[ck] Re: [patch v1] x86/perf: Add hardware performance events support for Zhaoxin CPU.

commit 3a4ac121c2cacbf97d493fa3bc42ead88657abe4 upstream. Zhaoxin CPU has provided facilities for monitoring performance via PMU (Performance Monitor Unit), but the functionality is unused so far. Therefore, add support for zhaoxin pmu to make performance related hardware events available. The PMU is mostly an Intel Architectural PerfMon-v2 with a novel errata for the ZXC line. It supports the following events: -----------------------------------------------------------------------------------------------------------------------------------   Event                      | Event  | Umask | Description                  | Select |       | -----------------------------------------------------------------------------------------------------------------------------------   cpu-cycles                 |  82h   |  00h  | unhalt core clock   instructions               |  00h   |  00h  | number of instructions at retirement.   cache-references           |  15h   |  05h  | number of fillq pushs at the current cycle.   cache-misses               |  1ah   |  05h  | number of l2 miss pushed by fillq.   branch-instructions        |  28h   |  00h  | counts the number of branch instructions retired.   branch-misses              |  29h   |  00h  | mispredicted branch instructions at retirement.   bus-cycles                 |  83h   |  00h  | unhalt bus clock   stalled-cycles-frontend    |  01h   |  01h  | Increments each cycle the # of Uops issued by the RAT to RS.   stalled-cycles-backend     |  0fh   |  04h  | RS0/1/2/3/45 empty   L1-dcache-loads            |  68h   |  05h  | number of retire/commit load.   L1-dcache-load-misses      |  4bh   |  05h  | retired load uops whose data source followed an L1 miss.   L1-dcache-stores           |  69h   |  06h  | number of retire/commit Store,no LEA   L1-dcache-store-misses     |  62h   |  05h  | cache lines in M state evicted out of L1D due to Snoop HitM or dirty line replacement.   L1-icache-loads            |  00h   |  03h  | number of l1i cache access for valid normal fetch,including un-cacheable access.   L1-icache-load-misses      |  01h   |  03h  | number of l1i cache miss for valid normal fetch,including un-cacheable miss.   L1-icache-prefetches       |  0ah   |  03h  | number of prefetch.   L1-icache-prefetch-misses  |  0bh   |  03h  | number of prefetch miss.   dTLB-loads                 |  68h   |  05h  | number of retire/commit load   dTLB-load-misses           |  2ch   |  05h  | number of load operations miss all level tlbs and cause a tablewalk.   dTLB-stores                |  69h   |  06h  | number of retire/commit Store,no LEA   dTLB-store-misses          |  30h   |  05h  | number of store operations miss all level tlbs and cause a tablewalk.   dTLB-prefetches            |  64h   |  05h  | number of hardware pte prefetch requests dispatched out of the prefetch FIFO.   dTLB-prefetch-misses       |  65h   |  05h  | number of hardware pte prefetch requests miss the l1d data cache.   iTLB-load                  |  00h   |  00h  | actually counter instructions.   iTLB-load-misses           |  34h   |  05h  | number of code operations miss all level tlbs and cause a tablewalk. ----------------------------------------------------------------------------------------------------------------------------------- Reported-by: kbuild test robot <lkp(a)intel.com> Signed-off-by: CodyYao-oc <CodyYao-oc(a)zhaoxin.com> Signed-off-by: Peter Zijlstra (Intel) <peterz(a)infradead.org> Link: https://lkml.kernel.org/r/1586747669-4827-1-git-send-email-CodyYao-oc@zhaox… Signed-off-by: LeoLiu-oc <LeoLiu-oc(a)zhaoxin.com> ---  arch/x86/events/Makefile               |    2 +  arch/x86/events/core.c                 |    4 +  arch/x86/events/perf_event.h           |   14 +-  arch/x86/events/zhaoxin/Makefile       |    3 +  arch/x86/events/zhaoxin/core.c         |  612 +++++++++++++  arch/x86/events/zhaoxin/uncore.c       | 1101 ++++++++++++++++++++++++  arch/x86/events/zhaoxin/uncore.h       |  308 +++++++  arch/x86/kernel/cpu/perfctr-watchdog.c |    8 +  8 files changed, 2051 insertions(+), 1 deletion(-)  create mode 100644 arch/x86/events/zhaoxin/Makefile  create mode 100644 arch/x86/events/zhaoxin/core.c  create mode 100644 arch/x86/events/zhaoxin/uncore.c  create mode 100644 arch/x86/events/zhaoxin/uncore.h diff --git a/arch/x86/events/Makefile b/arch/x86/events/Makefile index b8ccdb5c9244..ad4a7c789637 100644 --- a/arch/x86/events/Makefile +++ b/arch/x86/events/Makefile @@ -2,3 +2,5 @@ obj-y                    += core.o  obj-y                    += amd/  obj-$(CONFIG_X86_LOCAL_APIC)            += msr.o  obj-$(CONFIG_CPU_SUP_INTEL)        += intel/ +obj-$(CONFIG_CPU_SUP_CENTAUR)        += zhaoxin/ +obj-$(CONFIG_CPU_SUP_ZHAOXIN)        += zhaoxin/ diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c index 45c142c0cfd2..67e060dc98a2 100644 --- a/arch/x86/events/core.c +++ b/arch/x86/events/core.c @@ -1844,6 +1844,10 @@ static int __init init_hw_perf_events(void)          err = amd_pmu_init();          x86_pmu.name = "HYGON";          break; +    case X86_VENDOR_ZHAOXIN: +    case X86_VENDOR_CENTAUR: +        err = zhaoxin_pmu_init(); +        break;      default:          err = -ENOTSUPP;      } diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h index 363e5810fb34..46c1b34eb65b 100644 --- a/arch/x86/events/perf_event.h +++ b/arch/x86/events/perf_event.h @@ -654,9 +654,12 @@ struct x86_pmu {      struct event_constraint *event_constraints;      struct x86_pmu_quirk *quirks;      int        perfctr_second_write; -    bool        late_ack;      u64        (*limit_period)(struct perf_event *event, u64 l); +    /* PMI handler bits */ +            unsigned int    late_ack    :1,

+         enabled_ack        :1, +            counter_freezing    :1;      /*       * sysfs attrs       */ @@ -1160,3 +1163,12 @@ static inline int is_ht_workaround_enabled(void)      return 0;  }  #endif /* CONFIG_CPU_SUP_INTEL */ + +#if ((defined CONFIG_CPU_SUP_CENTAUR) || (defined CONFIG_CPU_ZHAOXIN)) +int zhaoxin_pmu_init(void); +#else +static inline int zhaoxin_pmu_init(void) +{ +    return 0; +} +#endif /*CONFIG_CPU_SUP_CENTAUR or CONFIG_CPU_SUP_ZHAOXIN*/ diff --git a/arch/x86/events/zhaoxin/Makefile b/arch/x86/events/zhaoxin/Makefile new file mode 100644 index 000000000000..767d6212bac1 --- /dev/null +++ b/arch/x86/events/zhaoxin/Makefile @@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0 +obj-y    += core.o +obj-y    += uncore.o diff --git a/arch/x86/events/zhaoxin/core.c b/arch/x86/events/zhaoxin/core.c new file mode 100644 index 000000000000..c2e5bdf3893d --- /dev/null +++ b/arch/x86/events/zhaoxin/core.c @@ -0,0 +1,612 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Zhaoxin PMU; + */ + +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt + +#include <linux/stddef.h> +#include <linux/types.h> +#include <linux/init.h> +#include <linux/slab.h> +#include <linux/export.h> +#include <linux/nmi.h> + +#include <asm/cpufeature.h> +#include <asm/hardirq.h> +#include <asm/apic.h> + +#include "../perf_event.h" + +/* + * Zhaoxin PerfMon, used on zxc and later. + */ +static u64 zx_pmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = { + +    [PERF_COUNT_HW_CPU_CYCLES]        = 0x0082, +    [PERF_COUNT_HW_INSTRUCTIONS]      = 0x00c0, +    [PERF_COUNT_HW_CACHE_REFERENCES]  = 0x0515, +    [PERF_COUNT_HW_CACHE_MISSES]      = 0x051a, +    [PERF_COUNT_HW_BUS_CYCLES]        = 0x0083, +}; + +static struct event_constraint zxc_event_constraints[] __read_mostly = { + +    FIXED_EVENT_CONSTRAINT(0x0082, 1), /* unhalted core clock cycles */ +    EVENT_CONSTRAINT_END +}; + +static struct event_constraint zxd_event_constraints[] __read_mostly = { + +    FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* retired instructions */ +    FIXED_EVENT_CONSTRAINT(0x0082, 1), /* unhalted core clock cycles */ +    FIXED_EVENT_CONSTRAINT(0x0083, 2), /* unhalted bus