead of replacing it with a constant? > > No, that does not work. First, the machine knows about its devices, but a > device should not know about the wiring of the global machine (just like in > real life). That means all devices should be either big or little endian and there should be no native endian ones. Why do we have those then? That's why this endianness property should either be removed from ops and only attached to it when added to a machine if needed or kept to show which machines it can be attached to: only big, little or both endian which is what it seems to be doing now. > Second, imagine a board with e.g. a big endian main CPU and a > little endian service processor - how should a device know the right > endianness here? How would that work with this series? So the proposed solution is to double the devices now marked as NATIVE_ENDIAN to have a big and a little endian variant for them so the board can choose? That does not exist in real as you wrote, there's only one device so then this is probably not the right way to model it. >> Or would that be too much overhead? If always looking up the endianness is >> not wanted could the ops declaration keep NATIVE_ENDIAN > > IMHO we should get rid of NATIVE_ENDIAN completely since there is no "native" > endian in multi-CPU boards. If we say NATIVE_ENDIAN means that the device can be attached to either big or little endian machine then we can keep this constant but when adding the ops to a memory region the board has to then decide which endianness it is and replace it with either big or little. Then we don't need two versions of the same device and NATIVE_ENDIAN means that the device can be used in both machines. In real life probably all devices can be used with either CPU and if they are accessed in little or big endian is only determinded by how they are wired on the board. So the device endianness only means what endianness the device expects for something (what exactly? e.g. a video chip may have a frame buffer and a registers area with different endianness). So this should be the board that decides this not the device. Therefore it may not need to be defined when MemoryRegionOps is defined at all (or only as a hint to show what the device expects normally) and then memory_region_init_io which takes the MemoryRegionOps should also take an endianness corresponding the board and set it at that point. It can warn if the device endianness does not match what the board sets but you can still connect a big endian device to a little endian CPU as long as the drivers write the right values or the data lines are connected the right way, the latter of which corresponds to NATIVE_ENDIAN now as the conversion is done by the wiring so drivers don't need to care. But if it's simpler to just double the few devices that need to be used this way then it's a possible solution but if there's a cleaner one with not much more complexity then maybe that should be considered, because the way to define these doubled devices is a bit confusing for new people (on top of that defining devices is already confusing with the lot of boiler plate code needed). So if this could be kept simpler that would be a good thing IMO. Regards, BALATON Zoltan
On 13/2/25 14:59, BALATON Zoltan wrote: > On Thu, 13 Feb 2025, Thomas Huth wrote: >> On 12/02/2025 23.34, BALATON Zoltan wrote: >>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>> On 12/2/25 17:23, BALATON Zoltan wrote: >>>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>>> On 12/2/25 14:53, Philippe Mathieu-Daudé wrote: >>>>>>> On 12/2/25 13:56, BALATON Zoltan wrote: >>>>>>>> On Wed, 12 Feb 2025, Philippe Mathieu-Daudé wrote: >>>>>>>>> On 12/2/25 12:37, Thomas Huth wrote: >>>>>>>>>> On 12/02/2025 12.24, Philippe Mathieu-Daudé wrote: >>>>>>>>>>> Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() >>>>>>>>>>> macros. >>>>>>>>>>> Endianness can be BIG, LITTLE or unspecified (default). >>>>>>>>>>> >>>>>>>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> >>>>>>>>>>> --- >>>>>>>>>>> qapi/common.json | 16 ++++++++++++++++ >>>>>>>>>>> include/hw/qdev-properties-system.h | 7 +++++++ >>>>>>>>>>> hw/core/qdev-properties-system.c | 11 +++++++++++ >>>>>>>>>>> 3 files changed, 34 insertions(+) >>>> >>>> >>>>>>>>>>> +{ 'enum': 'EndianMode', >>>>>>>>>>> + 'data': [ 'little', 'big', 'unspecified' ] } >>>>>>>>>> >>>>>>>>>> Should 'unspecified' come first? ... so that it gets the value >>>>>>>>>> 0, i.e. when someone forgets to properly initialize a related >>>>>>>>>> variable, the chances are higher that it ends up as >>>>>>>>>> "unspecified" than as "little" ? >>>>>>>>> >>>>>>>>> Hmm but then in this series the dual-endianness regions are >>>>>>>>> defined as: >>>>>>>>> >>>>>>>>> +static const MemoryRegionOps pic_ops[2] = { >>>>>>>>> + [0 ... 