Re: [PATCH RESEND 05/12] sh: DeviceTree support update

[Rich Felker]

On Tue, May 10, 2016 at 05:25:36PM +0900, Yoshinori Sato wrote:
> On Wed, 04 May 2016 12:10:05 +0900,
> Rich Felker wrote:
> > 
> > On Sun, May 01, 2016 at 02:08:29PM +0900, Yoshinori Sato wrote:
> > > Changes bellow
> > > - FDT setup timing fix.
> > > - chosen/bootargs support.
> > > - zImage support.
> > > - DT binding helper macro.
> > > 
> > > Signed-off-by: Yoshinori Sato <ysato@xxxxxxxxxxxxxxxxxxxx>
> > > ---
> > >  arch/sh/boards/of-generic.c                        | 23 +++++++++++-----------
> > >  arch/sh/boot/compressed/head_32.S                  |  5 +++--
> > >  arch/sh/boot/dts/include/dt-bindings               |  1 +
> > >  arch/sh/kernel/setup.c                             | 19 ++++++++++++++++++
> > >  include/dt-bindings/interrupt-controller/sh_intc.h |  2 ++
> > >  5 files changed, 36 insertions(+), 14 deletions(-)
> > >  create mode 120000 arch/sh/boot/dts/include/dt-bindings
> > >  create mode 100644 include/dt-bindings/interrupt-controller/sh_intc.h
> > > 
> > > diff --git a/arch/sh/boards/of-generic.c b/arch/sh/boards/of-generic.c
> > > index bf3a166..9570873 100644
> > > --- a/arch/sh/boards/of-generic.c
> > > +++ b/arch/sh/boards/of-generic.c
> > > @@ -112,29 +112,25 @@ static int noopi(void)
> > >  	return 0;
> > >  }
> > >  
> > > -static void __init sh_of_mem_reserve(void)
> > > +static void __init sh_of_mem_init(void)
> > >  {
> > >  	early_init_fdt_reserve_self();
> > >  	early_init_fdt_scan_reserved_mem();
> > >  }
> > >  
> > > -static void __init sh_of_time_init(void)
> > > -{
> > > -	pr_info("SH generic board support: scanning for clocksource devices\n");
> > > -	clocksource_probe();
> > > -}
> > 
> > Why did you remove this? Without it you won't get clock
> > event/clocksource devices from the device tree so the only way to have
> > a working timer interrupt is if the driver is hard-coded somewhere.
> 
> It not needed on Common Clock Framework.
> tmu define in dts.

It is needed. clocksources are something completely different from
"clk"s. A clocksource is the modern source of time data for the kernel
timekeeping system (without one, you're stuck using jiffies and very
low-res time), and the probe also gets clock_event_devices which are
the source of timer interrupts. Without this, unless you have a
hard-coded source of timer interrupt for the board, you won't get a
timer interrupt and the kernel will hang early in the boot process.

> > >  static void __init sh_of_setup(char **cmdline_p)
> > >  {
> > > -	unflatten_device_tree();
> > > -
> > > -	board_time_init = sh_of_time_init;
> > > +	struct device_node *cpu;
> > > +	int freq;
> > >  
> > >  	sh_mv.mv_name = of_flat_dt_get_machine_name();
> > >  	if (!sh_mv.mv_name)
> > >  		sh_mv.mv_name = "Unknown SH model";
> > >  
> > >  	sh_of_smp_probe();
> > > +	cpu = of_find_node_by_name(NULL, "cpu");
> > > +	if (!of_property_read_u32(cpu, "clock-frequency", &freq))
> > > +		preset_lpj = freq / 500;
> > >  }
> > 
> > I setup the DT-based pseudo-board to use the generic calibrate-delay
> > rather than hard-coding lpj. Ideally we could just get rid of bogomips
> > completely but there are probably still some things using it. Is there
> > a reason you prefer making up a value for lpj based on the cpu clock
> > rate?
> 
> clockevent initalize after calibrate delay.
> So don't work interrupt based calibrate.

Currently, it initializes before, but you removed the probe that
initializes it (above), clocksource_probe().

