Re: [PATCH] clk: mediatek: mt8173-mfgtop: do not use the clk API from power_on

[Ryan Brue <ryanbrue.dev@gmail.com>]

On 9/12/26 1:44 PM, sashiko-bot@kernel.org wrote:
> [Severity: High]
> Since this raw regmap write in clk_mt8173_mfgtop_power_on() in
> drivers/clk/mediatek/clk-mt8173-mfgtop.c directly asserts the gate in
> hardware, bypassing the Common Clock Framework (CCF) state tracking, the
> clock's enable_count remains 0.
>
> Because the mfg_26m clock is still fully registered via mfg_clks without the
> CLK_IGNORE_UNUSED flag, won't clk_disable_unused() forcibly disable this
> clock during late init?
>
> If a consumer has probed and powered on the domain before or during
> late_initcall_sync, clk_disable_unused() will check the hardware state, see
> that the bit is active while enable_count is 0, and turn it off by writing
> to set_ofs.
>
> This abruptly shuts off the internal power management clock while the genpd
> is still active, potentially leading to hardware hangs and failure of
> subsequent power-off operations.
Yep, that's a real gap. I'll add CLK_IGNORE_UNUSED to the mfg_26m gate. 
Thanks for catching this!
> [Severity: High]
> This raw regmap write in clk_mt8173_mfgtop_power_off() deasserts the gate
> directly, overriding any state assumed by CCF, which is part of the same
> framework bypass issue described above.
This shares the same fix as the first one; once CCF stops touching the 
gate, the raw set/clr writes from the callbacks are the only writers. 
I'll send out a v2.