Hi Claudiu,
On Fri, 23 May 2025 at 09:41, Claudiu Beznea <claudiu.beznea@xxxxxxxxx> wrote:
> On 22.05.2025 17:46, Geert Uytterhoeven wrote:
> > On Wed, 14 May 2025 at 11:04, Claudiu <claudiu.beznea@xxxxxxxxx> wrote:
> >> From: Claudiu Beznea <claudiu.beznea.uj@xxxxxxxxxxxxxx>
> >>
> >> The RZ/{G2L, V2L, G3S} CPG versions support a feature called MSTOP. Each
> >> module has one or more MSTOP bits associated with it, and these bits need
> >> to be configured along with the module clocks. Setting the MSTOP bits
> >> switches the module between normal and standby states.
> >>
> >> Previously, MSTOP support was abstracted through power domains
> >> (struct generic_pm_domain::{power_on, power_off} APIs). With this
> >> abstraction, the order of setting the MSTOP and CLKON bits was as follows:
> >>
> >> Previous Order:
> >> A/ Switching to Normal State (e.g., during probe):
> >> 1/ Clear module MSTOP bit
> >> 2/ Set module CLKON bit
> >>
> >> B/ Switching to Standby State (e.g., during remove):
> >> 1/ Clear CLKON bit
> >> 2/ Set MSTOP bit
> >>
> >> However, in some cases (when the clock is disabled through devres), the
> >> order may have been (due to the issue described in link section):
> >>
> >> 1/ Set MSTOP bit
> >> 2/ Clear CLKON bit
> >>
> >> Recently, the hardware team has suggested that the correct order to set
> >> the MSTOP and CLKON bits is:
> >>
> >> Updated Order:
> >> A/ Switching to Normal State (e.g., during probe):
> >> 1/ Set CLKON bit
> >> 2/ Clear MSTOP bit
> >>
> >> B/ Switching to Standby State (e.g., during remove):
> >> 1/ Set MSTOP bit
> >> 2/ Clear CLKON bit
> >>
> >> To prevent future issues due to incorrect ordering, the MSTOP setup has
> >> now been implemented in rzg2l_mod_clock_endisable(), ensuring compliance
> >> with the sequence suggested in Figure 41.5: Module Standby Mode Procedure
> >> from the RZ/G3S HW manual, Rev1.10.
> >>
> >> Additionally, since multiple clocks of a single module may be mapped to a
> >> single MSTOP bit, MSTOP setup is reference-counted.
> >>
> >> Furthermore, as all modules start in the normal state after reset, if the
> >> module clocks are disabled, the module state is switched to standby. This
> >> prevents keeping the module in an invalid state, as recommended by the
> >> hardware team.
> >>
> >> Link: https://lore.kernel.org/all/20250215130849.227812-1-claudiu.beznea.uj@xxxxxxxxxxxxxx/
> >> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@xxxxxxxxxxxxxx>
> >> ---
> >>
> >> Changes in v2:
> >> - udpated patch description to avoid plural in the configuration
> >> sequence description b/w MSTOP and CLK_ON
> >> - use atomic type to store the usage counter; s/refcnt/usecnt/g
> >> - moved MSTOP_OFF(), MSTOP_MASK() macros to rzg2l-cpg.c
> >> - dropped struct mstp_clock::critical and use clk_hw_get_flags()
> >> instead to get the clock flags
> >> - used unsigned int iterators in for loops
> >> - keep memory allocated for a single list for clocks sharing the
> >> same MSTOP by updating the rzg2l_mod_clock_add_shared_mstop_clk();
> >> - s/rzg2l_cpg_mstop_show/rzg2l_mod_clock_mstop_show/g,
> >> s/rzg2l_cpg_mstop/rzg2l_mod_clock_mstop/g,
> >> s/rzg2l_cpg_update_shared_mstop_clocks/rzg2l_mod_clock_update_shared_mstop_clks/g
> >> to keep the same naming conventions for functions handling mod clock MSTOP
> >> - use the newly added for_each_mstp_clk() macro all over the code
> >
> > Thanks for the update!
