Re: [PATCH v3] sh: avoid using IRQ0 on SH3/4

From: Sergey Shtylyov
Date: Tue May 03 2022 - 15:33:31 EST


On 5/2/22 11:56 PM, Sergey Shtylyov wrote:

[...]
>>>>> Using IRQ0 by the platform devices is going to be disallowed soon (see [1])
>>>>> and even now, when IRQ0 is about to be returned by platfrom_get_irq(), you
>>>>> see a big warning. The code supporting SH3/4 SoCs maps the IRQ #s starting
>>>>> at 0 -- modify that code to start the IRQ #s from 16 instead.
>>>>>
>>>>> The patch should mostly affect the AP-SH4A-3A/AP-SH4AD-0A boards as they
>>>>> indeed use IRQ0 for the SMSC911x compatible Ethernet chip...
>>
>>> As I told him in IRC, the problem is still that sh4 never gives me a shell
>>> prompt with this patch applied. I just reconfirmed it against current git:
>>>
>>> Freeing unused kernel image (initmem) memory: 124K
>>> This architecture does not have kernel memory protection.
>>> Run /init as init process
>>> mountpoint: dev/pts: No such file or directory
>>> 8139cp 0000:00:02.0 eth0: link up, 100Mbps, full-duplex, lpa 0x05E1
>>>
>>> It makes it partway through the init script, but it hangs with qemu-system-sh4
>>> stuck in a CPU-eating loop before finishing. Without the patch, I get a shell
>>> prompt.
>>
>> I regularly test on qemu rts7751r2d, but couldn't produce your
>> issue. Until I tried "ifconfig eth0 up", which causes a lock-up.
>> Interestingly, the 8139 irq was 112 with and without Sergey's patch,
>> so there must be an irq remapping missing.
>>
>> I also test regularly on landisk, where 8139 Ethernet works fine.
>> Turns out landisk uses arch/sh/drivers/pci/fixups-landisk.c to fixup
>> the irq...
>>
>> arch/sh/include/mach-common/mach/r2d.h has:
>> #define R2D_FPGA_IRQ_BASE 100
>> Subtracting 16 here does not help.
>
> Why subtract when you contrariwise need to add? :-)
>
>> With this (gmail-whitespace-damaged) patch:
>>
>> --- a/arch/sh/drivers/pci/fixups-rts7751r2d.c
>> +++ b/arch/sh/drivers/pci/fixups-rts7751r2d.c
>> @@ -31,9 +31,9 @@ static char lboxre2_irq_tab[] = {
>> int pcibios_map_platform_irq(const struct pci_dev *pdev, u8 slot, u8 pin)
>> {
>> if (mach_is_lboxre2())
>> - return lboxre2_irq_tab[slot];
>> + return lboxre2_irq_tab[slot] - 16;
>
> This table contains the values #define'd via evt2irq(), so
> shouldn't need to subtract anything...
>
>> else
>> - return rts7751r2d_irq_tab[slot];
>> + return rts7751r2d_irq_tab[slot] - 16;
>
> How about + 16?
>
>> }
>>
>> int pci_fixup_pcic(struct pci_channel *chan)
>>
>> it no longer crashes, but ifconfig still fails:
>>
>> / # ifconfig eth0 up
>> ifconfig: ioctl 0x8914 failed: Invalid argument
>
> I'm still not sure you used the correct IRQ #s...
>
>> Note that there are more implementations of pcibios_map_platform_irq()
>> that do not use evt2irq(), and thus are probably broken by this patch.
>
> That doesn't sound encouraging... :-/

Actually, of those only Dreamcast didn't use evt2irq()...
Now I'm wondering whether all #define *_IRQ_BASE should be visited as well...

>> Gr{oetje,eeting}s,
>>
>> Geert

MBR, Sergey