Message ID | 20241023085852.1061031-4-mark.cave-ayland@ilande.co.uk (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | next-cube: more tidy-ups and improvements | expand |
On 23/10/24 05:58, Mark Cave-Ayland wrote: > Change the start of the next.mmio memory region so that it follows on directly > after the next.dma memory region, adjusting the address offsets in > next_mmio_read() and next_mmio_write() accordingly. > > Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> > --- > hw/m68k/next-cube.c | 28 ++++++++++++++-------------- > 1 file changed, 14 insertions(+), 14 deletions(-) > @@ -897,7 +897,7 @@ static void next_pc_realize(DeviceState *dev, Error **errp) > qdev_init_gpio_in(dev, next_irq, NEXT_NUM_IRQS); > > memory_region_init_io(&s->mmiomem, OBJECT(s), &next_mmio_ops, s, > - "next.mmio", 0xd0000); > + "next.mmio", 0x9000); Please mention 0xd0000 was incorrect, otherwise: Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
On 24/10/2024 03:42, Philippe Mathieu-Daudé wrote: > On 23/10/24 05:58, Mark Cave-Ayland wrote: >> Change the start of the next.mmio memory region so that it follows on directly >> after the next.dma memory region, adjusting the address offsets in >> next_mmio_read() and next_mmio_write() accordingly. >> >> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> >> --- >> hw/m68k/next-cube.c | 28 ++++++++++++++-------------- >> 1 file changed, 14 insertions(+), 14 deletions(-) > > >> @@ -897,7 +897,7 @@ static void next_pc_realize(DeviceState *dev, Error **errp) >> qdev_init_gpio_in(dev, next_irq, NEXT_NUM_IRQS); >> memory_region_init_io(&s->mmiomem, OBJECT(s), &next_mmio_ops, s, >> - "next.mmio", 0xd0000); >> + "next.mmio", 0x9000); > > Please mention 0xd0000 was incorrect, otherwise: > > Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org> This one is a little more fuzzy, since I'm not even sure that the new value 0x9000 is correct (there isn't much in the way of documentation to support this) - effectively the purpose of truncating this memory region is to remove the overlap with another so that MMIO accesses are consistently directed to the same handler. I'll have a think about how I can re-word this better for a v2. ATB, Mark.
Am Wed, 23 Oct 2024 09:58:19 +0100 schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: > Change the start of the next.mmio memory region so that it follows on directly > after the next.dma memory region, adjusting the address offsets in > next_mmio_read() and next_mmio_write() accordingly. > > Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> > --- > hw/m68k/next-cube.c | 28 ++++++++++++++-------------- > 1 file changed, 14 insertions(+), 14 deletions(-) > > diff --git a/hw/m68k/next-cube.c b/hw/m68k/next-cube.c > index 4e8e55a8bd..e1d94c1ce0 100644 > --- a/hw/m68k/next-cube.c > +++ b/hw/m68k/next-cube.c > @@ -266,23 +266,23 @@ static uint64_t next_mmio_read(void *opaque, hwaddr addr, unsigned size) > uint64_t val; > > switch (addr) { > - case 0x7000: > + case 0x2000: > /* DPRINTF("Read INT status: %x\n", s->int_status); */ > val = s->int_status; > break; > > - case 0x7800: > + case 0x2800: > DPRINTF("MMIO Read INT mask: %x\n", s->int_mask); > val = s->int_mask; > break; > > - case 0xc000 ... 0xc003: > - val = extract32(s->scr1, (4 - (addr - 0xc000) - size) << 3, > + case 0x7000 ... 0x7003: > + val = extract32(s->scr1, (4 - (addr - 0x7000) - size) << 3, > size << 3); > break; > > - case 0xd000 ... 0xd003: > - val = extract32(s->scr2, (4 - (addr - 0xd000) - size) << 3, > + case 0x8000 ... 0x8003: > + val = extract32(s->scr2, (4 - (addr - 0x8000) - size) << 3, > size << 3); > break; > > @@ -301,25 +301,25 @@ static void next_mmio_write(void *opaque, hwaddr addr, uint64_t val, > NeXTPC *s = NEXT_PC(opaque); > > switch (addr) { > - case 0x7000: > + case 0x2000: > DPRINTF("INT Status old: %x new: %x\n", s->int_status, > (unsigned int)val); > s->int_status = val; > break; > > - case 0x7800: > + case 0x2800: > DPRINTF("INT Mask old: %x new: %x\n", s->int_mask, (unsigned int)val); > s->int_mask = val; > break; Could you please add comments at the end of the "case" lines, stating which mmio addresses are handled in each case? Otherwise, it's harder to grep for certain addresses later. E.g: case 0x2800: /* 0x2007800 */ > @@ -1000,7 +1000,7 @@ static void next_cube_init(MachineState *machine) > sysbus_create_simple(TYPE_NEXTFB, 0x0B000000, NULL); > > /* MMIO */ > - sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02000000); > + sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02005000); Why 0x02005000 and not directly 0x02007000 ? I think there is another range at 0x02006000 related to the ethernet controller, so directly going with 0x02007000 might cause less friction later when we add the NIC? Thanks, Thomas
