On Mar 16 16:30, Aaron Lindsay wrote: > I aim to eventually add raising interrupts on counter overflow, but that is > not > covered by this patchset. I think I have a reasonable grasp of the mechanics > of > *how* to raise them, but am curious if anyone has thoughts on how to determine > *when* to raise them - we don't want to call into PMU code every time an > instruction is executed to check if any instruction counters have overflowed, > etc. The main candidate I've seen for doing this so far would be to set up a > QEMUTimer, but I haven't fully explored it. Does that seem plausible? Any > other/better ideas?
I'm planning to post a full v4 of this patchset soon, pending a few review fixes, but I figured I'd throw out an early version of a patch to add interrupts on overflow in case it obviously has major issues that will need to be addressed. This patch sets up a QEMUTimer to get a callback when we expect counters to next overflow and triggers an interrupt at that time. Signed-off-by: Aaron Lindsay <alind...@codeaurora.org> --- target/arm/cpu.c | 11 +++++ target/arm/cpu.h | 7 +++ target/arm/helper.c | 129 ++++++++++++++++++++++++++++++++++++++++++++++++---- 3 files changed, 138 insertions(+), 9 deletions(-) diff --git a/target/arm/cpu.c b/target/arm/cpu.c index df27188..9108c6b 100644 --- a/target/arm/cpu.c +++ b/target/arm/cpu.c @@ -740,6 +740,12 @@ static void arm_cpu_finalizefn(Object *obj) QLIST_REMOVE(hook, node); g_free(hook); } +#ifndef CONFIG_USER_ONLY + if (arm_feature(&cpu->env, ARM_FEATURE_PMU)) { + timer_deinit(cpu->pmu_timer); + timer_free(cpu->pmu_timer); + } +#endif } static void arm_cpu_realizefn(DeviceState *dev, Error **errp) @@ -907,6 +913,11 @@ static void arm_cpu_realizefn(DeviceState *dev, Error **errp) arm_register_pre_el_change_hook(cpu, &pmu_pre_el_change, 0); arm_register_el_change_hook(cpu, &pmu_post_el_change, 0); + +#ifndef CONFIG_USER_ONLY + cpu->pmu_timer = timer_new(QEMU_CLOCK_VIRTUAL, 1, arm_pmu_timer_cb, + cpu); +#endif } else { cpu->pmceid0 = 0x00000000; cpu->pmceid1 = 0x00000000; diff --git a/target/arm/cpu.h b/target/arm/cpu.h index 5e6bbd3..bc0867f 100644 --- a/target/arm/cpu.h +++ b/target/arm/cpu.h @@ -703,6 +703,8 @@ struct ARMCPU { /* Timers used by the generic (architected) timer */ QEMUTimer *gt_timer[NUM_GTIMERS]; + /* Timer used by the PMU */ + QEMUTimer *pmu_timer; /* GPIO outputs for generic timer */ qemu_irq gt_timer_outputs[NUM_GTIMERS]; /* GPIO output for GICv3 maintenance interrupt signal */ @@ -934,6 +936,11 @@ void pmu_op_start(CPUARMState *env); void pmu_op_finish(CPUARMState *env); /** + * Called when a PMU counter is due to overflow + */ +void arm_pmu_timer_cb(void *opaque); + +/** * Functions to register as EL change hooks for PMU mode filtering */ void pmu_pre_el_change(ARMCPU *cpu, void *ignored); diff --git a/target/arm/helper.c b/target/arm/helper.c index 2147678..abe24dc 100644 --- a/target/arm/helper.c +++ b/target/arm/helper.c @@ -905,6 +905,7 @@ static const ARMCPRegInfo v6_cp_reginfo[] = { /* Definitions for the PMU registers */ #define PMCRN_MASK 0xf800 #define PMCRN_SHIFT 11 +#define PMCRLC 0x40 #define PMCRD 0x8 #define PMCRC 0x4 #define PMCRP 0x2 @@ -919,6 +920,8 @@ static const ARMCPRegInfo v6_cp_reginfo[] = { #define PMXEVTYPER_MT 0x02000000 #define PMXEVTYPER_EVTCOUNT 0x000003ff +#define PMEVCNTR_OVERFLOW_MASK ((uint64_t)1 << 31) + #define PMCCFILTR 0xf8000000 #define PMCCFILTR_M PMXEVTYPER_M #define PMCCFILTR_EL0 (PMCCFILTR | PMCCFILTR_M) @@ -934,6 +937,11 @@ typedef struct pm_event { /* Retrieve the current count of the underlying event. The programmed * counters hold a difference from the return value from this function */ uint64_t (*get_count)(CPUARMState *); + /* Return how many nanoseconds it will take (at a minimum) for count events + * to occur. A negative value indicates the counter will never overflow, or + * that the counter has otherwise arranged for the overflow bit to be set + * and the PMU interrupt to be raised on overflow. */ + int64_t (*ns_per_count)(uint64_t); } pm_event; static bool event_always_supported(CPUARMState *env) @@ -950,6 +958,11 @@ static uint64_t swinc_get_count(CPUARMState *env) return 0; } +static int64_t swinc_ns_per(uint64_t ignored) +{ + return -1; +} + /* * Return the underlying cycle count for the PMU cycle counters. If we're in * usermode, simply return 0. @@ -965,6 +978,11 @@ static uint64_t cycles_get_count(CPUARMState *env) } #ifndef CONFIG_USER_ONLY +static int64_t cycles_ns_per(uint64_t cycles) +{ + return ARM_CPU_FREQ/NANOSECONDS_PER_SECOND; +} + static bool instructions_supported(CPUARMState *env) { return use_icount == 1 /* Precise instruction counting */; @@ -974,22 +992,30 @@ static uint64_t instructions_get_count(CPUARMState *env) { return (uint64_t)cpu_get_icount_raw(); } + +static int64_t instructions_ns_per(uint64_t icount) +{ + return cpu_icount_to_ns((int64_t)icount); +} #endif #define SUPPORTED_EVENT_SENTINEL UINT16_MAX static const pm_event pm_events[] = { { .number = 0x000, /* SW_INCR */ .supported = event_always_supported, - .get_count = swinc_get_count + .get_count = swinc_get_count, + .ns_per_count = swinc_ns_per }, #ifndef CONFIG_USER_ONLY { .number = 0x008, /* INST_RETIRED, Instruction architecturally executed */ .supported = instructions_supported, - .get_count = instructions_get_count + .get_count = instructions_get_count, + .ns_per_count = instructions_ns_per }, { .number = 0x011, /* CPU_CYCLES, Cycle */ .supported = event_always_supported, - .get_count = cycles_get_count + .get_count = cycles_get_count, + .ns_per_count = cycles_ns_per }, #endif { .number = SUPPORTED_EVENT_SENTINEL } @@ -1168,6 +1194,13 @@ static inline bool pmu_counter_filtered(CPUARMState *env, uint64_t pmxevtyper) return false; } +static void pmu_update_irq(CPUARMState *env) +{ + ARMCPU *cpu = arm_env_get_cpu(env); + qemu_set_irq(cpu->pmu_interrupt, (env->cp15.c9_pmcr & PMCRE) && + (env->cp15.c9_pminten & env->cp15.c9_pmovsr)); +} + /* * Ensure c15_ccnt is the guest-visible count so that operations such as * enabling/disabling the counter or filtering, modifying the count itself, @@ -1186,7 +1219,18 @@ void pmccntr_op_start(CPUARMState *env) eff_cycles /= 64; } - env->cp15.c15_ccnt = eff_cycles - env->cp15.c15_ccnt_delta; + uint64_t new_pmccntr = eff_cycles - env->cp15.c15_ccnt_delta; + + unsigned int overflow_bit = (env->cp15.c9_pmcr & PMCRLC) ? 63 : 31; + uint64_t overflow_mask = (uint64_t)1 << overflow_bit; + if (!(new_pmccntr & overflow_mask) && + (env->cp15.c15_ccnt & overflow_mask)) { + env->cp15.c9_pmovsr |= (1 << 31); + new_pmccntr &= ~overflow_mask; + pmu_update_irq(env); + } + + env->cp15.c15_ccnt = new_pmccntr; } env->cp15.c15_ccnt_delta = cycles; } @@ -1200,13 +1244,25 @@ void pmccntr_op_finish(CPUARMState *env) { if (arm_ccnt_enabled(env) && !pmu_counter_filtered(env, env->cp15.pmccfiltr_el0)) { - uint64_t prev_cycles = env->cp15.c15_ccnt_delta; +#ifndef CONFIG_USER_ONLY + uint64_t delta = ((env->cp15.c9_pmcr & PMCRLC) ? + UINT64_MAX : UINT32_MAX) - (uint32_t)env->cp15.c15_ccnt; + int64_t overflow_in = cycles_ns_per(delta); + if (overflow_in >= 0) + { + int64_t overflow_at = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + + overflow_in; + ARMCPU *cpu = arm_env_get_cpu(env); + timer_mod_anticipate_ns(cpu->pmu_timer, overflow_at); + } +#endif + + uint64_t prev_cycles = env->cp15.c15_ccnt_delta; if (env->cp15.c9_pmcr & PMCRD) { /* Increment once every 64 processor clock cycles */ prev_cycles /= 64; } - env->cp15.c15_ccnt_delta = prev_cycles - env->cp15.c15_ccnt; } } @@ -1220,8 +1276,16 @@ static void pmevcntr_op_start(CPUARMState *env, uint8_t counter) if (pmu_counter_enabled(env, counter) && !pmu_counter_filtered(env, env->cp15.c14_pmevtyper[counter])) { - env->cp15.c14_pmevcntr[counter] = - count - env->cp15.c14_pmevcntr_delta[counter]; + + uint64_t new_pmevcntr = count - env->cp15.c14_pmevcntr_delta[counter]; + + if (!