Re: [PATCH v2 1/5] RISC-V: Remove per cpu clocksource

From: Palmer Dabbelt
Date: Mon Aug 26 2019 - 19:48:21 EST


On Fri, 16 Aug 2019 11:55:14 PDT (-0700), Atish Patra wrote:
On Fri, 2019-08-16 at 17:09 +0200, Daniel Lezcano wrote:
On 31/07/2019 03:24, Atish Patra wrote:
> There is only one clocksource in RISC-V. The boot cpu initializes
> that clocksource. No need to keep a percpu data structure.

That is not what is stated in the initial patch [1].

Can you clarify that ?


I think what I meant to say was "There is only one clocksource used in
RISC-V Linux" as it is guranteed that all the timers across all the
harts are synchronized within one tick of each other [2]. Apologies for not being verbose here.

However, reading the privilege specification(1.12-draft)
Section. 3.1.10 states that
"Accurate real-time clocks (RTCs) are relatively expensive to provide
(requiring a crystal or MEMS oscillator) and have to run even when the
rest of system is powered down, and so there is usually only one in a
system located in a different frequency/voltage domain from the
processors. Hence, the RTC must be shared by all the harts in a system"

This is different from the commit text in [1].

Perhaps I misunderstood something. @Palmer ?

This is one of those places the ISA has drifted around a bit: in the user ISA there is a time CSR, and CSRs are all per-hart state so logically there is a timer per hart. We used to actually build systems this way (with an SOC agent what would actively increment each CSR whenever the RTC fired), but it ended up being impractical for a bunch of reasons. There was never a way to actually write these time CSRs from supervisor mode, but machine-mode software could write them so it would have been possible to build system that had different time values on different harts.

As a result we ended up with per-CPU timers in Linux, but they never actually worked correctly: there's a bunch of per-CPU state in the driver, but nothing to actually enforce that timer reads go to the correct hart. For example, get the time on hart 0 you'd have to IPI over to that hart, do a local CSR read, and then IPI the time back. As a result the per-CPU state never really made any sense, but it kind of just hung around because it worked fine on the systems we were building (which always had time synced up anyway) and was closer to what the spec allowed -- we didn't IPI over because time was always synchronized on systems that actually existed and the IPIs are super slow, but the scaffolding stuck around.

As part of cleaning up the privileged ISA for ratification we decided to mandate that the time CSRs on every hart are always within a single tick of each other, effectively mandating a single time across the system. This was partially motivated by Linux, but mostly by a new approach we were taking to the hypervisor specification -- rather than a hypervisor mode, we decided to just extend supervisor mode to support fast nested virtualization, which means we now have "htimedelta" (a per-hart timer offset) rather than per-hart timers. This is more efficient because the per-state stays constant so we don't need to actively tick it, and since it makes the per-hart time state unnecessary we decided to drop that extra state.

The change to global time on RISC-V systems rendered the per-CPU timers defunct, but since they weren't really doing anything they just stuck around. The cleanup seems perfectly reasonable to me, modulo the issue I've pointed out below...



[2] https://elixir.bootlin.com/linux/v5.3-rc4/source/drivers/clocksource/timer-riscv.c#L44

Thanks

-- Daniel

[1] https://lkml.org/lkml/2018/8/4/51


> Signed-off-by: Atish Patra <atish.patra@xxxxxxx>
> ---
> drivers/clocksource/timer-riscv.c | 6 ++----
> 1 file changed, 2 insertions(+), 4 deletions(-)
> > diff --git a/drivers/clocksource/timer-riscv.c
> b/drivers/clocksource/timer-riscv.c
> index 5e6038fbf115..09e031176bc6 100644
> --- a/drivers/clocksource/timer-riscv.c
> +++ b/drivers/clocksource/timer-riscv.c
> @@ -55,7 +55,7 @@ static u64 riscv_sched_clock(void)
> return get_cycles64();
> }
> > -static DEFINE_PER_CPU(struct clocksource, riscv_clocksource) = {
> +static struct clocksource riscv_clocksource = {
> .name = "riscv_clocksource",
> .rating = 300,
> .mask = CLOCKSOURCE_MASK(64),
> @@ -92,7 +92,6 @@ void riscv_timer_interrupt(void)
> static int __init riscv_timer_init_dt(struct device_node *n)
> {
> int cpuid, hartid, error;
> - struct clocksource *cs;
> > hartid = riscv_of_processor_hartid(n);
> if (hartid < 0) {
> @@ -112,8 +111,7 @@ static int __init riscv_timer_init_dt(struct
> device_node *n)
> > pr_info("%s: Registering clocksource cpuid [%d] hartid [%d]\n",
> __func__, cpuid, hartid);
> - cs = per_cpu_ptr(&riscv_clocksource, cpuid);
> - error = clocksource_register_hz(cs, riscv_timebase);
> + error = clocksource_register_hz(&riscv_clocksource,
> riscv_timebase);

Someone's client has mangled the patches, but I think there's an issue here: we're still calling the init code for every "riscv" DT entry, but there's now only a single "struct clocksource". This will result in a single clocksource being initialized multiple times, which I assume is an issue.

> if (error) {
> pr_err("RISCV timer register failed [%d] for cpu =
> [%d]\n",
> error, cpuid);
>