1] = { >>>>>>>> >>>>>>>> This is already confusing as you'd have to know that 0 and 1 >>>>>>>> here means ENDIAN_MODE_LITTLE and ENDIAN_MODE_BIG (using those >>>>>>>> constants here as well might be clearer). It's easy to miss what >>>>>>>> this does so >>>>>> >>>>>> At this point 0 / 1 only mean "from the index #0 included to the >>>>>> index >>>>>> #1 included", 0 being the first one and 1 the last one. >>>>>> >>>>>>>> maybe repeating the definitions for each case would be longer >>>>>>>> but less confusing and then it does not matter what the values are. >>>>>> >>>>>> This is what I tried to do with: >>>>>> >>>>>> + [ENDIAN_MODE_BIG].endianness = DEVICE_BIG_ENDIAN, >>>>>> + [ENDIAN_MODE_LITTLE].endianness = DEVICE_LITTLE_ENDIAN, >>>>>> }; >>>>>> >>>>>> but in v7 we are back of picking an arbitrary value. >>>>>> >>>>>>>> Or what uses the ops.endianness now should look at the property >>>>>>>> introduced by this patch to avoid having to propagate it like >>>>>>>> below and drop the ops.endianness? Or it should move to the >>>>>>>> memory region and set when the ops are assigned? >>>>>>> >>>>>>> I'm not understanding well what you ask, but maybe the answer is >>>>>>> in v7 :) >>>>> >>>>> I'm not sure I understand it well either. I think what I was asking >>>>> about is the same as what Thomas asked if this could be avoided to >>>>> make it necessary to allocate two separate ops for this. Looks like >>>>> from now on this ops struct should really loose the endianness >>>>> value and this should be assigned when the ops is added to the io >>>>> region because that's where it decides which endianness is it based >>>>> on the property added in this series. But I don't know if that >>>>> could be done or would need deeper changes as what later uses this >>>>> ops struct might not have access to the property and if we have a >>>>> single ops struct it may need to be copied to set different >>>>> endianness intstead of just referencing it. So I'm not sure there's >>>>> a better way but I think this change makes an already cryptic >>>>> boiler plate even more confusing for people less knowledgeable >>>>> about QEMU and C programming so it makes even harder to write >>>>> devices. But as long as it's just a few devices that need to work >>>>> with different endianness then it might be OK. But wasn't that what >>>>> NATIVE_ENDIAN was meant for? What can't that be kept then? >>>> >>>> Moving toward a single binary able to run heterogeneous machines, we >>>> can't rely on a particular target endianness, so we need to remove >>>> DEVICE_NATIVE_ENDIAN. The endianness is a property a device / machine, >>>> not of the binary. >>> >>> So then can the behaviour of NATIVE_ENDIAN be changed to look at the >>> machine endianness instead of replacing it with a constant? >> >> No, that does not work. First, the machine knows about its devices, >> but a device should not know about the wiring of the global machine >> (just like in real life). > > That means all devices should be either big or little endian and there > should be no native endian ones. Why do we have those then? That's why > this endianness property should either be removed from ops and only > attached to it when added to a machine if needed or kept to show which > machines it can be attached to: only big, little or both endian which is > what it seems to be doing now. Well, devices don't have endianness, endianness is more a property of the bus [*] between cpu <-> devices, how the bytes are serialized. QEMU devices have an endianness property so the core memory can