On 26/10/2024 08:56, Thomas Huth wrote: > Am Wed, 23 Oct 2024 09:58:19 +0100 > schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: > >> Change the start of the next.mmio memory region so that it follows on directly >> after the next.dma memory region, adjusting the address offsets in >> next_mmio_read() and next_mmio_write() accordingly. >> >> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> >> --- >> hw/m68k/next-cube.c | 28 ++++++++++++++-------------- >> 1 file changed, 14 insertions(+), 14 deletions(-) >> >> diff --git a/hw/m68k/next-cube.c b/hw/m68k/next-cube.c >> index 4e8e55a8bd..e1d94c1ce0 100644 >> --- a/hw/m68k/next-cube.c >> +++ b/hw/m68k/next-cube.c >> @@ -266,23 +266,23 @@ static uint64_t next_mmio_read(void *opaque, hwaddr addr, unsigned size) >> uint64_t val; >> >> switch (addr) { >> - case 0x7000: >> + case 0x2000: >> /* DPRINTF("Read INT status: %x\n", s->int_status); */ >> val = s->int_status; >> break; >> >> - case 0x7800: >> + case 0x2800: >> DPRINTF("MMIO Read INT mask: %x\n", s->int_mask); >> val = s->int_mask; >> break; >> >> - case 0xc000 ... 0xc003: >> - val = extract32(s->scr1, (4 - (addr - 0xc000) - size) << 3, >> + case 0x7000 ... 0x7003: >> + val = extract32(s->scr1, (4 - (addr - 0x7000) - size) << 3, >> size << 3); >> break; >> >> - case 0xd000 ... 0xd003: >> - val = extract32(s->scr2, (4 - (addr - 0xd000) - size) << 3, >> + case 0x8000 ... 0x8003: >> + val = extract32(s->scr2, (4 - (addr - 0x8000) - size) << 3, >> size << 3); >> break; >> >> @@ -301,25 +301,25 @@ static void next_mmio_write(void *opaque, hwaddr addr, uint64_t val, >> NeXTPC *s = NEXT_PC(opaque); >> >> switch (addr) { >> - case 0x7000: >> + case 0x2000: >> DPRINTF("INT Status old: %x new: %x\n", s->int_status, >> (unsigned int)val); >> s->int_status = val; >> break; >> >> - case 0x7800: >> + case 0x2800: >> DPRINTF("INT Mask old: %x new: %x\n", s->int_mask, (unsigned int)val); >> s->int_mask = val; >> break; > > Could you please add comments at the end of the "case" lines, stating which > mmio addresses are handled in each case? Otherwise, it's harder to grep for > certain addresses later. E.g: > > case 0x2800: /* 0x2007800 */ If you think it will help? Presumably this is to aid with comparing with other source code/documentation? >> @@ -1000,7 +1000,7 @@ static void next_cube_init(MachineState *machine) >> sysbus_create_simple(TYPE_NEXTFB, 0x0B000000, NULL); >> >> /* MMIO */ >> - sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02000000); >> + sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02005000); > > Why 0x02005000 and not directly 0x02007000 ? Before this patch the output of "info mtree" shows as follows: (qemu) info mtree address-space: cpu-memory-0 address-space: memory 0000000000000000-ffffffffffffffff (prio 0, i/o): system 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom 0000000000000000-000000000001ffff 0000000001000000-000000000101ffff (prio 0, rom): next.rom 0000000002000000-0000000002004fff (prio 0, i/o): next.dma 0000000002000000-00000000020cffff (prio 0, i/o): next.mmio 000000000200e000-000000000200efff (prio 0, i/o): next.kbd 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem 0000000002100000-000000000211ffff (prio 0, i/o): next.scr 0000000002114000-000000000211400f (prio 0, i/o): esp-regs 0000000002118000-0000000002118003 (prio 0, i/o): escc 0000000004000000-0000000007ffffff (prio 0, ram): next.ram 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 @next.bmapmem 0000000000000000-000000000000003f All this patch does