[PATCH 1/1] target/riscv/kvm.c: Fix the hart bit setting of AIA

[PATCH 1/1] target/riscv/kvm.c: Fix the hart bit setting of AIA

[Yong-Xuan Wang]

The hart bit setting is different with Linux AIA driver[1] when the number
of hart is power of 2. For example, when the guest has 4 harts, the
estimated result of AIA driver is 2, whereas we pass 3 to RISC-V/KVM. Since
only 2 bits are needed to represent 4 harts, update the formula to get the
accurate result.

[1] https://lore.kernel.org/all/20240307140307.646078-1-apatel@ventanamicro.com/

Signed-off-by: Yong-Xuan Wang <yongxuan.wang@sifive.com>
---
 target/riscv/kvm/kvm-cpu.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/target/riscv/kvm/kvm-cpu.c b/target/riscv/kvm/kvm-cpu.c
index 6a6c6cae80f1..388c4ddaa145 100644
--- a/target/riscv/kvm/kvm-cpu.c
+++ b/target/riscv/kvm/kvm-cpu.c
@@ -1642,7 +1642,14 @@ void kvm_riscv_aia_create(MachineState *machine, uint64_t group_shift,
         }
     }
 
-    hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;
+
+    if (max_hart_per_socket > 1) {
+        max_hart_per_socket--;
+        hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;
+    } else {
+        hart_bits = 0;
+    }
+
     ret = kvm_device_access(aia_fd, KVM_DEV_RISCV_AIA_GRP_CONFIG,
                             KVM_DEV_RISCV_AIA_CONFIG_HART_BITS,
                             &hart_bits, true, NULL);
-- 
2.17.1

Re: [PATCH 1/1] target/riscv/kvm.c: Fix the hart bit setting of AIA

[Alistair Francis]

On Mon, Apr 15, 2024 at 4:53 PM Yong-Xuan Wang <yongxuan.wang@sifive.com> wrote:
>
> The hart bit setting is different with Linux AIA driver[1] when the number
> of hart is power of 2. For example, when the guest has 4 harts, the
> estimated result of AIA driver is 2, whereas we pass 3 to RISC-V/KVM. Since
> only 2 bits are needed to represent 4 harts, update the formula to get the
> accurate result.

I don't really follow this.

Do you mind re-wording it to talk about what the specification says?
Not what Linux does.

>
> [1] https://lore.kernel.org/all/20240307140307.646078-1-apatel@ventanamicro.com/
>
> Signed-off-by: Yong-Xuan Wang <yongxuan.wang@sifive.com>
> ---
>  target/riscv/kvm/kvm-cpu.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/target/riscv/kvm/kvm-cpu.c b/target/riscv/kvm/kvm-cpu.c
> index 6a6c6cae80f1..388c4ddaa145 100644
> --- a/target/riscv/kvm/kvm-cpu.c
> +++ b/target/riscv/kvm/kvm-cpu.c
> @@ -1642,7 +1642,14 @@ void kvm_riscv_aia_create(MachineState *machine, uint64_t group_shift,
>          }
>      }
>
> -    hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;
> +
> +    if (max_hart_per_socket > 1) {
> +        max_hart_per_socket--;

Assuming there are an even number of cores (which there usually are)
won't this always result in a

> +        hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;

1 being returned by find_last_bit()?

Alistair

> +    } else {
> +        hart_bits = 0;
> +    }
> +
>      ret = kvm_device_access(aia_fd, KVM_DEV_RISCV_AIA_GRP_CONFIG,
>                              KVM_DEV_RISCV_AIA_CONFIG_HART_BITS,
>                              &hart_bits, true, NULL);
> --
> 2.17.1
>
>

Re: [PATCH 1/1] target/riscv/kvm.c: Fix the hart bit setting of AIA

[Yong-Xuan Wang]

Hi Alistair,

On Mon, Apr 29, 2024 at 11:25 AM Alistair Francis <alistair23@gmail.com> wrote:
>
> On Mon, Apr 15, 2024 at 4:53 PM Yong-Xuan Wang <yongxuan.wang@sifive.com> wrote:
> >
> > The hart bit setting is different with Linux AIA driver[1] when the number
> > of hart is power of 2. For example, when the guest has 4 harts, the
> > estimated result of AIA driver is 2, whereas we pass 3 to RISC-V/KVM. Since
> > only 2 bits are needed to represent 4 harts, update the formula to get the
> > accurate result.
>
> I don't really follow this.
>
> Do you mind re-wording it to talk about what the specification says?
> Not what Linux does.
>

Sure!

> >
> > [1] https://lore.kernel.org/all/20240307140307.646078-1-apatel@ventanamicro.com/
> >
> > Signed-off-by: Yong-Xuan Wang <yongxuan.wang@sifive.com>
> > ---
> >  target/riscv/kvm/kvm-cpu.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/target/riscv/kvm/kvm-cpu.c b/target/riscv/kvm/kvm-cpu.c
> > index 6a6c6cae80f1..388c4ddaa145 100644
> > --- a/target/riscv/kvm/kvm-cpu.c
> > +++ b/target/riscv/kvm/kvm-cpu.c
> > @@ -1642,7 +1642,14 @@ void kvm_riscv_aia_create(MachineState *machine, uint64_t group_shift,
> >          }
> >      }
> >
> > -    hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;
> > +
> > +    if (max_hart_per_socket > 1) {
> > +        max_hart_per_socket--;
>
> Assuming there are an even number of cores (which there usually are)
> won't this always result in a
>
> > +        hart_bits = find_last_bit(&max_hart_per_socket, BITS_PER_LONG) + 1;
>
> 1 being returned by find_last_bit()?
>

find_last_bit() returns the position of the leftmost set bit. The
output will not be 1 when the given input is an odd number.

Regards,
Yong-Xuan


> Alistair
>
> > +    } else {
> > +        hart_bits = 0;
> > +    }
> > +
> >      ret = kvm_device_access(aia_fd, KVM_DEV_RISCV_AIA_GRP_CONFIG,
> >                              KVM_DEV_RISCV_AIA_CONFIG_HART_BITS,
> >                              &hart_bits, true, NULL);
> > --
> > 2.17.1
> >
> >