clock cycles */ +    EVENT_CONSTRAINT_END +}; + +static __initconst const u64 zxd_hw_cache_event_ids +                [PERF_COUNT_HW_CACHE_MAX] +                [PERF_COUNT_HW_CACHE_OP_MAX] +                [PERF_COUNT_HW_CACHE_RESULT_MAX] = { +[C(L1D)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0042, +        [C(RESULT_MISS)] = 0x0538, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = 0x0043, +        [C(RESULT_MISS)] = 0x0562, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(L1I)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0300, +        [C(RESULT_MISS)] = 0x0301, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = 0x030a, +        [C(RESULT_MISS)] = 0x030b, +    }, +}, +[C(LL)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(DTLB)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0042, +        [C(RESULT_MISS)] = 0x052c, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = 0x0043, +        [C(RESULT_MISS)] = 0x0530, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = 0x0564, +        [C(RESULT_MISS)] = 0x0565, +    }, +}, +[C(ITLB)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x00c0, +        [C(RESULT_MISS)] = 0x0534, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(BPU)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0700, +        [C(RESULT_MISS)] = 0x0709, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(NODE)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +}; + +static __initconst const u64 zxe_hw_cache_event_ids +                [PERF_COUNT_HW_CACHE_MAX] +                [PERF_COUNT_HW_CACHE_OP_MAX] +                [PERF_COUNT_HW_CACHE_RESULT_MAX] = { +[C(L1D)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0568, +        [C(RESULT_MISS)] = 0x054b, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = 0x0669, +        [C(RESULT_MISS)] = 0x0562, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(L1I)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0300, +        [C(RESULT_MISS)] = 0x0301, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = 0x030a, +        [C(RESULT_MISS)] = 0x030b, +    }, +}, +[C(LL)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0, +        [C(RESULT_MISS)] = 0x0, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = 0x0, +        [C(RESULT_MISS)] = 0x0, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = 0x0, +        [C(RESULT_MISS)] = 0x0, +    }, +}, +[C(DTLB)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0568, +        [C(RESULT_MISS)] = 0x052c, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = 0x0669, +        [C(RESULT_MISS)] = 0x0530, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = 0x0564, +        [C(RESULT_MISS)] = 0x0565, +    }, +}, +[C(ITLB)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x00c0, +        [C(RESULT_MISS)] = 0x0534, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(BPU)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = 0x0028, +        [C(RESULT_MISS)] = 0x0029, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +[C(NODE)] = { +    [C(OP_READ)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_WRITE)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +    [C(OP_PREFETCH)] = { +        [C(RESULT_ACCESS)] = -1, +        [C(RESULT_MISS)] = -1, +    }, +}, +}; + +static void zhaoxin_pmu_disable_all(void) +{ +    wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); +} + +static void zhaoxin_pmu_enable_all(int added) +{ +    wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, x86_pmu.intel_ctrl); +} + +static inline u64 zhaoxin_pmu_get_status(void) +{ +    u64 status; + +    rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); + +    return status; +} + +static inline void zhaoxin_pmu_ack_status(u64 ack) +{ +    wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); +} + +static inline void zxc_pmu_ack_status(u64 ack) +{ +    /* +     * ZXC needs global control enabled in order to clear status bits. +     */ +    zhaoxin_pmu_enable_all(0); +    zhaoxin_pmu_ack_status(ack); +    zhaoxin_pmu_disable_all(); +} + +static void zhaoxin_pmu_disable_fixed(struct hw_perf_event *hwc) +{ +    int idx = hwc->idx - INTEL_PMC_IDX_FIXED; +    u64 ctrl_val, mask; + +    mask = 0xfULL << (idx * 4); + +    rdmsrl(hwc->config_base, ctrl_val); +    ctrl_val &= ~mask; +    wrmsrl(hwc->config_base, ctrl_val); +} + +static void zhaoxin_pmu_disable_event(struct perf_event *event) +{ +    struct hw_perf_event *hwc = &event->hw; + +    if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { +        zhaoxin_pmu_disable_fixed(hwc); +        return; +    } + +    x86_pmu_disable_event(event); +} + +static void zhaoxin_pmu_enable_fixed(struct hw_perf_event *hwc) +{ +    int idx = hwc->idx - INTEL_PMC_IDX_FIXED; +    u64 ctrl_val, bits, mask; + +    /* +     * Enable IRQ generation (0x8), +     * and enable ring-3 counting (0x2) and ring-0 counting (0x1) +     * if requested: +     */ +    bits = 0x8ULL; +    if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) +        bits |= 0x2; +    if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) +        bits |= 0x1; + +    bits <<= (idx * 4); +    mask = 0xfULL << (idx * 4); + +    rdmsrl(hwc->config_base, ctrl_val); +    ctrl_val &= ~mask; +    ctrl_val |= bits; +    wrmsrl(hwc->config_base, ctrl_val); +} + +static void zhaoxin_pmu_enable_event(struct perf_event *event) +{ +    struct hw_perf_event *hwc = &event->hw; + +    if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { +        zhaoxin_pmu_enable_fixed(hwc); +        return; +    } + +    __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); +} + +/* + * This handler is triggered by the local APIC, so the APIC IRQ handling + * rules apply: + */ +static int zhaoxin_pmu_handle_irq(struct pt_regs *regs) +{ +    struct perf_sample_data data; +    struct cpu_hw_events *cpuc; +    int handled = 0; +    u64 status; +    int bit; + +    cpuc = this_cpu_ptr(&cpu_hw_events); +    apic_write(APIC_LVTPC, APIC_DM_NMI); +    zhaoxin_pmu_disable_all(); +    status = zhaoxin_pmu_get_status(); +    if (!status) +        goto done; + +again: +    if (x86_pmu.enabled_ack) +        zxc_pmu_ack_status(status); +    else +        zhaoxin_pmu_ack_status(status); + +    inc_irq_stat(apic_perf_irqs); + +    /* +     * CondChgd bit 63 doesn't mean any overflow status. Ignore +     * and clear the bit. +     */ +    if (__test_and_clear_bit(63, (unsigned long *)&status)) { +        if (!status) +            goto done; +    } + +    for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { +        struct perf_event *event = cpuc->events[bit]; + +        handled++; + +        if (!test_bit(bit, cpuc->active_mask)) +            continue; + +        x86_perf_event_update(event); +        perf_sample_data_init(&data, 0, event->hw.last_period); + +        if (!x86_perf_event_set_period(event)) +            continue; + +        if (perf_event_overflow(event, &data, regs)) +            x86_pmu_stop(event, 0); +    } + +    /* +     * Repeat if there is more work to be done: +     */ +    status = zhaoxin_pmu_get_status(); +    if (status) +        goto again; + +done: +    zhaoxin_pmu_enable_all(0); +    return handled; +} + +static u64 zhaoxin_pmu_event_map(int hw_event) +{ +    return zx_pmon_event_map[hw_event]; +} + +static struct event_constraint * +zhaoxin_get_event_constraints(struct cpu_hw_events *cpuc, int idx, +            struct perf_event *event) +{ +    struct event_constraint *c; + +    if (x86_pmu.event_constraints) { +        for_each_event_constraint(c, x86_pmu.event_constraints) { +            if ((event->hw.config & c->cmask) == c->code) +                return c; +        } +    } + +    return &unconstrained; +} + +PMU_FORMAT_ATTR(event,    "config:0-7"); +PMU_FORMAT_ATTR(umask,    "config:8-15"); +PMU_FORMAT_ATTR(edge,    "config:18"); +PMU_FORMAT_ATTR(inv,    "config:23"); +PMU_FORMAT_ATTR(cmask,    "config:24-31"); + +static struct attribute *zx_arch_formats_attr[] = { +    &format_attr_event.attr, +    &format_attr_umask.attr, +    &format_attr_edge.attr, +    &format_attr_inv.attr, +    &format_attr_cmask.attr, +    NULL, +}; + +static ssize_t zhaoxin_event_sysfs_show(char *page, u64 config) +{ +    u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); + +    return x86_event_sysfs_show(page, config, event); +} + +static const struct x86_pmu zhaoxin_pmu __initconst = { +    .name            = "zhaoxin_pmu", +    .handle_irq        = zhaoxin_pmu_handle_irq, +    .disable_all        = zhaoxin_pmu_disable_all, +    .enable_all        = zhaoxin_pmu_enable_all, +    .enable            = zhaoxin_pmu_enable_event, +    .disable        = zhaoxin_pmu_disable_event, +    .hw_config        = x86_pmu_hw_config, +    .schedule_events    = x86_schedule_events, +    .eventsel        = MSR_ARCH_PERFMON_EVENTSEL0, +    .perfctr        = MSR_ARCH_PERFMON_PERFCTR0, +    .event_map        = zhaoxin_pmu_event_map, +    .max_events        = ARRAY_SIZE(zx_pmon_event_map), +    .apic            = 1, +    /* +     * For zxd/zxe, read/write operation for PMCx MSR is 48 bits. +     */ +    .max_period        = (1ULL << 47) - 1, +    .get_event_constraints    = zhaoxin_get_event_constraints, + +    .format_attrs        = zx_arch_formats_attr, +    .events_sysfs_show    = zhaoxin_event_sysfs_show, +}; + +static const struct { int id; char *name; } zx_arch_events_map[] __initconst = { +    { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, +    { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, +    { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, +    { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, +    { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, +    { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, +    { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, +}; + +static __init void zhaoxin_arch_events_quirk(void) +{ +    int bit; + +    /* disable event that reported as not presend by cpuid */ +    for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(zx_arch_events_map)) { +        zx_pmon_event_map[zx_arch_events_map[bit].id] = 0; +        pr_warn("CPUID marked event: \'%s\' unavailable\n", +            zx_arch_events_map[bit].name); +    } +} + +__init int zhaoxin_pmu_init(void) +{ +    union cpuid10_edx edx; +    union cpuid10_eax eax; +    union cpuid10_ebx ebx; +    struct event_constraint *c; +    unsigned int unused; +    int version; + +    pr_info("Welcome to pmu!\n"); + +    /* +     * Check whether the Architectural PerfMon supports +     * hw_event or not. +     */ +    cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); + +    if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT - 1) +        return -ENODEV; + +    version = eax.split.version_id; +    if (version != 2) +        return -ENODEV; + +    x86_pmu = zhaoxin_pmu; +    pr_info("Version check pass!\n"); + +    x86_pmu.version            = version; +    x86_pmu.num_counters        = eax.split.num_counters; +    x86_pmu.cntval_bits        = eax.split.bit_width; +    x86_pmu.cntval_mask        = (1ULL << eax.split.bit_width) - 1; +    x86_pmu.events_maskl        = ebx.full; +    x86_pmu.events_mask_len        = eax.split.mask_length; + +    x86_pmu.num_counters_fixed = edx.split.num_counters_fixed; +    x86_add_quirk(zhaoxin_arch_events_quirk); + +    switch (boot_cpu_data.x86) { +    case 0x06: +        if (boot_cpu_data.x86_model == 0x0f || boot_cpu_data.x86_model == 0x19) { + +            x86_pmu.max_period = x86_pmu.cntval_mask >> 1; + +            /* Clearing status works only if the global control is enable on zxc. */ +            x86_pmu.enabled_ack = 1; + +            x86_pmu.event_constraints = zxc_event_constraints; +            zx_pmon_event_map[PERF_COUNT_HW_INSTRUCTIONS] = 0; +            zx_pmon_event_map[PERF_COUNT_HW_CACHE_REFERENCES] = 0; +            zx_pmon_event_map[PERF_COUNT_HW_CACHE_MISSES] = 0; +            zx_pmon_event_map[PERF_COUNT_HW_BUS_CYCLES] = 0; + +            pr_cont("C events, "); +     break; +        } +        return -ENODEV; + +    case 0x07: +        zx_pmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = +            X86_CONFIG(.event = 0x01, .umask = 0x01, .inv = 0x01, .cmask = 0x01); + +        zx_pmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = +            X86_CONFIG(.event = 0x0f, .umask = 0x04, .inv = 0, .cmask = 0); + +        switch (boot_cpu_data.x86_model) { +        case 0x1b: +            memcpy(hw_cache_event_ids, zxd_hw_cache_event_ids, +                   sizeof(hw_cache_event_ids)); + +            x86_pmu.event_constraints = zxd_event_constraints; + +            zx_pmon_event_map[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x0700; +            zx_pmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x0709; + +            pr_cont("D events, "); +            break; +        case 0x3b: +            memcpy(hw_cache_event_ids, zxe_hw_cache_event_ids, +                   sizeof(hw_cache_event_ids)); + +            x86_pmu.event_constraints = zxd_event_constraints; + +            zx_pmon_event_map[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x0028; +            zx_pmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x0029; + +            pr_cont("E events, "); +            break; +        default: +            return -ENODEV; +        } +        break; + +    default: +        return -ENODEV; +    } + +    x86_pmu.intel_ctrl = (1 << (x86_pmu.num_counters)) - 1; +    x86_pmu.intel_ctrl |= ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; + +    if (x86_pmu.event_constraints) { +        for_each_event_constraint(c, x86_pmu.event_constraints) { +            c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; +            c->weight += x86_pmu.num_counters; +        } +    } + +    return 0; +} diff --git a/arch/x86/events/zhaoxin/uncore.c b/arch/x86/events/zhaoxin/uncore.c new file mode 100644 index 000000000000..4c4ea01d23c8 --- /dev/null +++ b/arch/x86/events/zhaoxin/uncore.c @@ -0,0 +1,1101 @@ +// SPDX-License-Identifier: GPL-2.0-only +#include <linux/module.h> + +#include <asm/cpu_device_id.h> +#include "uncore.h" + +static struct zhaoxin_uncore_type *empty_uncore[] = { NULL, }; +static struct zhaoxin_uncore_type **uncore_msr_uncores = empty_uncore; + +/* mask of cpus that collect uncore events */ +static cpumask_t uncore_cpu_mask; + +/* constraint for the fixed counter */ +static struct event_constraint uncore_constraint_fixed = +    EVENT_CONSTRAINT(~0ULL, 1 << UNCORE_PMC_IDX_FIXED, ~0ULL); + +static int max_packages; + +/* CHX event control */ +#define CHX_UNC_CTL_EV_SEL_MASK            0x000000ff +#define CHX_UNC_CTL_UMASK_MASK            0x0000ff00 +#define CHX_UNC_CTL_EDGE_DET            (1 << 18) +#define CHX_UNC_CTL_EN                (1 << 22) +#define CHX_UNC_CTL_INVERT            (1 << 23) +#define CHX_UNC_CTL_CMASK_MASK            0xff000000 +#define CHX_UNC_FIXED_CTR_CTL_EN        (1 << 0) + +#define CHX_UNC_RAW_EVENT_MASK (CHX_UNC_CTL_EV_SEL_MASK | \ +                         CHX_UNC_CTL_UMASK_MASK | \ +                         CHX_UNC_CTL_EDGE_DET | \ +                         CHX_UNC_CTL_INVERT | \ +                         CHX_UNC_CTL_CMASK_MASK) + +/* CHX global control register */ +#define CHX_UNC_PERF_GLOBAL_CTL                 0x391 +#define CHX_UNC_FIXED_CTR                       0x394 +#define CHX_UNC_FIXED_CTR_CTRL                  0x395 + +/* CHX uncore global control */ +#define CHX_UNC_GLOBAL_CTL_EN_PC_ALL            ((1ULL << 4) - 1) +#define CHX_UNC_GLOBAL_CTL_EN_FC                (1ULL << 32) + +/* CHX uncore register */ +#define CHX_UNC_PERFEVTSEL0                     0x3c0 +#define CHX_UNC_UNCORE_PMC0                     0x3b0 + +DEFINE_UNCORE_FORMAT_ATTR(event, event, "config:0-7"); +DEFINE_UNCORE_FORMAT_ATTR(umask, umask, "config:8-15"); +DEFINE_UNCORE_FORMAT_ATTR(edge, edge, "config:18"); +DEFINE_UNCORE_FORMAT_ATTR(inv, inv, "config:23"); +DEFINE_UNCORE_FORMAT_ATTR(cmask8, cmask, "config:24-31"); + +ssize_t