> >
> >> --- a/drivers/clk/renesas/rzg2l-cpg.c
> >> +++ b/drivers/clk/renesas/rzg2l-cpg.c
> >
> >> @@ -1209,6 +1232,94 @@ struct mstp_clock {
> >> else if (((hw) = __clk_get_hw((priv)->clks[(priv)->num_core_clks + i])) && \
> >> ((mstp_clk) = to_mod_clock(hw)))
> >>
> >> +/* Need to be called with a lock held to avoid concurrent access to mstop->usecnt. */
> >> +static void rzg2l_mod_clock_module_set_state(struct mstp_clock *clock,
> >> + bool standby)
> >> +{
> >> + struct rzg2l_cpg_priv *priv = clock->priv;
> >> + struct mstop *mstop = clock->mstop;
> >> + bool update = false;
> >> + u32 value;
> >> +
> >> + if (!mstop)
> >> + return;
> >> +
> >> + value = MSTOP_MASK(mstop->conf) << 16;
> >> +
> >> + if (standby) {
> >> + unsigned int criticals = 0;
> >> +
> >> + for (unsigned int i = 0; i < clock->num_shared_mstop_clks; i++) {
> >> + struct mstp_clock *clk = clock->shared_mstop_clks[i];
> >> +
> >> + if (clk_hw_get_flags(&clk->hw) & CLK_IS_CRITICAL)
> >> + criticals++;
> >> + }
> >> +
> >> + /*
> >> + * If this is a shared MSTOP and it is shared with critical clocks,
> >> + * and the system boots up with this clock enabled but no driver
> >> + * uses it the CCF will disable it (as it is unused). As we don't
> >> + * increment reference counter for it at registration (to avoid
> >> + * messing with clocks enabled at probe but later used by drivers)
> >> + * do not set the MSTOP here too if it is shared with critical
> >> + * clocks and ref counted only by those critical clocks.
> >> + */
> >> + if (criticals && criticals == atomic_read(&mstop->usecnt))
> >> + return;
> >> +
> >> + value |= MSTOP_MASK(mstop->conf);
> >> +
> >> + /* Allow updates on probe when usecnt = 0. */
> >> + if (!atomic_read(&mstop->usecnt))
> >> + update = true;
> >> + else
> >> + update = atomic_dec_and_test(&mstop->usecnt);
> >> + } else {
> >> + atomic_inc(&mstop->usecnt);
> >> + update = true;
> >
> > Shouldn't the update be conditional, i.e.:
> >
> > if (!atomic_read(&mstop->usecnt))
> > update = true;
> > atomic_inc(&mstop->usecnt);
> >
> > ?
>
> Indeed, it should be conditional as you suggested.
>
> >
> >> + }
> >> +
> >> + if (update)
> >> + writel(value, priv->base + MSTOP_OFF(mstop->conf));
> >> +}
> >
> >> +static int rzg2l_mod_clock_update_shared_mstop_clks(struct rzg2l_cpg_priv *priv,
> >> + struct mstp_clock *clock)
> >> +{
> >> + struct mstp_clock *clk;
> >> + struct clk_hw *hw;
> >> +
> >> + if (!clock->mstop)
> >> + return 0;
> >> +
> >> + for_each_mstp_clk(clk, hw, priv) {
> >> + struct mstp_clock **new_clks;
> >> + int num_shared_mstop_clks;
> >> + bool found = false;
> >> +
> >> + if (clk->mstop != clock->mstop)
> >> + continue;
> >> +
> >> + num_shared_mstop_clks = clk->num_shared_mstop_clks;
> >> + for (unsigned int i = 0; i < num_shared_mstop_clks; i++) {
> >> + if (clk->shared_mstop_clks[i] == clock) {
> >> + found = true;
> >> + break;
> >> + }
> >> + }
> >> + if (found)
> >> + continue;
> >
> > Can this happen? With your current code, the answer is yes.