is move the start of next.mmio to 0x2005000 to avoid the overlap. > I think there is another range at 0x02006000 related to the ethernet > controller, so directly going with 0x02007000 might cause less friction > later when we add the NIC? By the end of the series, everything except the "global" registers in next.mmio have their own memory region (or empty-slot if the target is unknown) so that "info mtree" output looks like this: (qemu) info mtree address-space: cpu-memory-0 address-space: memory 0000000000000000-ffffffffffffffff (prio 0, i/o): system 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom 0000000000000000-000000000001ffff 0000000001000000-000000000101ffff (prio 0, rom): next.rom 0000000002000000-0000000002004fff (prio 0, i/o): next.dma 0000000002005000-000000000200dfff (prio 0, i/o): next.mmio 000000000200e000-000000000200efff (prio 0, i/o): next.kbd 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem 0000000002106000-000000000210601f (prio 0, i/o): next.en 0000000002110000-000000000211000f (prio -10000, i/o): empty-slot 0000000002112000-000000000211200f (prio -10000, i/o): empty-slot 0000000002114000-000000000211403f (prio 0, i/o): next.scsi 0000000002114000-000000000211400f (prio 0, i/o): esp-regs 0000000002114020-0000000002114021 (prio 0, i/o): csrs 0000000002114108-000000000211410b (prio 0, i/o): next.floppy 0000000002118000-0000000002118003 (prio 0, i/o): escc 0000000002118004-0000000002118013 (prio -10000, i/o): empty-slot 000000000211a000-000000000211a003 (prio 0, i/o): next.timer 0000000004000000-0000000007ffffff (prio 0, ram): next.ram 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 @next.bmapmem 0000000000000000-000000000000003f In this case next.en is a dummy memory region which can easily be replaced with a proper device implementation: see the final version of next-cube.c after the series at https://gitlab.com/mcayland/qemu/-/blob/next-cube-improvements/hw/m68k/next-cube.c. ATB, Mark.
Am Sat, 26 Oct 2024 22:13:25 +0100 schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: > On 26/10/2024 08:56, Thomas Huth wrote: > > > Am Wed, 23 Oct 2024 09:58:19 +0100 > > schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: > > > >> Change the start of the next.mmio memory region so that it follows on directly > >> after the next.dma memory region, adjusting the address offsets in > >> next_mmio_read() and next_mmio_write() accordingly. > >> > >> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> > >> --- > >> hw/m68k/next-cube.c | 28 ++++++++++++++-------------- > >> 1 file changed, 14 insertions(+), 14 deletions(-) > >> > >> diff --git a/hw/m68k/next-cube.c b/hw/m68k/next-cube.c > >> index 4e8e55a8bd..e1d94c1ce0 100644 > >> --- a/hw/m68k/next-cube.c > >> +++ b/hw/m68k/next-cube.c > >> @@ -266,23 +266,23 @@ static uint64_t next_mmio_read(void *opaque, hwaddr addr, unsigned size) > >> uint64_t val; > >> > >> switch (addr) { > >> - case 0x7000: > >> + case 0x2000: > >> /* DPRINTF("Read INT status: %x\n", s->int_status); */ > >> val = s->int_status; > >> break; > >> > >> - case 0x7800: > >> + case 0x2800: > >> DPRINTF("MMIO Read INT mask: %x\n", s->int_mask); > >> val = s->int_mask; > >> break; > >> > >> - case 0xc000 ... 0xc003: > >> - val = extract32(s->scr1, (4 - (addr - 0xc000) - size) << 3, > >> + case 0x7000 ... 0x7003: > >> + val = extract32(s->scr1, (4 - (addr - 0x7000) - size) << 3, > >> size << 3); > >> break; > >> > >> - case 0xd000 ... 0xd003: > >> - val = extract32(s->scr2, (4 - (addr - 0xd000) - size) << 3, > >> + case 0x8000 ... 