zx_uncore_event_show(struct kobject *kobj, struct kobj_attribute *attr, char *buf) +{ +    struct uncore_event_desc *event = +        container_of(attr, struct uncore_event_desc, attr); +    return sprintf(buf, "%s", event->config); +} + +/*chx uncore support */ +static void chx_uncore_msr_disable_event(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    wrmsrl(event->hw.config_base, 0); +} + +static u64 uncore_msr_read_counter(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    u64 count; + +    rdmsrl(event->hw.event_base, count); + +    return count; +} + +static void chx_uncore_msr_disable_box(struct zhaoxin_uncore_box *box) +{ +    wrmsrl(CHX_UNC_PERF_GLOBAL_CTL, 0); +} + +static void chx_uncore_msr_enable_box(struct zhaoxin_uncore_box *box) +{ +    wrmsrl(CHX_UNC_PERF_GLOBAL_CTL, CHX_UNC_GLOBAL_CTL_EN_PC_ALL | CHX_UNC_GLOBAL_CTL_EN_FC); +} + +static void chx_uncore_msr_enable_event(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    struct hw_perf_event *hwc = &event->hw; + +    if (hwc->idx < UNCORE_PMC_IDX_FIXED) +        wrmsrl(hwc->config_base, hwc->config | CHX_UNC_CTL_EN); +    else +        wrmsrl(hwc->config_base, CHX_UNC_FIXED_CTR_CTL_EN); +} + +static struct attribute *chx_uncore_formats_attr[] = { +    &format_attr_event.attr, +    &format_attr_umask.attr, +    &format_attr_edge.attr, +    &format_attr_inv.attr, +    &format_attr_cmask8.attr, +    NULL, +}; + +static struct attribute_group chx_uncore_format_group = { +    .name = "format", +    .attrs = chx_uncore_formats_attr, +}; + +static struct uncore_event_desc chx_uncore_events[] = { +    { /* end: all zeroes */ }, +}; + +static struct zhaoxin_uncore_ops chx_uncore_msr_ops = { +    .disable_box    = chx_uncore_msr_disable_box, +    .enable_box    = chx_uncore_msr_enable_box, +    .disable_event    = chx_uncore_msr_disable_event, +    .enable_event    = chx_uncore_msr_enable_event, +    .read_counter    = uncore_msr_read_counter, +}; + +static struct zhaoxin_uncore_type chx_uncore_box = { +    .name        = "", +    .num_counters   = 4, +    .num_boxes    = 1, +    .perf_ctr_bits    = 48, +    .fixed_ctr_bits    = 48, +    .event_ctl    = CHX_UNC_PERFEVTSEL0, +    .perf_ctr    = CHX_UNC_UNCORE_PMC0, +    .fixed_ctr    = CHX_UNC_FIXED_CTR, +    .fixed_ctl    = CHX_UNC_FIXED_CTR_CTRL, +    .event_mask    = CHX_UNC_RAW_EVENT_MASK, +    .event_descs    = chx_uncore_events, +    .ops        = &chx_uncore_msr_ops, +    .format_group    = &chx_uncore_format_group, +}; + +static struct zhaoxin_uncore_type *chx_msr_uncores[] = { +    &chx_uncore_box, +    NULL, +}; + +static struct zhaoxin_uncore_box *uncore_pmu_to_box(struct zhaoxin_uncore_pmu *pmu, int cpu) +{ +    unsigned int package_id = topology_logical_package_id(cpu); + +    /* +     * The unsigned check also catches the '-1' return value for non +     * existent mappings in the topology map. +     */ +    return package_id < max_packages ? pmu->boxes[package_id] : NULL; +} + +static void uncore_assign_hw_event(struct zhaoxin_uncore_box *box, +                   struct perf_event *event, int idx) +{ +    struct hw_perf_event *hwc = &event->hw; + +    hwc->idx = idx; +    hwc->last_tag = ++box->tags[idx]; + +    if (uncore_pmc_fixed(hwc->idx)) { +        hwc->event_base = uncore_fixed_ctr(box); +        hwc->config_base = uncore_fixed_ctl(box); +        return; +    } + +    hwc->config_base = uncore_event_ctl(box, hwc->idx); +    hwc->event_base  = uncore_perf_ctr(box, hwc->idx); +} + +void uncore_perf_event_update(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    u64 prev_count, new_count, delta; +    int shift; + +    if (uncore_pmc_fixed(event->hw.idx)) +        shift = 64 - uncore_fixed_ctr_bits(box); +    else +        shift = 64 - uncore_perf_ctr_bits(box); + +    /* the hrtimer might modify the previous event value */ +again: +    prev_count = local64_read(&event->hw.prev_count); +    new_count = uncore_read_counter(box, event); +    if (local64_xchg(&event->hw.prev_count, new_count) != prev_count) +        goto again; + +    delta = (new_count << shift) - (prev_count << shift); +    delta >>= shift; + +    local64_add(delta, &event->count); +} + +static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer) +{ +    struct zhaoxin_uncore_box *box; +    struct perf_event *event; +    unsigned long flags; +    int bit; + +    box = container_of(hrtimer, struct zhaoxin_uncore_box, hrtimer); +    if (!box->n_active || box->cpu != smp_processor_id()) +        return HRTIMER_NORESTART; +    /* +     * disable local interrupt to prevent uncore_pmu_event_start/stop +     * to interrupt the update process +     */ +    local_irq_save(flags); + +    /* +     * handle boxes with an active event list as opposed to active +     * counters +     */ +    list_for_each_entry(event, &box->active_list, active_entry) { +        uncore_perf_event_update(box, event); +    } + +    for_each_set_bit(bit, box->active_mask, UNCORE_PMC_IDX_MAX) +        uncore_perf_event_update(box, box->events[bit]); + +    local_irq_restore(flags); + +    hrtimer_forward_now(hrtimer, ns_to_ktime(box->hrtimer_duration)); +    return HRTIMER_RESTART; +} + +static void uncore_pmu_start_hrtimer(struct zhaoxin_uncore_box *box) +{ +    hrtimer_start(&box->hrtimer, ns_to_ktime(box->hrtimer_duration), +              HRTIMER_MODE_REL_PINNED); +} + +static void uncore_pmu_cancel_hrtimer(struct zhaoxin_uncore_box *box) +{ +    hrtimer_cancel(&box->hrtimer); +} + +static void uncore_pmu_init_hrtimer(struct zhaoxin_uncore_box *box) +{ +    hrtimer_init(&box->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); +    box->hrtimer.function = uncore_pmu_hrtimer; +} + +static struct zhaoxin_uncore_box *uncore_alloc_box(struct zhaoxin_uncore_type *type, +                         int node) +{ +    int i, size, numshared = type->num_shared_regs; +    struct zhaoxin_uncore_box *box; + +    size = sizeof(*box) + numshared * sizeof(struct zhaoxin_uncore_extra_reg); + +    box = kzalloc_node(size, GFP_KERNEL, node); +    if (!box) +        return NULL; + +    for (i = 0; i < numshared; i++) +        raw_spin_lock_init(&box->shared_regs[i].lock); + +    uncore_pmu_init_hrtimer(box); +    box->cpu = -1; +    box->package_id = -1; + +    /* set default hrtimer timeout */ +    box->hrtimer_duration = UNCORE_PMU_HRTIMER_INTERVAL; + +    INIT_LIST_HEAD(&box->active_list); + +    return box; +} + +static bool is_box_event(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    return &box->pmu->pmu == event->pmu; +} + +static struct event_constraint * +uncore_get_event_constraint(struct zhaoxin_uncore_box *box, struct perf_event *event) +{ +    struct zhaoxin_uncore_type *type = box->pmu->type; +    struct event_constraint *c; + +    if (type->ops->get_constraint) { +        c = type->ops->get_constraint(box, event); +        if (c) +            return c; +    } + +    if (event->attr.config == UNCORE_FIXED_EVENT) +        return &uncore_constraint_fixed; + +    if (type->constraints) { +        for_each_event_constraint(c, type->constraints) { +            if ((event->hw.config & c->cmask) == c->code) +                return c; +        } +    } + +    return &type->unconstrainted; +} + +static void uncore_put_event_constraint(struct zhaoxin_uncore_box *box, +                    struct perf_event *event) +{ +    if (box->pmu->type->ops->put_constraint) +        box->pmu->type->ops->put_constraint(box, event); +} + +static int uncore_assign_events(struct zhaoxin_uncore_box *box, int assign[], int n) +{ +    unsigned long used_mask[BITS_TO_LONGS(UNCORE_PMC_IDX_MAX)]; +    struct event_constraint *c; +    int i, wmin, wmax, ret = 0; +    struct hw_perf_event *hwc; + +    bitmap_zero(used_mask, UNCORE_PMC_IDX_MAX); + +    for (i = 0, wmin = UNCORE_PMC_IDX_MAX, wmax = 0; i < n; i++) { +        c = uncore_get_event_constraint(box, box->event_list[i]); +        box->event_constraint[i] = c; +        wmin = min(wmin, c->weight); +        wmax = max(wmax, c->weight); +    } + +    /* fastpath, try to reuse previous register */ +    for (i = 0; i < n; i++) { +        hwc = &box->event_list[i]->hw; +        c = box->event_constraint[i]; + +        /* never assigned */ +        if (hwc->idx == -1) +            break; + +        /* constraint still honored */ +        if (!test_bit(hwc->idx, c->idxmsk)) +            break; + +        /* not already used */ +        if (test_bit(hwc->idx, used_mask)) +            break; + +        __set_bit(hwc->idx, used_mask); +        if (assign) +            assign[i] = hwc->idx; +    } +    /* slow path */ +    if (i != n) +        ret = perf_assign_events(box->event_constraint, n, +                     wmin, wmax, n, assign); + +    if (!assign || ret) { +        for (i = 0; i < n; i++) +            uncore_put_event_constraint(box, box->event_list[i]); +    } +    return ret ? -EINVAL : 0; +} + +static void uncore_pmu_event_start(struct perf_event *event, int flags) +{ +    struct zhaoxin_uncore_box *box = uncore_event_to_box(event); +    int idx = event->hw.idx; + + +    if (WARN_ON_ONCE(idx == -1 || idx >= UNCORE_PMC_IDX_MAX)) +        return; + +    if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) +        return; + +    event->hw.state = 0; +    box->events[idx] = event; +    box->n_active++; +    __set_bit(idx, box->active_mask); + +    local64_set(&event->hw.prev_count, uncore_read_counter(box, event)); +    uncore_enable_event(box, event); + +    if (box->n_active == 1) { +        uncore_enable_box(box); +        uncore_pmu_start_hrtimer(box); +    } +} + +static void uncore_pmu_event_stop(struct perf_event *event, int flags) +{ +    struct zhaoxin_uncore_box *box = uncore_event_to_box(event); +    struct hw_perf_event *hwc = &event->hw; + +    if (__test_and_clear_bit(hwc->idx, box->active_mask)) { +        uncore_disable_event(box, event); +        box->n_active--; +        box->events[hwc->idx] = NULL; +        WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED); +        hwc->state |= PERF_HES_STOPPED; + +        if (box->n_active == 0) { +            uncore_disable_box(box); +            uncore_pmu_cancel_hrtimer(box); +        } +    } + +    if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) { +        /* +         * Drain the remaining delta count out of a event +         * that we are disabling: +         */ +        uncore_perf_event_update(box, event); +        hwc->state |= PERF_HES_UPTODATE; +    } +} + +static int +uncore_collect_events(struct zhaoxin_uncore_box *box, struct perf_event *leader, +              bool dogrp) +{ +    struct perf_event *event; +    int n, max_count; + +    max_count = box->pmu->type->num_counters; +    if (box->pmu->type->fixed_ctl) +        max_count++; + +    if (box->n_events >= max_count) +        return -EINVAL; + +    n = box->n_events; + +    if (is_box_event(box, leader)) { +        box->event_list[n] = leader; +        n++; +    } + +    if (!dogrp) +        return n; + +    for_each_sibling_event(event, leader) { +        if (!is_box_event(box, event) || +            event->state <= PERF_EVENT_STATE_OFF) +            continue; + +        if (n >= max_count) +            return -EINVAL; + +        box->event_list[n] = event; +        n++; +    } +    return n; +} + +static int uncore_pmu_event_add(struct perf_event *event, int flags) +{ +    struct zhaoxin_uncore_box *box = uncore_event_to_box(event); +    struct hw_perf_event *hwc = &event->hw; +    int assign[UNCORE_PMC_IDX_MAX]; +    int i, n, ret; + +    if (!box) +        return -ENODEV; + +    ret = n = uncore_collect_events(box, event, false); +    if (ret < 0) +        return ret; + +    hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED; +    if (!(flags & PERF_EF_START)) +        hwc->state |= PERF_HES_ARCH; + +    ret = uncore_assign_events(box, assign, n); +    if (ret) +        return ret; + +    /* save events moving to new counters */ +    for (i = 0; i < box->n_events; i++) { +        event = box->event_list[i]; +        hwc = &event->hw; + +        if (hwc->idx == assign[i] && +            hwc->last_tag == box->tags[assign[i]]) +            continue; +        /* +         * Ensure we don't accidentally enable a stopped +         * counter simply because we rescheduled. +         */ +        if (hwc->state & PERF_HES_STOPPED) +            hwc->state |= PERF_HES_ARCH; + +        uncore_pmu_event_stop(event, PERF_EF_UPDATE); +    } + +    /* reprogram moved events into new counters */ +    for (i = 0; i < n; i++) { +        event = box->event_list[i]; +        hwc = &event->hw; + +        if (hwc->idx != assign[i] || +            hwc->last_tag != box->tags[assign[i]]) +            uncore_assign_hw_event(box, event, assign[i]); +        else if (i < box->n_events) +            continue; + +        if (hwc->state & PERF_HES_ARCH) +            continue; + +        uncore_pmu_event_start(event, 0); +    } +    box->n_events = n; + +    return 0; +} + +static int uncore_validate_group(struct zhaoxin_uncore_pmu *pmu, +                struct perf_event *event) +{ +    struct perf_event *leader = event->group_leader; +    struct zhaoxin_uncore_box *fake_box; +    int ret = -EINVAL, n; + +    fake_box = uncore_alloc_box(pmu->type, NUMA_NO_NODE); +    if (!fake_box) +        return -ENOMEM; + +    fake_box->pmu = pmu; +    /* +     * the event is not yet connected with its +     * siblings therefore we must first collect +     * existing siblings, then add the new event +     * before we can simulate the scheduling +     */ +    n = uncore_collect_events(fake_box, leader, true); +    if (n < 0) +        goto out; + +    fake_box->n_events = n; +    n = uncore_collect_events(fake_box, event, false); +    if (n < 0) +        goto out; + +    fake_box->n_events = n; + +    ret = uncore_assign_events(fake_box, NULL, n); +out: +    kfree(fake_box); +    return ret; +} + +static void uncore_pmu_event_del(struct perf_event *event, int flags) +{ +    struct zhaoxin_uncore_box *box = uncore_event_to_box(event); +    int i; + +    uncore_pmu_event_stop(event, PERF_EF_UPDATE); + +    for (i = 0; i < box->n_events; i++) { +        if (event == box->event_list[i]) { +            uncore_put_event_constraint(box, event); + +            for (++i; i < box->n_events; i++) +                box->event_list[i - 1] = box->event_list[i]; + +            --box->n_events; +            break; +        } +    } + +    event->hw.idx = -1; +    event->hw.last_tag = ~0ULL; +} + +static void uncore_pmu_event_read(struct perf_event *event) +{ +    struct zhaoxin_uncore_box *box = uncore_event_to_box(event); + +    uncore_perf_event_update(box, event); +} + +static int uncore_pmu_event_init(struct perf_event *event) +{ +    struct zhaoxin_uncore_pmu *pmu; +    struct zhaoxin_uncore_box *box; +    struct hw_perf_event *hwc = &event->hw; +    int ret; + +    if (event->attr.type != event->pmu->type) +        return -ENOENT; + +    pmu = uncore_event_to_pmu(event); +    /* no device found for this pmu */ +    if (pmu->func_id < 0) +        return -ENOENT; + +    /* Sampling not supported yet */ +    if (hwc->sample_period) +        return -EINVAL; + +    /* +     * Place all uncore events for a particular physical package +     * onto a single cpu +     */ +    if (event->cpu < 0) +        return -EINVAL; +    box = uncore_pmu_to_box(pmu, event->cpu); +    if (!box || box->cpu < 0) +        return -EINVAL; +    event->cpu = box->cpu; +    event->pmu_private = box; + +    event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG; + +    event->hw.idx = -1; +    event->hw.last_tag = ~0ULL; +    event->hw.extra_reg.idx = EXTRA_REG_NONE; +    event->hw.branch_reg.idx = EXTRA_REG_NONE; + +    if (event->attr.config == UNCORE_FIXED_EVENT) { +        /* no fixed counter */ +        if (!pmu->type->fixed_ctl) +            return -EINVAL; +        /* +         * if there is only one fixed counter, only the first pmu +         * can access the fixed