(new_pmevcntr & PMEVCNTR_OVERFLOW_MASK) && + (env->cp15.c14_pmevcntr[counter] & PMEVCNTR_OVERFLOW_MASK)) { + env->cp15.c9_pmovsr |= (1 << counter); + new_pmevcntr &= ~PMEVCNTR_OVERFLOW_MASK; + pmu_update_irq(env); + } + env->cp15.c14_pmevcntr[counter] = new_pmevcntr; } env->cp15.c14_pmevcntr_delta[counter] = count; } @@ -1230,6 +1294,21 @@ static void pmevcntr_op_finish(CPUARMState *env, uint8_t counter) { if (pmu_counter_enabled(env, counter) && !pmu_counter_filtered(env, env->cp15.c14_pmevtyper[counter])) { +#ifndef CONFIG_USER_ONLY + uint16_t event = env->cp15.c14_pmevtyper[counter] & PMXEVTYPER_EVTCOUNT; + uint16_t event_idx = supported_event_map[event]; + uint64_t delta = UINT32_MAX - (uint32_t)env->cp15.c14_pmevcntr[counter]; + int64_t overflow_in = pm_events[event_idx].ns_per_count(delta); + + if (overflow_in >= 0) + { + int64_t overflow_at = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + + overflow_in; + ARMCPU *cpu = arm_env_get_cpu(env); + timer_mod_anticipate_ns(cpu->pmu_timer, overflow_at); + } +#endif + env->cp15.c14_pmevcntr_delta[counter] -= env->cp15.c14_pmevcntr[counter]; } @@ -1263,6 +1342,18 @@ void pmu_post_el_change(ARMCPU *cpu, void *ignored) pmu_op_finish(&cpu->env); } +void arm_pmu_timer_cb(void *opaque) { + ARMCPU *cpu = opaque; + + /* Update all the counter values based on the current underlying counts, + * triggering interrupts to be raised, if necessary. pmu_op_finish() also + * has the effect of setting the cpu->pmu_timer to the next earliest time a + * counter may expire. + */ + pmu_op_start(&cpu->env); + pmu_op_finish(&cpu->env); +} + static void pmcr_write(CPUARMState *env, const ARMCPRegInfo *ri, uint64_t value) { @@ -1300,7 +1391,21 @@ static void pmswinc_write(CPUARMState *env, const ARMCPRegInfo *ri, /* counter is SW_INCR */ (env->cp15.c14_pmevtyper[i] & PMXEVTYPER_EVTCOUNT) == 0x0) { pmevcntr_op_start(env, i); - env->cp15.c14_pmevcntr[i]++; + + /* Detect if this write causes an overflow since we can't predict + * PMSWINC overflows like we can for other events + */ + uint64_t new_pmswinc = env->cp15.c14_pmevcntr[i] + 1; + + if (!(new_pmswinc & PMEVCNTR_OVERFLOW_MASK) && + (env->cp15.c14_pmevcntr[i] & PMEVCNTR_OVERFLOW_MASK)) { + env->cp15.c9_pmovsr |= (1 << i); + new_pmswinc &= ~PMEVCNTR_OVERFLOW_MASK; + pmu_update_irq(env); + } + + env->cp15.c14_pmevcntr[i] = new_pmswinc; + pmevcntr_op_finish(env, i); } } @@ -1371,6 +1476,7 @@ static void pmcntenset_write(CPUARMState *env, const ARMCPRegInfo *ri, { value &= PMU_COUNTER_MASK(env); env->cp15.c9_pmcnten |= value; + pmu_update_irq(env); } static void pmcntenclr_write(CPUARMState *env, const ARMCPRegInfo *ri, @@ -1378,6 +1484,7 @@ static void pmcntenclr_write(CPUARMState *env, const ARMCPRegInfo *ri, { value &= PMU_COUNTER_MASK(env); env->cp15.c9_pmcnten &= ~value; + pmu_update_irq(env); } static void pmovsr_write(CPUARMState *env, const ARMCPRegInfo *ri, @@ -1385,6 +1492,7 @@ static void pmovsr_write(CPUARMState *env, const ARMCPRegInfo *ri, { value &= PMU_COUNTER_MASK(env); env->cp15.c9_pmovsr &= ~value; + pmu_update_irq(env); } static void pmovsset_write(CPUARMState *env, const ARMCPRegInfo *ri, @@ -1392,6 +1500,7 @@ static void pmovsset_write(CPUARMState *env, const ARMCPRegInfo *ri, { value &= PMU_COUNTER_MASK(env); env->cp15.c9_pmovsr |= value; + pmu_update_irq(env); } static void pmevtyper_write(CPUARMState *env, const ARMCPRegInfo *ri, @@ -1517,6 +1626,7 @@ static void pmintenset_write(CPUARMState *env, const ARMCPRegInfo *ri, /* We have no event counters so only the C bit can be changed */ value &= PMU_COUNTER_MASK(env); env->cp15.c9_pminten |= value; + pmu_update_irq(env); } static void pmintenclr_write(CPUARMState *env, const ARMCPRegInfo *ri, @@ -1524,6 +1634,7 @@ static void pmintenclr_write(CPUARMState *env, const ARMCPRegInfo *ri, { value &= PMU_COUNTER_MASK(env); env->cp15.c9_pminten &= ~value; + pmu_update_irq(env); } static void vbar_write(CPUARMState *env, const ARMCPRegInfo *ri, -- Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc. Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.