Re: [PATCH 2/2] speakup: only build serialio when ISA is enabled

Re: [PATCH 2/2] speakup: only build serialio when ISA is enabled

[Samuel Thibault]

Greg KH, le mar. 04 août 2020 13:58:17 +0200, a ecrit:
> 	ERROR: modpost: "spk_serial_io_ops" [drivers/accessibility/speakup/speakup_keypc.ko] undefined!
> 	ERROR: modpost: "spk_stop_serial_interrupt" [drivers/accessibility/speakup/speakup_keypc.ko] undefined!

Oh, indeed, these are not covered by the first patch.

I realize that KEYPC depends on ISA, but with COMPILE_TEST as
alternative. We can build serialio in the COMPILE_TEST case too, I'll
update the patches.

But then we'll still have:

> the riscv build issues.

Actually I was surprised by the riscv build issue: the issue is within
riscv's inb() implementation, serialio.c is only calling it.
arch/riscv/include/asm/io.h says:

#define PCI_IOBASE		((void __iomem *)PCI_IO_START)
[...]
#define inb(c)		({ u8  __v; __io_pbr(); __v = readb_cpu((void*)(PCI_IOBASE + (c))); __io_par(__v); __v; })

and thus yes it's arithmetic over a (void*) pointer, the caller cannot
do anything about it.

8250_port.c itself uses inb(), doesn't it get a warning as well? Or is
it getting compiled-out on riscv because of some Kconfig condition?  I
see that the whole drivers/tty/serial is under HAS_IOMEM, and that's the
only condition I can see for 8250_port.c (except SERIAL_8250 of course),
is that it, or is SERIAL_8250 just not enabled in the riscv bot?

Actually the warning seems new, when looking at the Debian build log:

https://buildd.debian.org/status/fetch.php?pkg=linux&arch=riscv64&ver=5.7.10-1&stamp=1595803499&raw=0

and looking for serialio.c, I do not see a warning, and its code hasn't
changed. Is the build bot compiler just more talkative?

Samuel

Re: [PATCH 2/2] speakup: only build serialio when ISA is enabled

[Greg KH]

On Tue, Aug 04, 2020 at 05:35:42PM +0200, Samuel Thibault wrote:
> Greg KH, le mar. 04 août 2020 13:58:17 +0200, a ecrit:
> > 	ERROR: modpost: "spk_serial_io_ops" [drivers/accessibility/speakup/speakup_keypc.ko] undefined!
> > 	ERROR: modpost: "spk_stop_serial_interrupt" [drivers/accessibility/speakup/speakup_keypc.ko] undefined!
> 
> Oh, indeed, these are not covered by the first patch.
> 
> I realize that KEYPC depends on ISA, but with COMPILE_TEST as
> alternative. We can build serialio in the COMPILE_TEST case too, I'll
> update the patches.
> 
> But then we'll still have:
> 
> > the riscv build issues.
> 
> Actually I was surprised by the riscv build issue: the issue is within
> riscv's inb() implementation, serialio.c is only calling it.
> arch/riscv/include/asm/io.h says:
> 
> #define PCI_IOBASE		((void __iomem *)PCI_IO_START)
> [...]
> #define inb(c)		({ u8  __v; __io_pbr(); __v = readb_cpu((void*)(PCI_IOBASE + (c))); __io_par(__v); __v; })
> 
> and thus yes it's arithmetic over a (void*) pointer, the caller cannot
> do anything about it.

And that's fine, math with pointers, even void ones, is ok.

I wonder why riscv was complaining about that.  It's not nice, but it is
valid C.

> 8250_port.c itself uses inb(), doesn't it get a warning as well? Or is
> it getting compiled-out on riscv because of some Kconfig condition?

Probably this.

> I
> see that the whole drivers/tty/serial is under HAS_IOMEM, and that's the
> only condition I can see for 8250_port.c (except SERIAL_8250 of course),
> is that it, or is SERIAL_8250 just not enabled in the riscv bot?
> 
> Actually the warning seems new, when looking at the Debian build log:
> 
> https://buildd.debian.org/status/fetch.php?pkg=linux&arch=riscv64&ver=5.7.10-1&stamp=1595803499&raw=0
> 
> and looking for serialio.c, I do not see a warning, and its code hasn't
> changed. Is the build bot compiler just more talkative?

It's more talkative at times, and maybe the riscv code changed too.

Let's see if any of their developers care about it :)

thanks,

greg k-h

Re: [PATCH 2/2] speakup: only build serialio when ISA is enabled

[Samuel Thibault]

Greg KH, le mar. 04 août 2020 17:49:52 +0200, a ecrit:
> > #define inb(c)		({ u8  __v; __io_pbr(); __v = readb_cpu((void*)(PCI_IOBASE + (c))); __io_par(__v); __v; })
> > 
> > and thus yes it's arithmetic over a (void*) pointer, the caller cannot
> > do anything about it.
> 
> And that's fine, math with pointers, even void ones, is ok.

C++ doesn't like it, but in general compilers provide some semantic for
it, yes.

> I wonder why riscv was complaining about that.  It's not nice, but it is
> valid C.

Strictly C speaking, no, but GNU C yes.

> Let's see if any of their developers care about it :)

Ok :) In the meanwhile I sent a v2 series that should fix the different
link cases, just not the particular riscv warning.

Samuel