counter +         */ +        if (pmu->type->single_fixed && pmu->pmu_idx > 0) +            return -EINVAL; + +        /* fixed counters have event field hardcoded to zero */ +        hwc->config = 0ULL; +    } else { +        hwc->config = event->attr.config & +                  (pmu->type->event_mask | ((u64)pmu->type->event_mask_ext << 32)); +        if (pmu->type->ops->hw_config) { +            ret = pmu->type->ops->hw_config(box, event); +            if (ret) +                return ret; +        } +    } + +    if (event->group_leader != event) +        ret = uncore_validate_group(pmu, event); +    else +        ret = 0; + +    return ret; +} + +static ssize_t uncore_get_attr_cpumask(struct device *dev, struct device_attribute *attr, char *buf) +{ +    return cpumap_print_to_pagebuf(true, buf, &uncore_cpu_mask); +} + +static DEVICE_ATTR(cpumask, S_IRUGO, uncore_get_attr_cpumask, NULL); + +static struct attribute *uncore_pmu_attrs[] = { +    &dev_attr_cpumask.attr, +    NULL, +}; + +static const struct attribute_group uncore_pmu_attr_group = { +    .attrs = uncore_pmu_attrs, +}; + +static void uncore_pmu_unregister(struct zhaoxin_uncore_pmu *pmu) +{ +    if (!pmu->registered) +        return; +    perf_pmu_unregister(&pmu->pmu); +    pmu->registered = false; +} + +static void uncore_free_boxes(struct zhaoxin_uncore_pmu *pmu) +{ +    int package; + +    for (package = 0; package < max_packages; package++) +        kfree(pmu->boxes[package]); +    kfree(pmu->boxes); +} + +static void uncore_type_exit(struct zhaoxin_uncore_type *type) +{ +    struct zhaoxin_uncore_pmu *pmu = type->pmus; +    int i; + +    if (pmu) { +        for (i = 0; i < type->num_boxes; i++, pmu++) { +            uncore_pmu_unregister(pmu); +            uncore_free_boxes(pmu); +        } +        kfree(type->pmus); +        type->pmus = NULL; +    } +    kfree(type->events_group); +    type->events_group = NULL; +} + +static void uncore_types_exit(struct zhaoxin_uncore_type **types) +{ +    for (; *types; types++) +        uncore_type_exit(*types); +} + +static int __init uncore_type_init(struct zhaoxin_uncore_type *type, bool setid) +{ +    struct zhaoxin_uncore_pmu *pmus; +    size_t size; +    int i, j; + +    pmus = kcalloc(type->num_boxes, sizeof(*pmus), GFP_KERNEL); +    if (!pmus) +        return -ENOMEM; + +    size = max_packages*sizeof(struct zhaoxin_uncore_box *); + +    for (i = 0; i < type->num_boxes; i++) { +        pmus[i].func_id    = setid ? i : -1; +        pmus[i].pmu_idx    = i; +        pmus[i].type    = type; +        pmus[i].boxes    = kzalloc(size, GFP_KERNEL); +        if (!pmus[i].boxes) +            goto err; +    } + +    type->pmus = pmus; +    type->unconstrainted = (struct event_constraint) +        __EVENT_CONSTRAINT(0, (1ULL << type->num_counters) - 1, +                0, type->num_counters, 0, 0); + +    if (type->event_descs) { +        struct { +            struct attribute_group group; +            struct attribute *attrs[]; +        } *attr_group; +        for (i = 0; type->event_descs[i].attr.attr.name; i++) +            ; + +        attr_group = kzalloc(struct_size(attr_group, attrs, i + 1), GFP_KERNEL); +        if (!attr_group) +            goto err; + +        attr_group->group.name = "events"; +        attr_group->group.attrs = attr_group->attrs; + +        for (j = 0; j < i; j++) +            attr_group->attrs[j] = &type->event_descs[j].attr.attr; + +        type->events_group = &attr_group->group; +    } + +    type->pmu_group = &uncore_pmu_attr_group; + +    return 0; + +err: +    for (i = 0; i < type->num_boxes; i++) +        kfree(pmus[i].boxes); +    kfree(pmus); + +    return -ENOMEM; +} + +static int __init +uncore_types_init(struct zhaoxin_uncore_type **types, bool setid) +{ +    int ret; + +    for (; *types; types++) { +        ret = uncore_type_init(*types, setid); +        if (ret) +            return ret; +    } +    return 0; +} + +static void uncore_change_type_ctx(struct zhaoxin_uncore_type *type, int old_cpu, +                   int new_cpu) +{ +    struct zhaoxin_uncore_pmu *pmu = type->pmus; +    struct zhaoxin_uncore_box *box; +    int i, package; + +    package = topology_logical_package_id(old_cpu < 0 ? new_cpu : old_cpu); +    for (i = 0; i < type->num_boxes; i++, pmu++) { +        box = pmu->boxes[package]; +        if (!box) +            continue; + +        if (old_cpu < 0) { +            WARN_ON_ONCE(box->cpu != -1); +            box->cpu = new_cpu; +            continue; +        } + +        WARN_ON_ONCE(box->cpu != old_cpu); +        box->cpu = -1; +        if (new_cpu < 0) +            continue; + +        uncore_pmu_cancel_hrtimer(box); +        perf_pmu_migrate_context(&pmu->pmu, old_cpu, new_cpu); +        box->cpu = new_cpu; +    } +} + +static void uncore_change_context(struct zhaoxin_uncore_type **uncores, +                  int old_cpu, int new_cpu) +{ +    for (; *uncores; uncores++) +        uncore_change_type_ctx(*uncores, old_cpu, new_cpu); +} + +static void uncore_box_unref(struct zhaoxin_uncore_type **types, int id) +{ +    struct zhaoxin_uncore_type *type; +    struct zhaoxin_uncore_pmu *pmu; +    struct zhaoxin_uncore_box *box; +    int i; + +    for (; *types; types++) { +        type = *types; +        pmu = type->pmus; +        for (i = 0; i < type->num_boxes; i++, pmu++) { +            box = pmu->boxes[id]; +            if (box && atomic_dec_return(&box->refcnt) == 0) +                uncore_box_exit(box); +        } +    } +} + +static int uncore_event_cpu_offline(unsigned int cpu) +{ +    int package, target; + +    /* Check if exiting cpu is used for collecting uncore events */ +    if (!cpumask_test_and_clear_cpu(cpu, &uncore_cpu_mask)) +        goto unref; +    /* Find a new cpu to collect uncore events */ +    target = cpumask_any_but(topology_core_cpumask(cpu), cpu); + +    /* Migrate uncore events to the new target */ +    if (target < nr_cpu_ids) +        cpumask_set_cpu(target, &uncore_cpu_mask); +    else +        target = -1; + +    uncore_change_context(uncore_msr_uncores, cpu, target); + +unref: +    /* Clear the references */ +    package = topology_logical_package_id(cpu); +    uncore_box_unref(uncore_msr_uncores, package); +    return 0; +} + +static int allocate_boxes(struct zhaoxin_uncore_type **types, +             unsigned int package, unsigned int cpu) +{ +    struct zhaoxin_uncore_box *box, *tmp; +    struct zhaoxin_uncore_type *type; +    struct zhaoxin_uncore_pmu *pmu; +    LIST_HEAD(allocated); +    int i; + +    /* Try to allocate all required boxes */ +    for (; *types; types++) { +        type = *types; +        pmu = type->pmus; +        for (i = 0; i < type->num_boxes; i++, pmu++) { +            if (pmu->boxes[package]) +                continue; +            box = uncore_alloc_box(type, cpu_to_node(cpu)); +            if (!box) +                goto cleanup; +            box->pmu = pmu; +            box->package_id = package; +            list_add(&box->active_list, &allocated); +        } +    } +    /* Install them in the pmus */ +    list_for_each_entry_safe(box, tmp, &allocated, active_list) { +        list_del_init(&box->active_list); +        box->pmu->boxes[package] = box; +    } +    return 0; + +cleanup: +    list_for_each_entry_safe(box, tmp, &allocated, active_list) { +        list_del_init(&box->active_list); +        kfree(box); +    } +    return -ENOMEM; +} + +static int uncore_box_ref(struct zhaoxin_uncore_type **types, +              int id, unsigned int cpu) +{ +    struct zhaoxin_uncore_type *type; +    struct zhaoxin_uncore_pmu *pmu; +    struct zhaoxin_uncore_box *box; +    int i, ret; + +    ret = allocate_boxes(types, id, cpu); +    if (ret) +        return ret; + +    for (; *types; types++) { +        type = *types; +        pmu = type->pmus; +        for (i = 0; i < type->num_boxes; i++, pmu++) { +            box = pmu->boxes[id]; +            if (box && atomic_inc_return(&box->refcnt) == 1) +                