> > But I think this loop and check can be removed...
> >
> >> +
> >> + if (!num_shared_mstop_clks)
> >> + new_clks = devm_kmalloc_array(priv->dev, 2, sizeof(*new_clks), GFP_KERNEL);
> >> + else
> >> + new_clks = devm_krealloc(priv->dev, clk->shared_mstop_clks,
> >> + (num_shared_mstop_clks + 1) * sizeof(*new_clks),
> >> + GFP_KERNEL);
> >> +
> >> + if (!new_clks)
> >> + return -ENOMEM;
> >> +
> >> + if (!num_shared_mstop_clks)
> >> + new_clks[num_shared_mstop_clks++] = clk;
> >> + if (clk != clock)
> >
> > This check is always true
>
> If I'm not wrong now, when adding the clock to it's own list, and the list
> is empty (!num_shared_mstop_clks check above is true), if this condition is
> missing the clock it will be added twice in its own list.
Sorry, I missed that this function is called _after_ the clock is
added to priv->clks[]. So one question and comment here:
1. Do you need a one-entry array (actual allocation is two entries)
for module clocks with an mstop entry that is not shared?
Perhaps for critical clocks? That could be handled in
rzg2l_mod_clock_module_set_state(), by explicitly checking
the clock's own critical flag if num_shared_mstop_clks is zero.
2. If rzg2l_mod_clock_update_shared_mstop_clks() would be called
_before_ the clock is added to priv->clks[], the clk != clock
check would not be needed.
> >> + new_clks[num_shared_mstop_clks++] = clock;
> >> +
> >> + for (unsigned int i = 0; i < num_shared_mstop_clks; i++) {
> >> + new_clks[i]->shared_mstop_clks = new_clks;
> >> + new_clks[i]->num_shared_mstop_clks = num_shared_mstop_clks;
> >> + }
> >
> > ... by adding a "break" here. The loop above has already updated the
> > shared_mstop_clks[] arrays for all clocks sharing the same mstop value.
>
> It may happen that the entries in the module clock array provided by the
> SoC specific drivers to not be sorted by module clock ID. That's the case
> with RZ/G2L IA55 clocks (from r9a07g044-cpg.c):
>
> static const struct {
> struct rzg2l_mod_clk common[79];
> #ifdef CONFIG_CLK_R9A07G054
> struct rzg2l_mod_clk drp[5];
> #endif
> } mod_clks = {
> .common = {
> // ...
>
> DEF_MOD("ia55_pclk", R9A07G044_IA55_PCLK, R9A07G044_CLK_P2,
> 0x518, 0, MSTOP(BUS_PERI_CPU, BIT(13))),
> DEF_MOD("ia55_clk", R9A07G044_IA55_CLK, R9A07G044_CLK_P1,
> 0x518, 1, MSTOP(BUS_PERI_CPU, BIT(13))),
>
> // ...
> };
>
> Where IDs are defined as:
>
> #define R9A07G044_IA55_CLK 8
> #define R9A07G044_IA55_PCLK 9
>
> These clocks share the same MSTOP bit.
>
> Because the ia55_pclk is the 1st clock registered (index 9) it will be
> added to priv->clks[base + 9].
>
> Next registered clock will be for ia55_clk, with index 8, it will be added
> to priv->clks[base + 8].
>
> for_each_mstp_clk() loops on clocks from priv->clks[] array. If a break
> will be done at the end of the for_each_mstp_clk() loop, at the end of the
> registration each of these clocks will have on it's shared_mstop_clks[]
> only references to itself.
If rzg2l_mod_clock_update_shared_mstop_clks() would be called _before_
the clock is added to priv->clks[], this issue could not happen, right?
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
