Re: [PATCH 3/3] RISC-V: Implement archrandom when Zkr is available

[Samuel Ortiz <sameo@rivosinc.com>]

Hi Conor,

On Tue, Jun 27, 2023 at 08:09:08PM +0100, Conor Dooley wrote:
> On Tue, Jun 27, 2023 at 04:37:44PM +0200, Samuel Ortiz wrote:
> > The Zkr extension is ratified and provides 16 bits of entropy seed when
> > reading the SEED CSR.
> > 
> > We can implement arch_get_random_seed_longs() by doing multiple csrrw to
> > that CSR and filling an unsigned long with valid entropy bits.
> > 
> > Signed-off-by: Samuel Ortiz <sameo@rivosinc.com>
> > ---
> >  arch/riscv/include/asm/archrandom.h | 66 +++++++++++++++++++++++++++++
> >  arch/riscv/include/asm/csr.h        |  9 ++++
> >  2 files changed, 75 insertions(+)
> >  create mode 100644 arch/riscv/include/asm/archrandom.h
> > 
> > diff --git a/arch/riscv/include/asm/archrandom.h b/arch/riscv/include/asm/archrandom.h
> > new file mode 100644
> > index 000000000000..3d01aab2800a
> > --- /dev/null
> > +++ b/arch/riscv/include/asm/archrandom.h
> > @@ -0,0 +1,66 @@
> > +/* SPDX-License-Identifier: GPL-2.0 */
> > +/*
> > + * Kernel interface for the RISCV arch_random_* functions
> > + *
> > + * Copyright (c) 2022 by Rivos Inc.
> > + *
> > + */
> > +
> > +#ifndef ASM_RISCV_ARCHRANDOM_H
> > +#define ASM_RISCV_ARCHRANDOM_H
> > +
> > +#include <asm/csr.h>
> > +
> > +#define PR_PREFIX "Zkr Extension: "
> 
> Does pr_fmt() not work for you?

It would, but since this is a header that e.g. gets included in
random.h, I would have to ifndef it first.
My v2 is less verbose and gets rid of this prefix anyways.

> Also, "Zkr Extension" doesn't really seem super helpful to a punter if
> they saw it in a log. Why not s/Zkr Extension/archrandom/, or similar?
> 
> > +#define SEED_RETRY_LOOPS 10
> > +
> > +static inline bool __must_check csr_seed_long(unsigned long *v)
> > +{
> > +	unsigned int retry = SEED_RETRY_LOOPS;
> > +	unsigned int needed_seeds = sizeof(unsigned long) / 2, valid_seeds = 0;
> > +	u16 *entropy = (u16 *)v;
> > +
> > +	do {
> > +		/*
> > +		 * The SEED CSR (0x015) must be accessed with a read-write
> > +		 * instruction. Moreover, implementations must ignore the write
> > +		 * value, its purpose is to signal polling for new seed.
> > +		 */
> 
> What relevance does the second half of this comment have to the kernel?

Not much, I will remove it with v2.

> > +		unsigned long csr_seed = csr_swap(CSR_SEED, 0);
> > +
> > +		switch (csr_seed & SEED_OPST_MASK) {
> > +		case SEED_OPST_ES16:
> > +			entropy[valid_seeds++] = csr_seed & SEED_ENTROPY_MASK;
> > +			if (valid_seeds == needed_seeds)
> > +				return true;
> > +			break;
> > +
> > +		case SEED_OPST_DEAD:
> > +			pr_err_once(PR_PREFIX "Unrecoverable error\n");
> > +			return false;
> > +
> > +		case SEED_OPST_BIST:
> > +			pr_info(PR_PREFIX "On going Built-in Self Test\n");
> 
> tiny nit, "On-going"? My OCD is bother by the capitalisation otherwise.

I removed that one with v2, as it may be a little too chatty.

> > +			fallthrough;
> > +
> > +		case SEED_OPST_WAIT:
> > +		default:
> > +			continue;
> > +		}
> > +
> > +	} while (--retry);
> > +
> > +	return false;
> > +}
> > +
> > +static inline size_t __must_check arch_get_random_longs(unsigned long *v, size_t max_longs)
> > +{
> > +	return 0;
> > +}
> > +
> > +static inline size_t __must_check arch_get_random_seed_longs(unsigned long *v, size_t max_longs)
> > +{
> > +	return max_longs && riscv_isa_extension_available(NULL, ZKR) && csr_seed_long(v) ? 1 : 0;
> 
> Could you please write this in a more readable way, even if that is
> going to be a lot more verbose? I know you copied it from x86, but
> that's not really an excuse ;)

Agreed :)
Fixed as well with v2.

> Also, is there a reason that you opted not to use the alternative-backed
> riscv_has_extension_[un]likely() here?

I did not know about them, I switched to it with v2 (Made it a likely
case).

Cheers,
Samuel.