0x8003: > >> + val = extract32(s->scr2, (4 - (addr - 0x8000) - size) << 3, > >> size << 3); > >> break; > >> > >> @@ -301,25 +301,25 @@ static void next_mmio_write(void *opaque, hwaddr addr, uint64_t val, > >> NeXTPC *s = NEXT_PC(opaque); > >> > >> switch (addr) { > >> - case 0x7000: > >> + case 0x2000: > >> DPRINTF("INT Status old: %x new: %x\n", s->int_status, > >> (unsigned int)val); > >> s->int_status = val; > >> break; > >> > >> - case 0x7800: > >> + case 0x2800: > >> DPRINTF("INT Mask old: %x new: %x\n", s->int_mask, (unsigned int)val); > >> s->int_mask = val; > >> break; > > > > Could you please add comments at the end of the "case" lines, stating which > > mmio addresses are handled in each case? Otherwise, it's harder to grep for > > certain addresses later. E.g: > > > > case 0x2800: /* 0x2007800 */ > > If you think it will help? Presumably this is to aid with comparing with other source > code/documentation? Yes, it will help with 1) debugging code that is running in the guest (so you can find IO addresses that it is accessing more easily) and 2) help when comparing the code with "Previous": https://sourceforge.net/p/previous/code/HEAD/tree/trunk/src/ioMemTabNEXT.c#l36 > >> @@ -1000,7 +1000,7 @@ static void next_cube_init(MachineState *machine) > >> sysbus_create_simple(TYPE_NEXTFB, 0x0B000000, NULL); > >> > >> /* MMIO */ > >> - sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02000000); > >> + sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02005000); > > > > Why 0x02005000 and not directly 0x02007000 ? > > Before this patch the output of "info mtree" shows as follows: > > (qemu) info mtree > address-space: cpu-memory-0 > address-space: memory > 0000000000000000-ffffffffffffffff (prio 0, i/o): system > 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom > 0000000000000000-000000000001ffff > 0000000001000000-000000000101ffff (prio 0, rom): next.rom > 0000000002000000-0000000002004fff (prio 0, i/o): next.dma > 0000000002000000-00000000020cffff (prio 0, i/o): next.mmio > 000000000200e000-000000000200efff (prio 0, i/o): next.kbd > 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem > 0000000002100000-000000000211ffff (prio 0, i/o): next.scr > 0000000002114000-000000000211400f (prio 0, i/o): esp-regs > 0000000002118000-0000000002118003 (prio 0, i/o): escc > 0000000004000000-0000000007ffffff (prio 0, ram): next.ram > 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video > 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 > @next.bmapmem 0000000000000000-000000000000003f > > All this patch does is move the start of next.mmio to 0x2005000 to avoid the overlap. > > > I think there is another range at 0x02006000 related to the ethernet > > controller, so directly going with 0x02007000 might cause less friction > > later when we add the NIC? > > By the end of the series, everything except the "global" registers in next.mmio have > their own memory region (or empty-slot if the target is unknown) so that "info mtree" > output looks like this: > > (qemu) info mtree > address-space: cpu-memory-0 > address-space: memory > 0000000000000000-ffffffffffffffff (prio 0, i/o): system > 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom > 0000000000000000-000000000001ffff > 0000000001000000-000000000101ffff (prio 0, rom): next.rom > 0000000002000000-0000000002004fff (prio 0, i/o): next.dma > 0000000002005000-000000000200dfff (prio 0, i/o): next.mmio > 000000000200e000-000000000200efff (prio 0, i/o): next.kbd > 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem > 0000000002106000-000000000210601f (prio 0, i/o): next.en > 0000000002110000-000000000211000f (prio -10000, i/o): empty-slot > 0000000002112000-000000000211200f (prio -10000, i/o): empty-slot > 0000000002114000-000000000211403f (prio 0, i/o): next.scsi > 0000000002114000-000000000211400f (prio 0, i/o): esp-regs > 0000000002114020-0000000002114021 (prio 0, i/o): csrs > 0000000002114108-000000000211410b (prio 0, i/o): next.floppy > 0000000002118000-0000000002118003 (prio 0, i/o): escc > 0000000002118004-0000000002118013 (prio -10000, i/o): empty-slot > 000000000211a000-000000000211a003 (prio 0, i/o): next.timer > 0000000004000000-0000000007ffffff (prio 0, ram): next.ram > 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video > 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 > @next.bmapmem 0000000000000000-000000000000003f > > In this case next.en is a dummy memory region which can easily be replaced with a > proper device implementation: see the final version of next-cube.c after the series > at https://gitlab.com/mcayland/qemu/-/blob/next-cube-improvements/hw/m68k/next-cube.c. If I get it right, the IO memory is mirrored at 0x2000000 and 0x2100000, so the ethernet region should show up in both, 0x2006000 and 0x2106000 in the end? ... those memory regions on the NeXT are very confusing, but at least this is what I get from https://sourceforge.net/p/previous/code/HEAD/tree/trunk/src/cpu/memory.c#l1167 (IO_bank is mapped twice, unless it's the 030 Cube) So I think we should make sure that we can mirror the ethernet registers at 0x2006000, too? Thomas
