Message-ID: <CA+Ln22E4FPexE1R2dmV=u9U5UFWsAz=8kXgqBntEBgabnUEF+Q@mail.gmail.com>
Date:   Tue, 11 Aug 2020 18:53:44 +0200
From:   Tomasz Figa <tomasz.figa@...il.com>
To:     Sylwester Nawrocki <s.nawrocki@...sung.com>
Cc:     "open list:COMMON CLK FRAMEWORK" <linux-clk@...r.kernel.org>,
        Chanwoo Choi <cw00.choi@...sung.com>,
        Stephen Boyd <sboyd@...nel.org>,
        Mike Turquette <mturquette@...libre.com>,
        "moderated list:SAMSUNG SOC CLOCK DRIVERS" 
        <linux-samsung-soc@...r.kernel.org>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        Bartlomiej Zolnierkiewicz <b.zolnierkie@...sung.com>,
        Marek Szyprowski <m.szyprowski@...sung.com>
Subject: Re: [PATCH v2] clk: samsung: Prevent potential endless loop in the
 PLL set_rate ops

2020年8月11日(火) 18:45 Sylwester Nawrocki <s.nawrocki@...sung.com>:
>
> Hi Tomasz,
>
> On 11.08.2020 14:59, Tomasz Figa wrote:
> > 2020年8月11日(火) 13:25 Sylwester Nawrocki <s.nawrocki@...sung.com>:
> >>
> >> In the .set_rate callback for some PLLs there is a loop polling state
> >> of the PLL lock bit and it may become an endless loop when something
> >> goes wrong with the PLL. For some PLLs there is already (a duplicated)
> >> code for polling with timeout. This patch replaces that code with
> >> the readl_relaxed_poll_timeout_atomic() macro and moves it to a common
> >> helper function, which is then used for all the PLLs. The downside
> >> of switching to the common macro is that we drop the cpu_relax() call.
> >
> > Tbh. I'm not sure what effect was exactly expected from cpu_relax() in
> > the functions which already had timeout handling. Could someone shed
> > some light on this?
> >
> >> Using a common helper function rather than the macro directly allows
> >> to avoid repeating the error message in the code and to avoid the object
> >> code size increase due to inlining.
> >>
> >> Signed-off-by: Sylwester Nawrocki <s.nawrocki@...sung.com>
> >> ---
> >> Changes for v2:
> >>  - use common readl_relaxed_poll_timeout_atomic() macro
> >> ---
> >>  drivers/clk/samsung/clk-pll.c | 92 +++++++++++++++----------------------------
> >>  1 file changed, 32 insertions(+), 60 deletions(-)
> >>
> >> diff --git a/drivers/clk/samsung/clk-pll.c b/drivers/clk/samsung/clk-pll.c
> >> index ac70ad7..c3c1efe 100644
> >> --- a/drivers/clk/samsung/clk-pll.c
> >> +++ b/drivers/clk/samsung/clk-pll.c
> >> @@ -9,13 +9,14 @@
>
> >> -#define PLL_TIMEOUT_MS         10
> >> +#define PLL_TIMEOUT_US         10000U
> >
> > I'm also wondering if 10ms is the universal value that would cover the
> > oldest PLLs as well, but my loose recollection is that they should
> > still lock much faster than that. Could you double check that in the
> > documentation?
>
> Thanks for your comments.
>
> The oldest PLLs have a hard coded 300 us waiting time for PLL lock and
> are not affected by the patch.
> I have checked some of the PLLs and maximum observed lock time was around
> 370 us and most of the time it was just a few us.
>
> We calculate the lock time in each set_rate op, in the oscillator cycle
> units, as a product of current P divider value and a constant PLL type
> specific LOCK_FACTOR. Maximum possible P value is 64, maximum possible
> LOCK_FACTOR is 3000. Assuming minimum VCO frequency of 24 MHz (which
> I think will usually be much higher than that) maximum lock time
> would be (64 x 3000) / 24 MHz = 8 ms. I think we can leave the current
> 10 ms value.

