Hi,

On Thu, Apr 26, 2012 at 02:32:56PM -0600, Stephen Warren wrote:
> On 04/26/2012 02:26 PM, Felipe Balbi wrote:
> > Hi,
> > 
> > On Thu, Apr 26, 2012 at 01:55:13PM -0600, Stephen Warren wrote:
> >> On 04/25/2012 05:07 AM, Hiroshi DOYU wrote:
> >>> Add extern func, "tegra_ahb_enable_smmu()" to inform AHB that SMMU is
> >>> ready.
> >>
> >>> +#ifdef CONFIG_ARCH_TEGRA_3x_SOC
> >>> +static int __tegra_ahb_enable_smmu(struct device *dev, void *data)
> >> ...
> >>> +int tegra_ahb_enable_smmu(void)
> >>> +{
> >>> +	return driver_for_each_device(&tegra_ahb_driver.driver, NULL, NULL,
> >>> +				      __tegra_ahb_enable_smmu);
> >>> +}
> >>> +EXPORT_SYMBOL(tegra_ahb_enable_smmu);
> >>> +#endif
> >>
> >> That looks like a neat solution to avoid having a global device object.
> > 
> > except that it won't work always. If you happen to have two AHB bridges,
> > each using a separate smmu but only one smmu is ready, this will set
> > SMMU_INIT_DONE on both bridges.
> 
> There is only 1.

that's why there's a "if you happen to have" statement. If you stick to
this "there is only 1" argument, why do you even make this into a
platform driver ? Just stick the entire code hidden on the
machine_init() code. Drivers a supposed to be able to instantiated
multiple times and always work, this method won't work if tegra99999
ends up with two AHB bridges/SMMUs

> >> However, if that driver_for_each_device finds no devices, the function
> >> still succeeds. That doesn't seem right, and doesn't allow e.g. the SMMU
> >> to defer its probe until the AHB driver has completed.
> >>
> >> Perhaps add a local int variable to tegra_ahb_enable_smmu(), pass the
> >> address to __tegra_ahb_enable_smmu, and have it increment the int. Then,
> >> after calling driver_for_each_device,:
> >>
> >> if (!ahb_device_count)
> >>     return -EPROBE_DEFER
> >> if (WARN_ON(ahb_device_count != 1))
> >>     return -EINVAL;
> >> return 0;
> > 
> > that would look, well, weird. Why don't you just different initcall
> > leves for this ? Maybe smmu goes into postcore_initcall() and tegra_ahb
> > goes into postcore_initcall_sync() ?? then you know that SMMU will be
> > ready by the time you call tegra_ahb probe. Well, unless smmu's probe
> > fail, but then again, IIUC it won't work anyway...
> 
> Uggh. I'd rather all these devices just got instantiated from device
> tree and relied on deferred probe to manage any ordering, rather than
> playing complex games with multiple initcall levels (and in the end
> probably having to invent more and more initcall levels to correctly
> represent all the dependencies).

then do that... it'll be better than current trickery with
driver_for_each_device() and my initcall trickery ;-)

-- 
balbi