> > >  static int sh_of_irq_demux(int irq)
> > > @@ -167,8 +163,7 @@ static struct sh_machine_vector __initmv sh_of_generic_mv = {
> > >  	.mv_init_irq	= sh_of_init_irq,
> > >  	.mv_clk_init	= sh_of_clk_init,
> > >  	.mv_mode_pins	= noopi,
> > > -	.mv_mem_init	= noop,
> > > -	.mv_mem_reserve	= sh_of_mem_reserve,
> > > +	.mv_mem_init	= sh_of_mem_init,
> > 
> > Is there a reason for this renaming? The function seems to be dealing
> > purely with reserving memory ranges.
> 
> mv_mem_reserve too late call in MMU system.

OK.

> > >  	if (!dt_virt || !early_init_dt_scan(dt_virt)) {
> > >  		pr_crit("Error: invalid device tree blob"
> > > @@ -267,8 +276,13 @@ void __ref sh_fdt_init(phys_addr_t dt_phys)
> > >  
> > >  void __init setup_arch(char **cmdline_p)
> > >  {
> > > +#ifdef CONFIG_OF
> > > +	unflatten_device_tree();
> > > +#endif
> > >  	enable_mmu();
> > 
> > Was this moved to setup_arch to have it before enable_mmu? I think
> > that makes sense.
> 
> early_init_dt_alloc_memory_arch used physical address.
> It override on sh-specific, can after enable_mmu.
> But I don't feel an advantage.

I dont think we should be putting sh-specific code in
early_init_dt_alloc_memory. If calling unflatten_device_tree before
enable_mmu avoids the need to do that, it seems like the right choice.
Is this how other archs do it? I think we should try to be as
consistent as possible with general practice across the kernel.

> > > +#ifndef CONFIG_OF
> > >  	ROOT_DEV = old_decode_dev(ORIG_ROOT_DEV);
> > >  
> > >  	printk(KERN_NOTICE "Boot params:\n"
> > > @@ -290,6 +304,7 @@ void __init setup_arch(char **cmdline_p)
> > >  
> > >  	if (!MOUNT_ROOT_RDONLY)
> > >  		root_mountflags &= ~MS_RDONLY;
> > > +#endif
> > 
> > Do these boot params only make sense for non-DT setups?
> > 
> > > +#if !defined(CONFIG_OF) || defined(USE_BUILTIN_DTB)
> > >  	/* Save unparsed command line copy for /proc/cmdline */
> > >  	memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
> > >  	*cmdline_p = command_line;
> > > +#else
> > > +	*cmdline_p = boot_command_line;
> > > +#endif
> > 
> > I think this is wrong -- it causes the builtin command line and
> > bootloader-provided command line to be ignored on DT kernels. Do you
> > just want to deprecate builtin and bootloader-provided command lines?
> > Or is it just a side effect of adding support for chosen/bootarg?
> 
> Yes. I think zero-page passing only non-DT.
> DT using chosen/bootargs.

This precludes having the bootloader provide a DTB from ROM, since it
needs to be able to patch the DTB with user preferences, and also
requires DTB-patching code to be added to the bootloader. At the very
least we should retain the ability to have a builtin command line in
the kernel that's appended to the one from the DTB (and possibly allow
the order to be changed), but I think we should also allow the
bootloader to pass in a command line like in the non-DTB setup.

DTB and command line are very different things (hardware description
vs user preference, and OS-generic vs OS-specific) and the whole
chosen/bootargs system is something of a hack.

> > >  	parse_early_param();
> > >  
> > > diff --git a/include/dt-bindings/interrupt-controller/sh_intc.h b/include/dt-bindings/interrupt-controller/sh_intc.h
> > > new file mode 100644
> > > index 0000000..8c9dcdc
> > > --- /dev/null
> > > +++ b/include/dt-bindings/interrupt-controller/sh_intc.h
> > > @@ -0,0 +1,2 @@
> > > +#define evt2irq(evt)		(((evt) >> 5) - 16)
> > > +#define irq2evt(irq)		(((irq) + 16) << 5)
> > 
> > This seems unrelated to other things in this patch.
> 
> It using DT define.

Ah, I see, it's used later in the dts itself. But why not just put the
actual IRQ numbers in the dts directly? The evt numbers seem like an
implementation detail, whereas device tree bindings should be stable
and independent of kernel sources.

If evt numbers are really what makes sense to have in the dts, then
sh_intc should probably define its irq_domain so that it maps evt
numbers to irq numbers. But that doesn't seem necessary.

Rich