On Mon, Aug 10, 2015 at 09:26:05PM +0800, Zhichao Huang wrote: > We trap debug register accesses from guest all the time, and read the > BCR/WCR to indicate whether the guest has enabled any break/watch points > or not. > > Signed-off-by: Zhichao Huang <zhichao.huang@xxxxxxxxxx> > --- > arch/arm/include/asm/kvm_asm.h | 2 ++ > arch/arm/kvm/coproc.c | 75 +++++++++++++++++++++++++++++++++++++++--- > 2 files changed, 72 insertions(+), 5 deletions(-) > > diff --git a/arch/arm/include/asm/kvm_asm.h b/arch/arm/include/asm/kvm_asm.h > index 5b1c3eb..e9e1f0a 100644 > --- a/arch/arm/include/asm/kvm_asm.h > +++ b/arch/arm/include/asm/kvm_asm.h > @@ -65,7 +65,9 @@ > #define NR_CP14_REGS 66 /* Number of regs (incl. invalid) */ > > #define KVM_ARM_DEBUG_HOST_INUSE_SHIFT 0 > +#define KVM_ARM_DEBUG_GUEST_INUSE_SHIFT 1 > #define KVM_ARM_DEBUG_HOST_INUSE (1 << KVM_ARM_DEBUG_HOST_INUSE_SHIFT) > +#define KVM_ARM_DEBUG_GUEST_INUSE (1 << KVM_ARM_DEBUG_GUEST_INUSE_SHIFT) > > #define ARM_EXCEPTION_RESET 0 > #define ARM_EXCEPTION_UNDEFINED 1 > diff --git a/arch/arm/kvm/coproc.c b/arch/arm/kvm/coproc.c > index b37afd6..d9dcd28b 100644 > --- a/arch/arm/kvm/coproc.c > +++ b/arch/arm/kvm/coproc.c > @@ -220,7 +220,22 @@ bool access_vm_reg(struct kvm_vcpu *vcpu, > return true; > } > > -static bool trap_debug32(struct kvm_vcpu *vcpu, > +/* Indicate whether the guest has enabled any break/watch points or not. */ > +static bool guest_debug_in_use(struct kvm_vcpu *vcpu) > +{ > + unsigned int i; > + > + for (i = 0; i < ARM_MAX_BRP; i++) > + if (vcpu->arch.cp14[cp14_DBGBCR0 + i] & 0x1) > + return true; > + for (i = 0; i < ARM_MAX_WRP; i++) > + if (vcpu->arch.cp14[cp14_DBGWCR0 + i] & 0x1) > + return true; > + > + return false; > +} > + > +static bool __trap_debug32(struct kvm_vcpu *vcpu, > const struct coproc_params *p, > const struct coproc_reg *r) > { > @@ -232,6 +247,56 @@ static bool trap_debug32(struct kvm_vcpu *vcpu, > return true; > } > > +/* > + * We want to avoid world-switching all the DBG registers all the > + * time: > + * > + * When we are about to run a guest, we have the following cases: > + * > + * 1) Neither the host nor the guest has configured any [WB]points > + * 2) Only the host has configured any [WB]points > + * 3) Only the guest has configured any [WB]points > + * 4) Both the host and the guest have configured any [WB]points > + * > + * - In case (1), KVM should enable trapping and swtich the register switch > + * state on guest accesses. > + * - In cases (2), (3), and (4) we must switch the register state on each > + * entry/exit. > + * > + * For ARMv7, if the CONFIG_HAVE_HW_BREAKPOINT is set, ARM_DSCR_MDBGEN > + * is always set(ARM64 use it to indicate that debug registers are actively add space ^ > + * in use). While I like having the overall explanation of the flow somewhere, I feel this is a strange place to put it. Perhaps there is a more suitable location? > + * > + * - We add a function reading the break/watch control variables directly to > + * indicate whether the host has enabled any break/watch points or not. > + * We only call the function upon guest entry, after preempt_disable() and > + * local_irq_disable(), so there is no race for it. This paragraph of the