On 27/10/2024 11:24, Thomas Huth wrote: > Am Sat, 26 Oct 2024 22:13:25 +0100 > schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: > >> On 26/10/2024 08:56, Thomas Huth wrote: >> >>> Am Wed, 23 Oct 2024 09:58:19 +0100 >>> schrieb Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk>: >>> >>>> Change the start of the next.mmio memory region so that it follows on directly >>>> after the next.dma memory region, adjusting the address offsets in >>>> next_mmio_read() and next_mmio_write() accordingly. >>>> >>>> Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> >>>> --- >>>> hw/m68k/next-cube.c | 28 ++++++++++++++-------------- >>>> 1 file changed, 14 insertions(+), 14 deletions(-) >>>> >>>> diff --git a/hw/m68k/next-cube.c b/hw/m68k/next-cube.c >>>> index 4e8e55a8bd..e1d94c1ce0 100644 >>>> --- a/hw/m68k/next-cube.c >>>> +++ b/hw/m68k/next-cube.c >>>> @@ -266,23 +266,23 @@ static uint64_t next_mmio_read(void *opaque, hwaddr addr, unsigned size) >>>> uint64_t val; >>>> >>>> switch (addr) { >>>> - case 0x7000: >>>> + case 0x2000: >>>> /* DPRINTF("Read INT status: %x\n", s->int_status); */ >>>> val = s->int_status; >>>> break; >>>> >>>> - case 0x7800: >>>> + case 0x2800: >>>> DPRINTF("MMIO Read INT mask: %x\n", s->int_mask); >>>> val = s->int_mask; >>>> break; >>>> >>>> - case 0xc000 ... 0xc003: >>>> - val = extract32(s->scr1, (4 - (addr - 0xc000) - size) << 3, >>>> + case 0x7000 ... 0x7003: >>>> + val = extract32(s->scr1, (4 - (addr - 0x7000) - size) << 3, >>>> size << 3); >>>> break; >>>> >>>> - case 0xd000 ... 0xd003: >>>> - val = extract32(s->scr2, (4 - (addr - 0xd000) - size) << 3, >>>> + case 0x8000 ... 0x8003: >>>> + val = extract32(s->scr2, (4 - (addr - 0x8000) - size) << 3, >>>> size << 3); >>>> break; >>>> >>>> @@ -301,25 +301,25 @@ static void next_mmio_write(void *opaque, hwaddr addr, uint64_t val, >>>> NeXTPC *s = NEXT_PC(opaque); >>>> >>>> switch (addr) { >>>> - case 0x7000: >>>> + case 0x2000: >>>> DPRINTF("INT Status old: %x new: %x\n", s->int_status, >>>> (unsigned int)val); >>>> s->int_status = val; >>>> break; >>>> >>>> - case 0x7800: >>>> + case 0x2800: >>>> DPRINTF("INT Mask old: %x new: %x\n", s->int_mask, (unsigned int)val); >>>> s->int_mask = val; >>>> break; >>> >>> Could you please add comments at the end of the "case" lines, stating which >>> mmio addresses are handled in each case? Otherwise, it's harder to grep for >>> certain addresses later. E.g: >>> >>> case 0x2800: /* 0x2007800 */ >> >> If you think it will help? Presumably this is to aid with comparing with other source >> code/documentation? > > Yes, it will help with 1) debugging code that is running in the guest (so > you can find IO addresses that it is accessing more easily) and 2) help > when comparing the code with "Previous": > > https://sourceforge.net/p/previous/code/HEAD/tree/trunk/src/ioMemTabNEXT.c#l36 > >>>> @@ -1000,7 +1000,7 @@ static void next_cube_init(MachineState *machine) >>>> sysbus_create_simple(TYPE_NEXTFB, 0x0B000000, NULL); >>>> >>>> /* MMIO */ >>>> - sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02000000); >>>> + sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02005000); >>> >>> Why 0x02005000 and not directly 0x02007000 ? >> >> Before this patch the output of "info mtree" shows as follows: >> >> (qemu) info mtree >> address-space: cpu-memory-0 >> address-space: memory >> 0000000000000000-ffffffffffffffff (prio 0, i/o): system >> 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom >> 0000000000000000-000000000001ffff >> 0000000001000000-000000000101ffff (prio 0, rom): next.rom >> 0000000002000000-0000000002004fff (prio 0, i/o): next.dma >> 0000000002000000-00000000020cffff (prio 0, i/o): next.mmio >> 000000000200e000-000000000200efff (prio 0, i/o): next.kbd >> 