On 22/02/2024 01:42, Sam Protsenko wrote: > On Tue, Feb 20, 2024 at 5:04 AM Krzysztof Kozlowski > <krzysztof.kozlowski@xxxxxxxxxx> wrote: >> >> On 16/02/2024 23:32, Sam Protsenko wrote: >>> Abstract CPU clock registers by keeping their offsets in a dedicated >>> chip specific structure to accommodate for oncoming Exynos850 support, >>> which has different offsets for cluster 0 and cluster 1. This rework >>> also makes it possible to use exynos_set_safe_div() for all chips, so >>> exynos5433_set_safe_div() is removed here to reduce the code >>> duplication. >>> >> >> So that's the answer why you could not use flags anymore - you need an >> enum, not a bitmap. Such short explanation should be in previous commits >> justifying moving reg layout to new property. > > Will do, thanks. > >> >>> No functional change. >>> >>> Signed-off-by: Sam Protsenko <semen.protsenko@xxxxxxxxxx> >>> --- >>> drivers/clk/samsung/clk-cpu.c | 156 +++++++++++++++++++--------------- >>> 1 file changed, 86 insertions(+), 70 deletions(-) >>> >>> diff --git a/drivers/clk/samsung/clk-cpu.c b/drivers/clk/samsung/clk-cpu.c >>> index 04394d2166c9..744b609c222d 100644 >>> --- a/drivers/clk/samsung/clk-cpu.c >>> +++ b/drivers/clk/samsung/clk-cpu.c >>> @@ -44,12 +44,14 @@ typedef int (*exynos_rate_change_fn_t)(struct clk_notifier_data *ndata, >>> >>> /** >>> * struct exynos_cpuclk_chip - Chip specific data for CPU clock >>> + * @regs: register offsets for CPU related clocks >>> * @pre_rate_cb: callback to run before CPU clock rate change >>> * @post_rate_cb: callback to run after CPU clock rate change >>> */ >>> struct exynos_cpuclk_chip { >>> - exynos_rate_change_fn_t pre_rate_cb; >>> - exynos_rate_change_fn_t post_rate_cb; >>> + const void * const regs; >> >> Why this is void? >> > > Different chips can have very different register layout. For example, > older Exynos chips usually keep multiple CPU divider ratios in one > single register, whereas more modern chips have a dedicated register > for each divider clock. Also, old chips usually split divider ratio vs > DIV clock status between different registers, but in modern chips they > both live in one single register. Having (void *) makes it possible to > keep pointers to different structures, and each function for the > particular chip can "know" which exactly structure is stored there, > casting (void *) to a needed type. Another way to do that would be to > have "one-size-fits-all" structure with all possible registers for all > possible chips. I don't know, I just didn't like that for a couple of > reasons, so decided to go with (void *). > > I'll add some explanation in the commit message in v2. Currently the one-size-fits-all seems feasible, even if few fields are not matching, so I would prefer to go this approach. Best regards, Krzysztof