access them using the proper ordering, see: $ git grep -wW memory_region_big_endian system/memory.c:356:static bool memory_region_big_endian(MemoryRegion *mr) system/memory.c-357-{ system/memory.c-358-#if TARGET_BIG_ENDIAN system/memory.c-359- return mr->ops->endianness != DEVICE_LITTLE_ENDIAN; system/memory.c-360-#else system/memory.c-361- return mr->ops->endianness == DEVICE_BIG_ENDIAN; system/memory.c-362-#endif system/memory.c-363-} -- system/memory.c=521=static MemTxResult access_with_adjusted_size(hwaddr addr, system/memory.c-522- uint64_t *value, system/memory.c-523- unsigned size, system/memory.c-524- unsigned access_size_min, system/memory.c-525- unsigned access_size_max, system/memory.c-526- MemTxResult (*access_fn) system/memory.c-527- (MemoryRegion *mr, system/memory.c-528- hwaddr addr, system/memory.c-529- uint64_t *value, system/memory.c-530- unsigned size, system/memory.c-531- signed shift, system/memory.c-532- uint64_t mask, system/memory.c-533- MemTxAttrs attrs), system/memory.c-534- MemoryRegion *mr, system/memory.c-535- MemTxAttrs attrs) system/memory.c-536-{ system/memory.c-537- uint64_t access_mask; system/memory.c-538- unsigned access_size; system/memory.c-539- unsigned i; system/memory.c-540- MemTxResult r = MEMTX_OK; system/memory.c-541- bool reentrancy_guard_applied = false; system/memory.c-542- system/memory.c-543- if (!access_size_min) { system/memory.c-544- access_size_min = 1; system/memory.c-545- } system/memory.c-546- if (!access_size_max) { system/memory.c-547- access_size_max = 4; system/memory.c-548- } system/memory.c-549- system/memory.c-550- /* Do not allow more than one simultaneous access to a device's IO Regions */ system/memory.c-551- if (mr->dev && !mr->disable_reentrancy_guard && system/memory.c-552- !mr->ram_device && !mr->ram && !mr->rom_device && !mr->readonly) { system/memory.c-553- if (mr->dev->mem_reentrancy_guard.engaged_in_io) { system/memory.c-554- warn_report_once("Blocked re-entrant IO on MemoryRegion: " system/memory.c-555- "%s at addr: 0x%" HWADDR_PRIX, system/memory.c-556- memory_region_name(mr), addr); system/memory.c-557- return MEMTX_ACCESS_ERROR; system/memory.c-558- } system/memory.c-559- mr->dev->mem_reentrancy_guard.engaged_in_io = true; system/memory.c-560- reentrancy_guard_applied = true; system/memory.c-561- } system/memory.c-562- system/memory.c-563- /* FIXME: support unaligned access? */ system/memory.c-564- access_size = MAX(MIN(size, access_size_max), access_size_min); system/memory.c-565- access_mask = MAKE_64BIT_MASK(0, access_size * 8); system/memory.c:566: if (memory_region_big_endian(mr)) { system/memory.c-567- for (i = 0; i < size; i += access_size) { system/memory.c-568- r |= access_fn(mr, addr + i, value, access_size, system/memory.c-569- (size - access_size - i) * 8, access_mask, attrs); system/memory.c-570- } system/memory.c-571- } else { system/memory.c-572- for (i = 0; i < size; i += access_size) { system/memory.c-573- r |= access_fn(mr, addr + i, value, access_size, i * 8, system/memory.c-574- access_mask, attrs); system/memory.c-575- } system/memory.c-576- } system/memory.c-577- if (mr->dev && reentrancy_guard_applied) { system/memory.c-578- mr->dev->mem_reentrancy_guard.engaged_in_io = false; system/memory.c-579- } system/memory.c-580- return r; system/memory.c-581-} [*] See chapter 2 "The Basics of Endianness", in particular section 2.2 'Endianness Definition by Bus Specification': https://www.intel.com/content/dam/www/public/us/en/documents/application-notes/ixp4xx-ixc1100-big-endian-little-endian-modes-note.pdf >> Second, imagine a board with e.g. a big endian main CPU and a little >> endian service processor - how should a device know the right >> endianness here? > > How would that work with this series? So the proposed solution is to > double the devices now marked as NATIVE_ENDIAN to have a big and a > little endian variant for them so the board can choose? That does not > exist in real as you wrote, there's only one device so then this is > probably not the right way to model it. > >>> Or would that be too much overhead? If always looking up the >>> endianness is not wanted could the ops declaration keep NATIVE_ENDIAN >> >> IMHO we should get rid of NATIVE_ENDIAN completely since there is no >> "native" endian in multi-CPU boards. > > If we say NATIVE_ENDIAN means that the device can be attached to either > big or little endian machine then we can keep this constant but when > adding the ops to a memory region the board has to then decide which > endianness it is and replace it with either big or little. Then we don't > need two versions of the same device and NATIVE_ENDIAN means that the > device can be used in both machines. > > In real life probably all devices can be used with either CPU and if > they are accessed in little or big endian is only determinded by how > they are wired on the board. So the device endianness only means what > endianness the device expects for something (what exactly? e.g. a video > chip may have a frame buffer and a registers area with different > endianness). So this should be the board that decides this not the > device. Therefore it may not need to be defined when MemoryRegionOps is > defined at all (or only as a hint to show what the device expects > normally) and then memory_region_init_io which takes the MemoryRegionOps > should also take an endianness corresponding the board and set it at > that point. It can warn if the device endianness does not match what the > board sets but you can still connect a big endian device to a little > endian CPU as long as the drivers write the right values or the data > lines are connected the right way, the latter of which corresponds to > NATIVE_ENDIAN now as the conversion is done by the wiring so drivers > don't need to care. > > But if it's simpler to just double the few devices that need to be used > this way then it's a possible solution but if there's a cleaner one with > not much more complexity then maybe that should be considered, because > the way to define these doubled devices is a bit confusing for new > people (on top of that defining devices is already confusing with the > lot of boiler plate code needed). So if this could be kept simpler that > would be a good thing IMO. > > Regards, > BALATON Zoltan
On 13/02/2025 14.59, BALATON Zoltan wrote: > On Thu, 13 Feb 2025, Thomas Huth wrote: >> On 12/02/2025 23.34, BALATON Zoltan wrote: [...] >>> So then can the behaviour of NATIVE_ENDIAN be changed to look at the >>> machine endianness instead of replacing it with a constant? >> >> No, that does not work. First, the machine knows about its devices, but a >> device should not know about the wiring of the global machine (just like >> in real life). > > That means all devices should be either big or little endian and there > should be no native endian ones. Why do we have those then? Some device can indeed be either big or little endian - think of devices that are synthesized in an FPGA for example. But in most cases, it rather depends on the bus wiring. Anyway, we need a config knob to allow either the one or the other endianness for certain devices. > That's why this > endianness property should either be removed from ops and only attached to > it when added to a machine if needed or kept to show which machines it can > be attached to: only big, little or both endian which is what it seems to be > doing now. Again, devices should not know about machines, not the other way round. So the device should offer a config switch (property) and the machine should set it to the value that it needs. >> Second, imagine a board with e.g. a big endian main CPU and a little >> endian service processor - how should a device know the right endianness >> here? > > How would that work with this series? So the proposed solution is to double > the devices now marked as NATIVE_ENDIAN to have a big and a little endian > variant for them so the board can choose? This is not doubling the devices. It just introduces a config property to let the machine switch the endianness. > That does not exist in real as you > wrote, there's only one device so then this is probably not the right way to > model it. Some devices can exist in both, big and little endian variants. We could of course create two devices for this, but that's nonsense if it can simply be handled by a property