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem >> 0000000002100000-000000000211ffff (prio 0, i/o): next.scr >> 0000000002114000-000000000211400f (prio 0, i/o): esp-regs >> 0000000002118000-0000000002118003 (prio 0, i/o): escc >> 0000000004000000-0000000007ffffff (prio 0, ram): next.ram >> 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video >> 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 >> @next.bmapmem 0000000000000000-000000000000003f >> >> All this patch does is move the start of next.mmio to 0x2005000 to avoid the overlap. >> >>> I think there is another range at 0x02006000 related to the ethernet >>> controller, so directly going with 0x02007000 might cause less friction >>> later when we add the NIC? >> >> By the end of the series, everything except the "global" registers in next.mmio have >> their own memory region (or empty-slot if the target is unknown) so that "info mtree" >> output looks like this: >> >> (qemu) info mtree >> address-space: cpu-memory-0 >> address-space: memory >> 0000000000000000-ffffffffffffffff (prio 0, i/o): system >> 0000000000000000-000000000001ffff (prio 0, rom): alias next.rom2 @next.rom >> 0000000000000000-000000000001ffff >> 0000000001000000-000000000101ffff (prio 0, rom): next.rom >> 0000000002000000-0000000002004fff (prio 0, i/o): next.dma >> 0000000002005000-000000000200dfff (prio 0, i/o): next.mmio >> 000000000200e000-000000000200efff (prio 0, i/o): next.kbd >> 00000000020c0000-00000000020c003f (prio 0, ram): next.bmapmem >> 0000000002106000-000000000210601f (prio 0, i/o): next.en >> 0000000002110000-000000000211000f (prio -10000, i/o): empty-slot >> 0000000002112000-000000000211200f (prio -10000, i/o): empty-slot >> 0000000002114000-000000000211403f (prio 0, i/o): next.scsi >> 0000000002114000-000000000211400f (prio 0, i/o): esp-regs >> 0000000002114020-0000000002114021 (prio 0, i/o): csrs >> 0000000002114108-000000000211410b (prio 0, i/o): next.floppy >> 0000000002118000-0000000002118003 (prio 0, i/o): escc >> 0000000002118004-0000000002118013 (prio -10000, i/o): empty-slot >> 000000000211a000-000000000211a003 (prio 0, i/o): next.timer >> 0000000004000000-0000000007ffffff (prio 0, ram): next.ram >> 000000000b000000-000000000b1cb0ff (prio 0, ram): next-video >> 00000000820c0000-00000000820c003f (prio 0, ram): alias next.bmapmem2 >> @next.bmapmem 0000000000000000-000000000000003f >> >> In this case next.en is a dummy memory region which can easily be replaced with a >> proper device implementation: see the final version of next-cube.c after the series >> at https://gitlab.com/mcayland/qemu/-/blob/next-cube-improvements/hw/m68k/next-cube.c. > > If I get it right, the IO memory is mirrored at 0x2000000 and 0x2100000, > so the ethernet region should show up in both, 0x2006000 and 0x2106000 in > the end? ... those memory regions on the NeXT are very confusing, but at > least this is what I get from > https://sourceforge.net/p/previous/code/HEAD/tree/trunk/src/cpu/memory.c#l1167 > (IO_bank is mapped twice, unless it's the 030 Cube) That's definitely not something I've seen before, but that certainly looks like the case in Previous. I wonder whether this is something that has been measured on real hardware? > So I think we should make sure that we can mirror the ethernet registers at > 0x2006000, too? Yes, that's possible. I haven't really given a great deal of thought to how the series reflects real hardware, but I can confirm that each commit is structured such that it always boots the kernel to the same point as current master. The address decoding looks odd, but if we were to decide to do this then I'd be inclined to create a memory subregion for the next-pc device, move everything from 0x2000000-0x20fffff into it, and then create an alias from the 0x2100000-0x21fffff region back onto it. Would it be okay to leave such a change to a follow-up series? There is already a huge amount of value in this series in terms of detangling the memory regions and sorting out the IRQ wiring, and I think the final decision as to how to implement it would depend a lot upon what happens once the DMA controller has been reworked. ATB, Mark.
