From: Tao Cui KVM_GET/SET_SREGS has so far returned -ENOIOCTLCMD with an empty struct kvm_sregs, so migrating a vCPU's CSR state takes one KVM_GET/SET_ONE_REG ioctl per register -- thousands of syscalls for a large VM. Define struct kvm_sregs as a flat array of 0x184 entries (the core CSR range 0x0-0x183, up to DMWIN3) and implement the two ioctls: - get_sregs does a single vcpu_load/put, which also pulls pending interrupts into ESTAT, avoiding the per-register load/put side-effect that made ONE_REG snapshots of ESTAT order-sensitive. - set_sregs writes all CSRs via _kvm_setcsr in one pass, propagates _kvm_setcsr errors, and clears KVM_LARCH_HWCSR_USABLE up front so a mid-loop failure still forces the next vcpu_load() to reload from SW, matching KVM_SET_ONE_REG. CSRs above 0x183 (debug, breakpoint, PMU) stay on KVM_GET/SET_ONE_REG. struct kvm_sregs was empty, so enlarging it -- and the _IOWR-derived ioctl number -- breaks no userspace; userspace built against the updated UAPI header can adopt the bulk ioctl. On a Loongson 3A6000, snapshotting one vCPU's core CSR range drops from 388 KVM_GET_ONE_REG calls (~2 ms) to a single KVM_GET_SREGS ioctl (~5 us). Signed-off-by: Tao Cui --- arch/loongarch/include/uapi/asm/kvm.h | 3 ++ arch/loongarch/kvm/vcpu.c | 43 +++++++++++++++++++++++++-- 2 files changed, 44 insertions(+), 2 deletions(-) diff --git a/arch/loongarch/include/uapi/asm/kvm.h b/arch/loongarch/include/uapi/asm/kvm.h index cd0b5c11ca9c..dc091f29963d 100644 --- a/arch/loongarch/include/uapi/asm/kvm.h +++ b/arch/loongarch/include/uapi/asm/kvm.h @@ -124,7 +124,10 @@ struct kvm_sync_regs { }; /* dummy definition */ +#define KVM_LOONGARCH_NR_SREGS 0x184 + struct kvm_sregs { + __u64 csr[KVM_LOONGARCH_NR_SREGS]; }; struct kvm_iocsr_entry { diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c index 20c207d80e31..0890b68efa2a 100644 --- a/arch/loongarch/kvm/vcpu.c +++ b/arch/loongarch/kvm/vcpu.c @@ -1001,12 +1001,51 @@ static int kvm_set_reg(struct kvm_vcpu *vcpu, const struct kvm_one_reg *reg) int kvm_arch_vcpu_ioctl_get_sregs(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs) { - return -ENOIOCTLCMD; + int i; + unsigned long estat, gintc; + struct loongarch_csrs *csr = vcpu->arch.csr; + + /* + * Pull pending interrupts into ESTAT with a single vcpu_load/put so + * the ESTAT value read below matches the current interrupt state. + * This also avoids the per-register load/put side-effect that makes + * the ONE_REG path's ESTAT snapshot order-sensitive. The SW CSR + * reads run under vcpu->mutex, which serialises this ioctl. + */ + preempt_disable(); + vcpu_load(vcpu); + kvm_deliver_intr(vcpu); + vcpu->arch.aux_inuse &= ~KVM_LARCH_SWCSR_LATEST; + vcpu_put(vcpu); + preempt_enable(); + + for (i = 0; i < KVM_LOONGARCH_NR_SREGS; i++) { + if (i == LOONGARCH_CSR_ESTAT) { + gintc = kvm_read_sw_gcsr(csr, LOONGARCH_CSR_GINTC) & KVM_GINTC_IRQ_MASK; + estat = kvm_read_sw_gcsr(csr, LOONGARCH_CSR_ESTAT) & ~KVM_ESTAT_EXTI_MASK; + sregs->csr[i] = estat | (gintc << VIP_DELTA); + } else { + sregs->csr[i] = kvm_read_sw_gcsr(csr, i); + } + } + + return 0; } int kvm_arch_vcpu_ioctl_set_sregs(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs) { - return -ENOIOCTLCMD; + int i, ret; + + /* Clear first so a failing _kvm_setcsr still forces a HW reload. */ + vcpu->arch.aux_inuse &= ~KVM_LARCH_HWCSR_USABLE; + + for (i = 0; i < KVM_LOONGARCH_NR_SREGS; i++) { + ret = _kvm_setcsr(vcpu, i, sregs->csr[i]); + if (ret) + return ret; + } + + return 0; } int kvm_arch_vcpu_ioctl_get_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs) -- 2.43.0