Quoting Théo Lebrun (2024-10-23 04:08:31) > On Thu Oct 17, 2024 at 8:48 PM CEST, Stephen Boyd wrote: > > Quoting Théo Lebrun (2024-10-07 06:49:19) > > > +/* Required early for UART. */ > > > > I still don't get this. UART isn't an early device. It's only the > > interrupt controller and the timer that matter. Does MIPS do something > > special for UARTs? > > Our hardware has a PL011. That is AMBA stuff; they get probed before > platform devices by of_platform_bus_create(). "pll-per" on EyeQ5 must > be available at that time. > > In concrete terms, if we don't register pll-per on EyeQ5 at > of_clk_init(), we stare at void because the serial fails probing. > I haven't digged into why EPROBE_DEFER doesn't do its job. Anyway we > don't want our serial to stall for some time during our boot process. > Ok thanks for the details. It sounds like there's a bug in there somewhere. Eventually this should be removed. Can you dump_stack() in clk_get() when the "pll-per" clk is claimed? I suspect of_clk_get_hw_from_clkspec() is seeing NULL if of_clk_hw_onecell_get() is being used and the clk_hw pointer isn't set yet. NULL is a valid clk and it will be returned to the consumer. You'll want to write a custom 'get' function for of_clk_add_hw_provider() that returns -EPROBE_DEFER for any clk that isn't registered early. Then the AMBA stuff should defer probe until the "full" clk provider is registered.