diff --git a/hw/m68k/next-cube.c b/hw/m68k/next-cube.c index 4e8e55a8bd..e1d94c1ce0 100644 --- a/hw/m68k/next-cube.c +++ b/hw/m68k/next-cube.c @@ -266,23 +266,23 @@ static uint64_t next_mmio_read(void *opaque, hwaddr addr, unsigned size) uint64_t val; switch (addr) { - case 0x7000: + case 0x2000: /* DPRINTF("Read INT status: %x\n", s->int_status); */ val = s->int_status; break; - case 0x7800: + case 0x2800: DPRINTF("MMIO Read INT mask: %x\n", s->int_mask); val = s->int_mask; break; - case 0xc000 ... 0xc003: - val = extract32(s->scr1, (4 - (addr - 0xc000) - size) << 3, + case 0x7000 ... 0x7003: + val = extract32(s->scr1, (4 - (addr - 0x7000) - size) << 3, size << 3); break; - case 0xd000 ... 0xd003: - val = extract32(s->scr2, (4 - (addr - 0xd000) - size) << 3, + case 0x8000 ... 0x8003: + val = extract32(s->scr2, (4 - (addr - 0x8000) - size) << 3, size << 3); break; @@ -301,25 +301,25 @@ static void next_mmio_write(void *opaque, hwaddr addr, uint64_t val, NeXTPC *s = NEXT_PC(opaque); switch (addr) { - case 0x7000: + case 0x2000: DPRINTF("INT Status old: %x new: %x\n", s->int_status, (unsigned int)val); s->int_status = val; break; - case 0x7800: + case 0x2800: DPRINTF("INT Mask old: %x new: %x\n", s->int_mask, (unsigned int)val); s->int_mask = val; break; - case 0xc000 ... 0xc003: + case 0x7000 ... 0x7003: DPRINTF("SCR1 Write: %x\n", (unsigned int)val); - s->scr1 = deposit32(s->scr1, (4 - (addr - 0xc000) - size) << 3, + s->scr1 = deposit32(s->scr1, (4 - (addr - 0x7000) - size) << 3, size << 3, val); break; - case 0xd000 ... 0xd003: - s->scr2 = deposit32(s->scr2, (4 - (addr - 0xd000) - size) << 3, + case 0x8000 ... 0x8003: + s->scr2 = deposit32(s->scr2, (4 - (addr - 0x8000) - size) << 3, size << 3, val); next_scr2_led_update(s); next_scr2_rtc_update(s); @@ -897,7 +897,7 @@ static void next_pc_realize(DeviceState *dev, Error **errp) qdev_init_gpio_in(dev, next_irq, NEXT_NUM_IRQS); memory_region_init_io(&s->mmiomem, OBJECT(s), &next_mmio_ops, s, - "next.mmio", 0xd0000); + "next.mmio", 0x9000); memory_region_init_io(&s->scrmem, OBJECT(s), &next_scr_ops, s, "next.scr", 0x20000); sysbus_init_mmio(sbd, &s->mmiomem); @@ -1000,7 +1000,7 @@ static void next_cube_init(MachineState *machine) sysbus_create_simple(TYPE_NEXTFB, 0x0B000000, NULL); /* MMIO */ - sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02000000); + sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 0, 0x02005000); /* BMAP IO - acts as a catch-all for now */ sysbus_mmio_map(SYS_BUS_DEVICE(pcdev), 1, 0x02100000);
Change the start of the next.mmio memory region so that it follows on directly after the next.dma memory region, adjusting the address offsets in next_mmio_read() and next_mmio_write() accordingly. Signed-off-by: Mark Cave-Ayland <mark.cave-ayland@ilande.co.uk> --- hw/m68k/next-cube.c | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-)