Message-ID: <mhng-53fb904e-d47e-4887-b661-560a8dd25313@palmer-si-x1c4>
Date:   Wed, 22 Nov 2017 09:11:34 -0800 (PST)
From:   Palmer Dabbelt <palmer@...ive.com>
To:     mark.rutland@....com
CC:     robh+dt@...nel.org, devicetree@...r.kernel.org,
        patches@...ups.riscv.org, linux-kernel@...r.kernel.org
Subject:     Re: [patches] Re: [PATCH] dt-bindings: Add an enable method to RISC-V

On Tue, 21 Nov 2017 03:04:52 PST (-0800), mark.rutland@....com wrote:
> Hi Palmer,
>
> On Mon, Nov 20, 2017 at 11:50:22AM -0800, Palmer Dabbelt wrote:
>> RISC-V doesn't currently specify a mechanism for enabling or disabling
>> CPUs.  Instead, we assume that all CPUs are enabled on boot, and if
>> someone wants to save power we instead put a CPU to sleep via a WFI
>> loop.
>>
>> This patch adds "enable-method" to the RISC-V CPU binding, which
>> currently only has the value "none".  This allows us to change the
>> enable method in the future.
>
> I think you might want to be a bit more explicit about what this means,
> and this could do with a better name, as "none" sounds like the CPU is
> unusable, rather than it having been placed within the kernel already by
> the FW/bootloader (which IIUC is what happens currently).

It was proposed to make "enable-method" optional, and have the lack of an 
enable method signify the current scheme.  The current scheme is that the 
bootloader starts every hart at the kernel's entry point.

Calling this "always-enabled" was also suggested, which seems fine to me.

> As previosuly commented, I also really think you'll want to define a
> simple boot protocol (like PPC spin-table) whereby the kernel can bring
> each CPU into the kernel independently. That will save you a lot of pain
> in future with things like kexec, suspend/resume, etc.
>
> For arm64 we had a spin-table clone (implemented in our boot-wrapper
> firmware) that allowed us to bring CPUs into the kernel explicitly.
> However, we made the mistake of allowing CPUs to share a mailbox, and we
> couldn't tell how many CPUs were stuck in the kernel at any point in
> time (rendering kexec, suspend, etc impossible).

This is actually why I'm kind of pushing back on this: because we don't know 
how we're actually going to handle this, I don't want to go build an interface 
to the firmware that might be broken.  Essentially what we're doing now is just 
keeping the spin table entirely within Linux, so we can change this interface 
whenever we want.  The start of our kernel looks like

  _start(char *dtb_pointer, long hartid)
    if (atomic_increment_return(hart_lottery) == 0)
      start_kernel()
    else
      while (READ_ONCE(__cpu_up_has_turned_on_hart[hartid]) == 0)
        wait_for_interrupt()
      smp_callin()

If I understand correctly, this is essentially what the spin tables are doing 
in arm64.  Our mechanism is a bit different because we can expose a much more 
complicated interface here, but since the interface can change (it's a 
kernel-internal interface, not a firmware->kernel interface) that's the natural 
thing to do.

While I haven't actually gone through and looked at any of this (and I admit I 
have only a vague idea of how it works), I think this should work fine for 
kexec, CPU hotplug, and suspend.  kexec is easy: the fresh kernel's image will 
boot exactly like a regular one, as all the harts can just jump to the entry 
point at the same time.  Since "hart_lottery" is initialized to 0 by the ELF 
there isn't anything special required to make it work.

Actually turning off harts will require us to add an interface that does so, 
which will probably happen via an SBI call.  We haven't actually designed the 
interface yet, but I'm assuming it'll just reset the hart.  In general, we like 
to make any interface that sleeps also work as a NOP, so for now let's just 
pretend that this interface does nothing and go straight to_start.  This should 
map pretty well, our __cpu_down could just be the mirror of __cpu_up

  __cpu_down(int hartid)
    __cpu_up_has_turned_on_hart[hartid] = false;
    atomic_decrement(hart_lottery);
    __sbi_suspend_hart();
    jump _start

That should cover hotplug, and then suspend is just a matter of hotplugging out 
the last CPU.  I assume that lots of our stuff will blow up when we start 
removing harts at runtime, but that'll all happen regardless of how we wake 
them up.  There's also a bit of a race here (bringing up a hart while the last 
one is suspending), and that counter overflows, but those seem solvable.

Does that sound sane?  If not, I'd be happy to go and design a spin table 
firmware interface.  We just like to avoid inventing external interfaces until 
we really know what we're doing :).

> Thanks,
> Mark.
>
>> CC: Mark Rutland <mark.rutland@....com>
>> Signed-off-by: Palmer Dabbelt <palmer@...ive.com>
>> ---
>>  Documentation/devicetree/bindings/riscv/cpus.txt | 7 +++++++
>>  1 file changed, 7 insertions(+)
>>
>> diff --git a/Documentation/devicetree/bindings/riscv/cpus.txt b/Documentation/devicetree/bindings/riscv/cpus.txt
>> index adf7b7af5dc3..dd9e1ae197e2 100644
>> --- a/Documentation/devicetree/bindings/riscv/cpus.txt
>> +++ b/Documentation/devicetree/bindings/riscv/cpus.txt
>> @@ -82,6 +82,11 @@ described below.
>>                  Value type: <string>
>>                  Definition: Contains the RISC-V ISA string of this hart.  These
>>                              ISA strings are defined by the RISC-V ISA manual.
>> +        - cpu-enable-method:
>> +		Usage: required
>> +		Value type: <stringlist>
>> +		Definition: Must be one of
>> +			"none": This CPU's state cannot be changed.
>>
>>  Example: SiFive Freedom U540G Development Kit
>>  ---------------------------------------------
>> @@ -105,6 +110,7 @@ Linux is allowed to run on.
>>                          reg = <0>;
>>                          riscv,isa = "rv64imac";
>>                          status = "disabled";
>> +                        enable-method = "none";
>>                          L10: interrupt-controller {
>>                                  #interrupt-cells = <1>;
>>                                  compatible = "riscv,cpu-intc";
>> @@ -130,6 +136,7 @@ Linux is allowed to run on.
>>                          reg = <1>;
>>                          riscv,isa = "rv64imafdc";
>>                          status = "okay";
>> +                        enable-method = "none";
>>                          tlb-split;
>>                          L13: interrupt-controller {
>>                                  #interrupt-cells = <1>;
>> --
>> 2.13.6
>>

Powered by blists - more mailing lists