Re: [PATCH 1/3] riscv: Enable cbo.zero only when all harts support Zicboz

[Conor Dooley]

On Fri, Jun 07, 2024 at 01:35:00PM -0700, Deepak Gupta wrote:
> On Wed, Jun 05, 2024 at 01:56:45PM -0700, Samuel Holland wrote:
> > Currently, we enable cbo.zero for usermode on each hart that supports
> > the Zicboz extension. This means that the [ms]envcfg CSR value may
> > differ between harts. Other features, such as pointer masking and CFI,
> > require setting [ms]envcfg bits on a per-thread basis. The combination
> > of these two adds quite some complexity and overhead to context
> > switching, as we would need to maintain two separate masks for the
> > per-hart and per-thread bits. Andrew Jones, who originally added Zicboz
> > support, writes[1][2]:
> > 
> >  I've approached Zicboz the same way I would approach all
> >  extensions, which is to be per-hart. I'm not currently aware of
> >  a platform that is / will be composed of harts where some have
> >  Zicboz and others don't, but there's nothing stopping a platform
> >  like that from being built.
> > 
> >  So, how about we add code that confirms Zicboz is on all harts.
> >  If any hart does not have it, then we complain loudly and disable
> >  it on all the other harts. If it was just a hardware description
> >  bug, then it'll get fixed. If there's actually a platform which
> >  doesn't have Zicboz on all harts, then, when the issue is reported,
> >  we can decide to not support it, support it with defconfig, or
> >  support it under a Kconfig guard which must be enabled by the user.
> > 
> > Let's follow his suggested solution and require the extension to be
> > available on all harts, so the envcfg CSR value does not need to change
> > when a thread migrates between harts. Since we are doing this for all
> > extensions with fields in envcfg, the CSR itself only needs to be saved/
> > restored when it is present on all harts.
> > 
> > This should not be a regression as no known hardware has asymmetric
> > Zicboz support, but if anyone reports seeing the warning, we will
> > re-evaluate our solution.
> > 
> > Link: https://lore.kernel.org/linux-riscv/20240322-168f191eeb8479b2ea169a5e@orel/ [1]
> > Link: https://lore.kernel.org/linux-riscv/20240323-28943722feb57a41fb0ff488@orel/ [2]
> > Signed-off-by: Samuel Holland <samuel.holland@xxxxxxxxxx>
> > ---
> > 
> > arch/riscv/kernel/cpufeature.c | 7 ++++++-
> > arch/riscv/kernel/suspend.c    | 4 ++--
> > 2 files changed, 8 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/riscv/kernel/cpufeature.c b/arch/riscv/kernel/cpufeature.c
> > index 5ef48cb20ee1..2879e26dbcd8 100644
> > --- a/arch/riscv/kernel/cpufeature.c
> > +++ b/arch/riscv/kernel/cpufeature.c
> > @@ -27,6 +27,8 @@
> > 
> > #define NUM_ALPHA_EXTS ('z' - 'a' + 1)
> > 
> > +static bool any_cpu_has_zicboz;
> > +
> > unsigned long elf_hwcap __read_mostly;
> > 
> > /* Host ISA bitmap */
> > @@ -92,6 +94,7 @@ static bool riscv_isa_extension_check(int id)
> > 			pr_err("Zicboz disabled as cboz-block-size present, but is not a power-of-2\n");
> > 			return false;
> > 		}
> > +		any_cpu_has_zicboz = true;
> > 		return true;
> > 	case RISCV_ISA_EXT_INVALID:
> > 		return false;
> > @@ -724,8 +727,10 @@ unsigned long riscv_get_elf_hwcap(void)
> > 
> > void riscv_user_isa_enable(void)
> > {
> > -	if (riscv_cpu_has_extension_unlikely(smp_processor_id(), RISCV_ISA_EXT_ZICBOZ))
> > +	if (riscv_has_extension_unlikely(RISCV_ISA_EXT_ZICBOZ))
> > 		csr_set(CSR_ENVCFG, ENVCFG_CBZE);
> > +	else if (any_cpu_has_zicboz)
> > +		pr_warn_once("Zicboz disabled as it is unavailable on some harts\n");
> 
> `riscv_has_extension_unlikely` will check bitmap `riscv_isa[0]` which I think gets populated
> by boot cpu (correct me if I am wrong here). So as long boot processor has the extension, it'll
> try to set it on CPU which doesn't have it.
> 
> How about doing this
> 
> `riscv_fill_hwcap_from_isa_string` checks and enables bitmap for all CPUs.
> So make a check there and if any of the CPU dont have `Zicboz`, then set a global bool
> `zicboz_cpu_not_homogenous`.

That is what riscv_fill_hwcap.*() already does, we track both what each
cpu has and what is common across all cpus.
riscv_has_extension_[un]likely() is a test for whether all cpus have the
extension.

> Now in `riscv_user_isa_enable`, check following
> 
> If `zicboz_cpu_not_homogenous` is set, then you already detected that some of the CPUs don't
> have support for `Zicboz` and thus you wouldn't set for CPU which even has the support and
> print a warning message.
> 
> If `zicboz_cpu_not_homogenous` is clear, then that means all CPUs support the feature.
> You simply enable it on hart.

Attachment: signature.asc
Description: PGP signature