instead. >>> Or would that be too much overhead? If always looking up the endianness >>> is not wanted could the ops declaration keep NATIVE_ENDIAN >> >> IMHO we should get rid of NATIVE_ENDIAN completely since there is no >> "native" endian in multi-CPU boards. > > If we say NATIVE_ENDIAN means that the device can be attached to either big > or little endian machine then we can keep this constant but when adding the > ops to a memory region the board has to then decide which endianness it is > and replace it with either big or little. Then we don't need two versions of > the same device and NATIVE_ENDIAN means that the device can be used in both > machines. Well, it's currently the devices that are calling memory_region_init_io(). And since memory_region_init_io() does not copy the MemoryRegionOps struct, we need two implementations right now for this, one for big and one for little endian. So I think Philippe's series here is fine. But feel free to suggest clean up patches on top if you think that the memory_region_init_io() needs to be handled differently in QEMU everywhere. Thomas
On Thu, 13 Feb 2025, Thomas Huth wrote: > On 13/02/2025 14.59, BALATON Zoltan wrote: >> On Thu, 13 Feb 2025, Thomas Huth wrote: >>> On 12/02/2025 23.34, BALATON Zoltan wrote: > [...] >>>> So then can the behaviour of NATIVE_ENDIAN be changed to look at the >>>> machine endianness instead of replacing it with a constant? >>> >>> No, that does not work. First, the machine knows about its devices, but a >>> device should not know about the wiring of the global machine (just like >>> in real life). >> >> That means all devices should be either big or little endian and there >> should be no native endian ones. Why do we have those then? > > Some device can indeed be either big or little endian - think of devices that > are synthesized in an FPGA for example. But in most cases, it rather depends > on the bus wiring. Anyway, we need a config knob to allow either the one or > the other endianness for certain devices. > >> That's why this endianness property should either be removed from ops and >> only attached to it when added to a machine if needed or kept to show which >> machines it can be attached to: only big, little or both endian which is >> what it seems to be doing now. > > Again, devices should not know about machines, not the other way round. So > the device should offer a config switch (property) and the machine should set > it to the value that it needs. That would mean this endianness in ops should be set when the memory region is mapped in the machine not when defining the device. So all device ops should really be NATIVE_ENDIAN and only assigned an endianness when they are mapped based on the machine/bus/cpu they are connected to. >>> Second, imagine a board with e.g. a big endian main CPU and a little >>> endian service processor - how should a device know the right endianness >>> here? >> >> How would that work with this series? So the proposed solution is to double >> the devices now marked as NATIVE_ENDIAN to have a big and a little endian >> variant for them so the board can choose? > > This is not doubling the devices. It just introduces a config property to let > the machine switch the endianness. Yes, I've oversimpified, each ops has its own endianness config not the device. But since the endianness is not a property of the device or its regions but the bus it's connected to, this config switch may be at the wrong place now. >> That does not exist in real as you wrote, there's only one device so then >> this is probably not the right way to model it. > > Some devices can exist in both, big and little endian variants. We could of > course create two devices for this, but that's nonsense if it can simply be > handled by a property instead. > >>>> Or would that be too much overhead? If always looking up the endianness >>>> is not wanted could the ops declaration keep NATIVE_ENDIAN >>> >>> IMHO we should get rid of NATIVE_ENDIAN completely since there is no >>> "native" endian in multi-CPU boards. >> >> If we say NATIVE_ENDIAN means that the device can be attached to either big >> or little endian machine then we can keep