Sounds good to me. Thanks!

>
> But there is other issue, it seems we can't really use the ktime API
> in the set_rate callbacks, as these could be called early, before the
> clocksource is initialized and ktime doesn't work yet. Below trace
> is from a dump_stack() added to the samsung_pll_lock_wait() callback.
> The PLL rate setting is triggered by assigned-clock* properties in
> the clock supplier node.
> I think we need to switch to a simple udelay() loop, as is done in
> clk-tegra210 driver for instance.
>
> [    0.000000] Hardware name: Samsung Exynos (Flattened Device Tree)
> [    0.000000] [<c0111e9c>] (unwind_backtrace) from [<c010d0ec>] (show_stack+0x10/0x14)
> [    0.000000] [<c010d0ec>] (show_stack) from [<c051d890>] (dump_stack+0xac/0xd8)
> [    0.000000] [<c051d890>] (dump_stack) from [<c0578d94>] (samsung_pll_lock_wait+0x14/0x174)
> [    0.000000] [<c0578d94>] (samsung_pll_lock_wait) from [<c057319c>] (clk_change_rate+0x1a8/0x8ac)
> [    0.000000] [<c057319c>] (clk_change_rate) from [<c0573aec>] (clk_core_set_rate_nolock+0x24c/0x268)
> [    0.000000] [<c0573aec>] (clk_core_set_rate_nolock) from [<c0573b38>] (clk_set_rate+0x30/0x64)
> [    0.000000] [<c0573b38>] (clk_set_rate) from [<c0577df8>] (of_clk_set_defaults+0x214/0x384)
> [    0.000000] [<c0577df8>] (of_clk_set_defaults) from [<c0572f34>] (of_clk_add_hw_provider+0x98/0xd8)
> [    0.000000] [<c0572f34>] (of_clk_add_hw_provider) from [<c1120278>] (samsung_clk_of_add_provider+0x1c/0x30)
> [    0.000000] [<c1120278>] (samsung_clk_of_add_provider) from [<c1121844>] (exynos5250_clk_of_clk_init_driver+0x1f4/0x240)
> [    0.000000] [<c1121844>] (exynos5250_clk_of_clk_init_driver) from [<c11200d0>] (of_clk_init+0x16c/0x218)
> [    0.000000] [<c11200d0>] (of_clk_init) from [<c1104bdc>] (time_init+0x24/0x30)
> [    0.000000] [<c1104bdc>] (time_init) from [<c1100d20>] (start_kernel+0x3b0/0x520)

Yeah... I should've thought about this. Interestingly enough, some of
the existing implementations in drivers/clk/samsung/clk-pll.c use the
ktime API. I guess they are lucky enough not to be called too early,
i.e. are not needed for the initialization of timers.

Best regards,
Tomasz


> [    0.000000] [<c1100d20>] (start_kernel) from [<00000000>] (0x0)
> [    0.000000] samsung_pll_lock_wait: PLL fout_epll, lock time: 0 us, ret: 0
> [    0.000000] Exynos5250: clock setup completed, armclk=1700000000
> [    0.000000] Switching to timer-based delay loop, resolution 41ns
> [    0.000000] clocksource: mct-frc: mask: 0xffffffff max_cycles: 0xffffffff, max_idle_ns: 79635851949 ns
> [    0.000003] sched_clock: 32 bits at 24MHz, resolution 41ns, wraps every 89478484971ns
> [    0.000032] genirq: irq_chip COMBINER did not update eff. affinity mask of irq 49
> [    0.000523] arch_timer: cp15 timer(s) running at 24.00MHz (virt).
> [    0.000536] clocksource: arch_sys_counter: mask: 0xffffffffffffff max_cycles: 0x588fe9dc0, max_idle_ns: 440795202592 ns
> [    0.000551] sched_clock: 56 bits at 24MHz, resolution 41ns, wraps every 4398046511097ns
>
> --
> Regards,
> Sylwester

Powered by blists - more mailing lists