On 6/27/23 18:19, Philippe Mathieu-Daudé wrote:
Hi Daniel, On 27/6/23 18:31, Daniel Henrique Barboza wrote:As it is today it's not possible to use '-cpu host' if the RISC-V host has RVH enabled. This is the resulting error: $ sudo ./qemu/build/qemu-system-riscv64 \ -machine virt,accel=kvm -m 2G -smp 1 \ -nographic -snapshot -kernel ./guest_imgs/Image \ -initrd ./guest_imgs/rootfs_kvm_riscv64.img \ -append "earlycon=sbi root=/dev/ram rw" \ -cpu host qemu-system-riscv64: H extension requires priv spec 1.12.0 This happens because we're checking for priv spec for all CPUs, and since we're not setting env->priv_ver for the 'host' CPU, it's being default to zero (i.e. PRIV_SPEC_1_10_0). In reality env->priv_ver does not make sense when running with the KVM 'host' CPU. It's used to gate certain CSRs/extensions during translation to make them unavailable if the hart declares an older spec version. It doesn't have any other use. E.g. OpenSBI version 1.2 retrieves the spec checking if the CSR_MCOUNTEREN, CSR_MCOUNTINHIBIT and CSR_MENVCFG CSRs are available [1]. 'priv_ver' is just one example. We're doing a lot of feature validation and setup during riscv_cpu_realize() that it doesn't apply KVM CPUs. Validating the feature set for those CPUs is a KVM problem that should be handled in KVM specific code. The new riscv_cpu_realize_features() helper contains all validation logic that are not applicable to KVM CPUs. riscv_cpu_realize() verifies if we're dealing with a KVM CPU and, if not, execute the new helper to proceed with the usual realize() logic for all other CPUs. [1] lib/sbi/sbi_hart.c, hart_detect_features() Signed-off-by: Daniel Henrique Barboza <dbarb...@ventanamicro.com> Reviewed-by: Alistair Francis <alistair.fran...@wdc.com> Reviewed-by: Andrew Jones <ajo...@ventanamicro.com> --- target/riscv/cpu.c | 43 +++++++++++++++++++++++++++++++++---------- 1 file changed, 33 insertions(+), 10 deletions(-) diff --git a/target/riscv/cpu.c b/target/riscv/cpu.c index fb8458bf74..e515dde208 100644 --- a/target/riscv/cpu.c +++ b/target/riscv/cpu.c @@ -331,6 +331,15 @@ static void set_satp_mode_default_map(RISCVCPU *cpu) } #endif +static bool riscv_running_kvm(void) +{ +#ifndef CONFIG_USER_ONLY + return kvm_enabled(); +#else + return false; +#endif +}Instead of this, ...@@ -1369,6 +1370,28 @@ static void riscv_cpu_realize(DeviceState *dev, Error **errp) } } #endif +} + +static void riscv_cpu_realize(DeviceState *dev, Error **errp) +{ + CPUState *cs = CPU(dev); + RISCVCPU *cpu = RISCV_CPU(dev); + RISCVCPUClass *mcc = RISCV_CPU_GET_CLASS(dev); + Error *local_err = NULL; + + cpu_exec_realizefn(cs, &local_err); + if (local_err != NULL) { + error_propagate(errp, local_err); + return; + } + + if (!riscv_running_kvm()) {... why not simply do: #ifndef CONFIG_USER_ONLY if (!kvm_enabled()) {+ riscv_cpu_realize_features(dev, &local_err); + if (local_err != NULL) { + error_propagate(errp, local_err); + return; + } + }#endif ?
Because a month back, when I first wrote this series, I was getting a lot of failed linux-user builds because kvm_enabled() was being called in a linux-user context. I got a little tired of it and created this wrapper that included the #ifndef. After all patches are applied we have 3 places where this function is called. There's a chance that some of them are being called in a sysemu emulation only and we could make it do with just a kvm_enabled(). I guess it doesn't hurt to remove this function and handle each case one by one.
If riscv_cpu_realize_features() is for all but KVM, then the name isn't ideal.
I guess we could rename it to riscv_cpu_realize_tcg(). We don't have any other accelerators aside from KVM and TCG, and there's a high chance that only TCG would care about this code even with other acccelerators in place. In fact, this check could also be made with "if (tcg_enabled())" instead of !kvm_enabled().This goes in line with a change you posted in qemu-ppc a few days ago. Thanks, Daniel