commenting doesn't seem to fit with the rest, and didn't we cover this already in a previous patch? > + * > + * - We trap debug register accesses from guest all the time, and read the > + * BCR/WCR to indicate whether the guest has enabled any break/watch points > + * or not. do we trap all the time? not if we're switching the state I suppose? > + * > + * For this, we can keep track of the host/guest use of debug registers, > + * and skip the save/restore dance when neither the host nor the guest has > + * configured any [WB]points. > + */ > +static bool trap_debug32(struct kvm_vcpu *vcpu, > + const struct coproc_params *p, > + const struct coproc_reg *r) > +{ > + __trap_debug32(vcpu, p, r); > + > + if (p->is_write) { > + if ((vcpu->arch.cp14[r->reg] & 0x1) || > + guest_debug_in_use(vcpu)) > + vcpu->arch.debug_flags |= KVM_ARM_DEBUG_GUEST_INUSE; > + else > + vcpu->arch.debug_flags &= ~KVM_ARM_DEBUG_GUEST_INUSE; I don't understand this logic, if we are enabling one of the w/b points or if there was already an enabled w/b point, then we set the flag, but if you disable a single one then you clear the flag? It looks to me like you're mixing two approaches here; either read through all the registers whenever you need to know to set the flag or not, or you keep track of this on every read/write of the registers. > + } > + > + return true; > +} > + > /* DBGIDR (RO) Debug ID */ > static bool trap_dbgidr(struct kvm_vcpu *vcpu, > const struct coproc_params *p, > @@ -419,13 +484,13 @@ static const struct coproc_reg cp15_regs[] = { > #define DBG_BCR_BVR_WCR_WVR(n) \ > /* DBGBVRn */ \ > { CRn( 0), CRm((n)), Op1( 0), Op2( 4), is32, \ > - trap_debug32, reset_val, (cp14_DBGBVR0 + (n)), 0 }, \ > + __trap_debug32, reset_val, (cp14_DBGBVR0 + (n)), 0 }, \ > /* DBGBCRn */ \ > { CRn( 0), CRm((n)), Op1( 0), Op2( 5), is32, \ > trap_debug32, reset_val, (cp14_DBGBCR0 + (n)), 0 }, \ > /* DBGWVRn */ \ > { CRn( 0), CRm((n)), Op1( 0), Op2( 6), is32, \ > - trap_debug32, reset_val, (cp14_DBGWVR0 + (n)), 0 }, \ > + __trap_debug32, reset_val, (cp14_DBGWVR0 + (n)), 0 }, \ > /* DBGWCRn */ \ > { CRn( 0), CRm((n)), Op1( 0), Op2( 7), is32, \ > trap_debug32, reset_val, (cp14_DBGWCR0 + (n)), 0 } > @@ -462,7 +527,7 @@ static const struct coproc_reg cp14_regs[] = { > /* DBGDSAR (64bit) */ > { CRn( 0), CRm( 2), Op1( 0), Op2( 0), is64, trap_raz_wi }, > /* DBGDSCRext */ > - { CRn( 0), CRm( 2), Op1( 0), Op2( 2), is32, trap_debug32, > + { CRn( 0), CRm( 2), Op1( 0), Op2( 2), is32, __trap_debug32, > reset_val, cp14_DBGDSCRext, 0 }, > DBG_BCR_BVR_WCR_WVR(2), > /* DBGDTRTXext */ > @@ -474,7 +539,7 @@ static const struct coproc_reg cp14_regs[] = { > DBG_BCR_BVR_WCR_WVR(5), > DBG_BCR_BVR_WCR_WVR(6), > /* DBGVCR */ > - { CRn( 0), CRm( 7), Op1( 0), Op2( 0), is32, trap_debug32 }, > + { CRn( 0), CRm( 7), Op1( 0), Op2( 0), is32, __trap_debug32 }, > DBG_BCR_BVR_WCR_WVR(7), > DBG_BCR_BVR_WCR_WVR(8), > DBG_BCR_BVR_WCR_WVR(9), > -- > 1.7.12.4 > So __trap_debug32 is for the non-control registers and trap-debug32 is for the control registers? I think specifically naming the control register function trap_debug_cr would be cleaner in that case. Thanks, -Christoffer -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html