this constant but when adding the >> ops to a memory region the board has to then decide which endianness it is >> and replace it with either big or little. Then we don't need two versions >> of the same device and NATIVE_ENDIAN means that the device can be used in >> both machines. > > Well, it's currently the devices that are calling memory_region_init_io(). That answers my question. Then it seems it's not so simple to set endianness when the device is mapped because it would need to be done somewhere else. It may still be possible but might be too much work to find it out. > And since memory_region_init_io() does not copy the MemoryRegionOps struct, > we need two implementations right now for this, one for big and one for > little endian. So I think Philippe's series here is fine. It could copy the MemoryRegionOps when needed but seems it's not memory_region_init_io that would need to handle that. Given the current way it's implemented doubling the ops region may be the simplest even if not the correct way to do it so it's OK if there's no simple alternative that's more correct. > But feel free to > suggest clean up patches on top if you think that the memory_region_init_io() > needs to be handled differently in QEMU everywhere. I think so as I've described above but not enough to try to solve it so I'm OK with Philippe's series if there's no other way to make it less like a workaround for something that could be done clearer. Looks like other way might be too complex for now. Regards, BALATON Zoltan
diff --git a/qapi/common.json b/qapi/common.json index 6ffc7a37890..217feaaf683 100644 --- a/qapi/common.json +++ b/qapi/common.json @@ -212,3 +212,19 @@ ## { 'struct': 'HumanReadableText', 'data': { 'human-readable-text': 'str' } } + +## +# @EndianMode: +# +# An enumeration of three options: little, big, and unspecified +# +# @little: Little endianness +# +# @big: Big endianness +# +# @unspecified: Endianness not specified +# +# Since: 10.0 +## +{ 'enum': 'EndianMode', + 'data': [ 'little', 'big', 'unspecified' ] } diff --git a/include/hw/qdev-properties-system.h b/include/hw/qdev-properties-system.h index 7ec37f6316c..ead4dfc2f02 100644 --- a/include/hw/qdev-properties-system.h +++ b/include/hw/qdev-properties-system.h @@ -30,6 +30,7 @@ extern const PropertyInfo qdev_prop_pcie_link_speed; extern const PropertyInfo qdev_prop_pcie_link_width; extern const PropertyInfo qdev_prop_cpus390entitlement; extern const PropertyInfo qdev_prop_iothread_vq_mapping_list; +extern const PropertyInfo qdev_prop_endian_mode; #define DEFINE_PROP_PCI_DEVFN(_n, _s, _f, _d) \ DEFINE_PROP_SIGNED(_n, _s, _f, _d, qdev_prop_pci_devfn, int32_t) @@ -97,4 +98,10 @@ extern const PropertyInfo qdev_prop_iothread_vq_mapping_list; DEFINE_PROP(_name, _state, _field, qdev_prop_iothread_vq_mapping_list, \ IOThreadVirtQueueMappingList *) +#define DEFINE_PROP_ENDIAN(_name, _state, _field, _default) \ + DEFINE_PROP_UNSIGNED(_name, _state, _field, _default, \ + qdev_prop_endian_mode, EndianMode) +#define DEFINE_PROP_ENDIAN_NODEFAULT(_name, _state, _field) \ + DEFINE_PROP_ENDIAN(_name, _state, _field, ENDIAN_MODE_UNSPECIFIED) + #endif diff --git a/hw/core/qdev-properties-system.c b/hw/core/qdev-properties-system.c index a96675beb0d..89f954f569e 100644 --- a/hw/core/qdev-properties-system.c +++ b/hw/core/qdev-properties-system.c @@ -1283,3 +1283,14 @@ const PropertyInfo qdev_prop_iothread_vq_mapping_list = { .set = set_iothread_vq_mapping_list, .release = release_iothread_vq_mapping_list, }; + +/* --- Endian modes */ + +const PropertyInfo qdev_prop_endian_mode = { + .name = "EndianMode", + .description = "Endian mode, big/little/unspecified", + .enum_table = &EndianMode_lookup, + .get = qdev_propinfo_get_enum, + .set = qdev_propinfo_set_enum, + .set_default_value = qdev_propinfo_set_default_value_enum, +};
Introduce the EndianMode type and the DEFINE_PROP_ENDIAN() macros. Endianness can be BIG, LITTLE or unspecified (default). Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org> --- qapi/common.json | 16 ++++++++++++++++ include/hw/qdev-properties-system.h | 7 +++++++ hw/core/qdev-properties-system.c | 11 +++++++++++ 3 files changed, 34 insertions(+)