uncore_box_init(box); +        } +    } +    return 0; +} + +static int uncore_event_cpu_online(unsigned int cpu) +{ +    int package, target, msr_ret; + +    package = topology_logical_package_id(cpu); +    msr_ret = uncore_box_ref(uncore_msr_uncores, package, cpu); + +    if (msr_ret) +        return -ENOMEM; + +    /* +     * Check if there is an online cpu in the package +     * which collects uncore events already. +     */ +    target = cpumask_any_and(&uncore_cpu_mask, topology_core_cpumask(cpu)); +    if (target < nr_cpu_ids) +        return 0; + +    cpumask_set_cpu(cpu, &uncore_cpu_mask); + +    if (!msr_ret) +        uncore_change_context(uncore_msr_uncores, -1, cpu); + +    return 0; +} + +static int uncore_pmu_register(struct zhaoxin_uncore_pmu *pmu) +{ +    int ret; + +    if (!pmu->type->pmu) { +        pmu->pmu = (struct pmu) { +            .attr_groups    = pmu->type->attr_groups, +            .task_ctx_nr    = perf_invalid_context, +            .event_init    = uncore_pmu_event_init, +            .add        = uncore_pmu_event_add, +            .del        = uncore_pmu_event_del, +            .start        = uncore_pmu_event_start, +            .stop        = uncore_pmu_event_stop, +            .read        = uncore_pmu_event_read, +            .module        = THIS_MODULE, +        }; +    } else { +        pmu->pmu = *pmu->type->pmu; +        pmu->pmu.attr_groups = pmu->type->attr_groups; +    } + +    if (pmu->type->num_boxes == 1) { +        if (strlen(pmu->type->name) > 0) +            sprintf(pmu->name, "uncore_%s", pmu->type->name); +        else +            sprintf(pmu->name, "uncore"); +    } else { +        sprintf(pmu->name, "uncore_%s_%d", pmu->type->name, +            pmu->pmu_idx); +    } + +    ret = perf_pmu_register(&pmu->pmu, pmu->name, -1); +    if (!ret) +        pmu->registered = true; +    return ret; +} + +static int __init type_pmu_register(struct zhaoxin_uncore_type *type) +{ +    int i, ret; + +    for (i = 0; i < type->num_boxes; i++) { +        ret = uncore_pmu_register(&type->pmus[i]); +        if (ret) +            return ret; +    } +    return 0; +} + +static int __init uncore_msr_pmus_register(void) +{ +    struct zhaoxin_uncore_type **types = uncore_msr_uncores; +    int ret; + +    for (; *types; types++) { +        ret = type_pmu_register(*types); +        if (ret) +            return ret; +    } +    return 0; +} + +static int __init uncore_cpu_init(void) +{ +    int ret; + +    ret = uncore_types_init(uncore_msr_uncores, true); +    if (ret) +        goto err; + +    ret = uncore_msr_pmus_register(); +    if (ret) +        goto err; +    return 0; +err: +    uncore_types_exit(uncore_msr_uncores); +    uncore_msr_uncores = empty_uncore; +    return ret; +} + + +#define CENTAUR_UNCORE_MODEL_MATCH(model, init)    \ +    { X86_VENDOR_CENTAUR, 7, model, X86_FEATURE_ANY, (unsigned long)&init } + +#define ZHAOXIN_UNCORE_MODEL_MATCH(model, init)    \ +    { X86_VENDOR_ZHAOXIN, 7, model, X86_FEATURE_ANY, (unsigned long)&init } + +struct zhaoxin_uncore_init_fun { +    void    (*cpu_init)(void); +}; + +void chx_uncore_cpu_init(void) +{ +    uncore_msr_uncores = chx_msr_uncores; +} + +static const struct zhaoxin_uncore_init_fun chx_uncore_init __initconst = { +    .cpu_init = chx_uncore_cpu_init, +}; + +static const struct x86_cpu_id zhaoxin_uncore_match[] __initconst = { +    CENTAUR_UNCORE_MODEL_MATCH(ZHAOXIN_FAM7_CHX001, chx_uncore_init), +    CENTAUR_UNCORE_MODEL_MATCH(ZHAOXIN_FAM7_CHX002, chx_uncore_init), +    ZHAOXIN_UNCORE_MODEL_MATCH(ZHAOXIN_FAM7_CHX001, chx_uncore_init), +    ZHAOXIN_UNCORE_MODEL_MATCH(ZHAOXIN_FAM7_CHX002, chx_uncore_init), +    {}, +}; + +MODULE_DEVICE_TABLE(x86cpu, zhaoxin_uncore_match); + +static int __init zhaoxin_uncore_init(void) +{ +    const struct x86_cpu_id *id; +    struct zhaoxin_uncore_init_fun *uncore_init; +    int cret = 0, ret; + +    id = x86_match_cpu(zhaoxin_uncore_match); + +    if (!id) +        return -ENODEV; + +    if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) +        return -ENODEV; + +    max_packages = topology_max_packages(); + +    pr_info("welcome to uncore!\n"); + +    uncore_init = (struct zhaoxin_uncore_init_fun *)id->driver_data; + +    if (uncore_init->cpu_init) { +        uncore_init->cpu_init(); +        cret = uncore_cpu_init(); +    } + +    if (cret) +        return -ENODEV; + +    ret = cpuhp_setup_state(CPUHP_AP_PERF_X86_UNCORE_ONLINE, +                "perf/x86/zhaoxin/uncore:online", +                uncore_event_cpu_online, +                uncore_event_cpu_offline); +    if (ret) +        goto err; + +    pr_info("uncore init success!\n"); +    return 0; +err: +    uncore_types_exit(uncore_msr_uncores); +    return ret; +} +module_init(zhaoxin_uncore_init); + +static void __exit zhaoxin_uncore_exit(void) +{ +    cpuhp_remove_state(CPUHP_AP_PERF_X86_UNCORE_ONLINE); +    uncore_types_exit(uncore_msr_uncores); +} +module_exit(zhaoxin_uncore_exit); + +MODULE_LICENSE("GPL"); diff --git a/arch/x86/events/zhaoxin/uncore.h b/arch/x86/events/zhaoxin/uncore.h new file mode 100644 index 000000000000..3521123dc95d --- /dev/null +++ b/arch/x86/events/zhaoxin/uncore.h @@ -0,0 +1,308 @@ +// SPDX-License-Identifier: GPL-2.0-only +#include <linux/slab.h> +#include <linux/pci.h> +#include <asm/apicdef.h> +#include <linux/io-64-nonatomic-lo-hi.h> + +#include <linux/perf_event.h> +#include "../perf_event.h" + +#define ZHAOXIN_FAM7_CHX001        0x1b +#define ZHAOXIN_FAM7_CHX002        0x3b + +#define UNCORE_PMU_NAME_LEN        32 +#define UNCORE_PMU_HRTIMER_INTERVAL    (60LL * NSEC_PER_SEC) +#define UNCORE_CHX_IMC_HRTIMER_INTERVAL (5ULL * NSEC_PER_SEC) + + +#define UNCORE_FIXED_EVENT              0xff +#define UNCORE_PMC_IDX_MAX_GENERIC      4 +#define UNCORE_PMC_IDX_MAX_FIXED        1 +#define UNCORE_PMC_IDX_FIXED UNCORE_PMC_IDX_MAX_GENERIC + +#define UNCORE_PMC_IDX_MAX              (UNCORE_PMC_IDX_FIXED + 1) + +struct zhaoxin_uncore_ops; +struct zhaoxin_uncore_pmu; +struct zhaoxin_uncore_box; +struct uncore_event_desc; + +struct zhaoxin_uncore_type { +    const char *name; +    int num_counters; +    int num_boxes; +    int perf_ctr_bits; +    int fixed_ctr_bits; +    unsigned perf_ctr; +    unsigned event_ctl; +    unsigned event_mask; +    unsigned event_mask_ext; +    unsigned fixed_ctr; +    unsigned fixed_ctl; +    unsigned box_ctl; +    unsigned msr_offset; +    unsigned num_shared_regs:8; +    unsigned single_fixed:1; +    unsigned pair_ctr_ctl:1; +    unsigned *msr_offsets; +    struct event_constraint unconstrainted; +    struct event_constraint *constraints; +    struct zhaoxin_uncore_pmu *pmus; +    struct zhaoxin_uncore_ops *ops; +    struct uncore_event_desc *event_descs; +    const struct attribute_group *attr_groups[4]; +    struct pmu *pmu; /* for custom pmu ops */ +}; + +#define pmu_group attr_groups[0] +#define format_group attr_groups[1] +#define events_group attr_groups[2] + +struct zhaoxin_uncore_ops { +    void (*init_box)(struct zhaoxin_uncore_box *); +    void (*exit_box)(struct zhaoxin_uncore_box *); +    void (*disable_box)(struct zhaoxin_uncore_box *); +    void (*enable_box)(struct zhaoxin_uncore_box *); +    void (*disable_event)(struct zhaoxin_uncore_box *, struct perf_event *); +    void (*enable_event)(struct zhaoxin_uncore_box *, struct perf_event *); +    u64 (*read_counter)(struct zhaoxin_uncore_box *, struct perf_event *); +    int (*hw_config)(struct zhaoxin_uncore_box *, struct perf_event *); +    struct event_constraint *(*get_constraint)(struct zhaoxin_uncore_box *, +                           struct perf_event *); +    void (*put_constraint)(struct zhaoxin_uncore_box *, struct perf_event *); +}; + +struct zhaoxin_uncore_pmu { +    struct pmu            pmu; +    char                name[UNCORE_PMU_NAME_LEN]; +    int                pmu_idx; +    int                func_id; +    bool                registered; +    atomic_t            activeboxes; +    struct zhaoxin_uncore_type    *type; +    struct zhaoxin_uncore_box    **boxes; +}; + +struct zhaoxin_uncore_extra_reg { +    raw_spinlock_t lock; +    u64 config, config1, config2; +    atomic_t ref; +}; + +struct zhaoxin_uncore_box { +    int pci_phys_id; +    int package_id;    /*Package ID */ +    int n_active;    /* number of active events */ +    int n_events; +    int cpu;    /* cpu to collect events */ +    unsigned long flags; +    atomic_t refcnt; +    struct perf_event *events[UNCORE_PMC_IDX_MAX]; +    struct perf_event *event_list[UNCORE_PMC_IDX_MAX]; +    struct event_constraint *event_constraint[UNCORE_PMC_IDX_MAX]; +    unsigned long active_mask[BITS_TO_LONGS(UNCORE_PMC_IDX_MAX)]; +    u64 tags[UNCORE_PMC_IDX_MAX]; +    struct pci_dev *pci_dev; +    struct zhaoxin_uncore_pmu *pmu; +    u64 hrtimer_duration; /* hrtimer timeout for this box */ +    struct hrtimer hrtimer; +    struct list_head list; +    struct list_head active_list; +    void __iomem *io_addr; +    struct zhaoxin_uncore_extra_reg shared_regs[0]; +}; + +#define UNCORE_BOX_FLAG_INITIATED    0 + +struct uncore_event_desc { +    struct kobj_attribute attr; +    const char *config; +}; + +ssize_t zx_uncore_event_show(struct kobject *kobj, +            struct kobj_attribute *attr, char *buf); + +#define ZHAOXIN_UNCORE_EVENT_DESC(_name, _config)            \ +{                                \ +    .attr    = __ATTR(_name, 0444, zx_uncore_event_show, NULL),    \ +    .config    = _config,                    \ +} + +#define DEFINE_UNCORE_FORMAT_ATTR(_var, _name, _format)     \ +static ssize_t __uncore_##_var##_show(struct kobject *kobj,     \ +                struct kobj_attribute *attr,        \ +                char *page)                \ +{                                    \ +    BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);            \ +    return sprintf(page, _format "\n");                \ +}                                    \ +static struct kobj_attribute format_attr_##_var =            \ +    __ATTR(_name, 0444, __uncore_##_var##_show, NULL) + +static inline bool uncore_pmc_fixed(int idx) +{ +    return idx == UNCORE_PMC_IDX_FIXED; +} + +static inline unsigned uncore_msr_box_offset(struct zhaoxin_uncore_box *box) +{ +    struct zhaoxin_uncore_pmu *pmu = box->pmu; + +    return pmu->type->msr_offsets ? +        pmu->type->msr_offsets[pmu->pmu_idx] : +        pmu->type->msr_offset * pmu->pmu_idx; +} + +static inline unsigned uncore_msr_box_ctl(struct zhaoxin_uncore_box *box) +{ +    if (!box->pmu->type->box_ctl) +        return 0; +    return box->pmu->type->box_ctl + uncore_msr_box_offset(box); +} + +static inline unsigned uncore_msr_fixed_ctl(struct zhaoxin_uncore_box *box) +{ +    if (!box->pmu->type->fixed_ctl) +        return 0; +    return box->pmu->type->fixed_ctl + uncore_msr_box_offset(box); +} + +static inline unsigned uncore_msr_fixed_ctr(struct zhaoxin_uncore_box *box) +{ +    return box->pmu->type->fixed_ctr + uncore_msr_box_offset(box); +} + +static inline +unsigned uncore_msr_event_ctl(struct zhaoxin_uncore_box *box, int idx) +{ +    return box->pmu->type->event_ctl + +        (box->pmu->type->pair_ctr_ctl ? 2 * idx : idx) + +        uncore_msr_box_offset(box); +} + +static inline +unsigned uncore_msr_perf_ctr(struct zhaoxin_uncore_box *box, int idx) +{ +    return box->pmu->type->perf_ctr + +        (box->pmu->type->pair_ctr_ctl ? 2 * idx : idx) + +        uncore_msr_box_offset(box); +} + +static inline +unsigned uncore_fixed_ctl(struct zhaoxin_uncore_box *box) +{ +    return uncore_msr_fixed_ctl(box); +} + +static inline +unsigned uncore_fixed_ctr(struct zhaoxin_uncore_box *box) +{ +    return uncore_msr_fixed_ctr(box); +} + +static inline +unsigned uncore_event_ctl(struct zhaoxin_uncore_box *box, int idx) +{ +    return uncore_msr_event_ctl(box, idx); +} + +static inline +unsigned uncore_perf_ctr(struct zhaoxin_uncore_box *box, int idx) +{ +    return uncore_msr_perf_ctr(box, idx); +} + +static inline int uncore_perf_ctr_bits(struct zhaoxin_uncore_box *box) +{ +    return box->pmu->type->perf_ctr_bits; +} + +static inline int uncore_fixed_ctr_bits(struct zhaoxin_uncore_box *box) +{ +    return box->pmu->type->fixed_ctr_bits; +} + +static inline int uncore_num_counters(struct zhaoxin_uncore_box *box) +{ +    return box->pmu->type->num_counters; +} + +static inline void uncore_disable_box(struct zhaoxin_uncore_box *box) +{ +    if (box->pmu->type->ops->disable_box) +        box->pmu->type->ops->disable_box(box); +} + +static inline void uncore_enable_box(struct zhaoxin_uncore_box *box) +{ +    if (box->pmu->type->ops->enable_box) +        box->pmu->type->ops->enable_box(box); +} + +static inline void uncore_disable_event(struct zhaoxin_uncore_box *box, +                struct perf_event *event) +{ +    box->pmu->type->ops->disable_event(box, event); +} + +static inline void uncore_enable_event(struct zhaoxin_uncore_box *box, +                struct perf_event *event) +{ +    box->pmu->type->ops->enable_event(box, event); +} + +static inline u64 uncore_read_counter(struct zhaoxin_uncore_box *box, +                struct perf_event *event) +{ +    return box->pmu->type->ops->read_counter(box, event); +} + +static inline void uncore_box_init(struct zhaoxin_uncore_box *box) +{ +    if (!test_and_set_bit(UNCORE_BOX_FLAG_INITIATED, &box->flags)) { +        if (box->pmu->type->ops->init_box) +            box->pmu->type->ops->init_box(box); +    } +} + +static inline void uncore_box_exit(struct zhaoxin_uncore_box *box) +{ +    if (test_and_clear_bit(UNCORE_BOX_FLAG_INITIATED, &box->flags)) { +        if (box->pmu->type->ops->exit_box) +            box->pmu->type->ops->exit_box(box); +    } +} + +static inline bool uncore_box_is_fake(struct zhaoxin_uncore_box *box) +{ +    return (box->package_id < 0); +} + +static inline struct zhaoxin_uncore_pmu *uncore_event_to_pmu(struct perf_event *event) +{ +    return container_of(event->pmu, struct zhaoxin_uncore_pmu, pmu); +} + +static inline struct zhaoxin_uncore_box *uncore_event_to_box(struct perf_event *event) +{ +    return event->pmu_private; +} + + +static struct zhaoxin_uncore_box *uncore_pmu_to_box(struct zhaoxin_uncore_pmu *pmu, int cpu); +static u64 uncore_msr_read_counter(struct zhaoxin_uncore_box *box, struct perf_event *event); + +static void uncore_pmu_start_hrtimer(struct zhaoxin_uncore_box *box); +static void uncore_pmu_cancel_hrtimer(struct zhaoxin_uncore_box *box); +static void uncore_pmu_event_start(struct perf_event *event, int flags); +static void uncore_pmu_event_stop(struct perf_event *event, int flags); +static int uncore_pmu_event_add(struct perf_event *event, int flags); +static void uncore_pmu_event_del(struct perf_event *event, int flags); +static void uncore_pmu_event_read(struct perf_event *event); +static void uncore_perf_event_update(struct zhaoxin_uncore_box *box, struct perf_event *event); +struct event_constraint * +uncore_get_constraint(struct zhaoxin_uncore_box *box, struct perf_event *event); +void uncore_put_constraint(struct zhaoxin_uncore_box *box, struct perf_event *event); +u64 uncore_shared_reg_config(struct zhaoxin_uncore_box *box, int idx); + +void chx_uncore_cpu_init(void); diff --git a/arch/x86/kernel/cpu/perfctr-watchdog.c b/arch/x86/kernel/cpu/perfctr-watchdog.c index 9556930cd8c1..a548d9104604 100644 --- a/arch/x86/kernel/cpu/perfctr-watchdog.c +++ b/arch/x86/kernel/cpu/perfctr-watchdog.c @@ -63,6 +63,10 @@ static inline unsigned int nmi_perfctr_msr_to_bit(unsigned int msr)          case 15:              return msr - MSR_P4_BPU_PERFCTR0;          } +        break; +    case X86_VENDOR_ZHAOXIN: +    case X86_VENDOR_CENTAUR: +        return msr - MSR_ARCH_PERFMON_PERFCTR0;      }      return 0;  } @@ -92,6 +96,10 @@ static inline unsigned int nmi_evntsel_msr_to_bit(unsigned int msr)          case 15:              return msr - MSR_P4_BSU_ESCR0;          } +        break; +    case X86_VENDOR_ZHAOXIN: +    case X86_VENDOR_CENTAUR: +        return msr - MSR_ARCH_PERFMON_EVENTSEL0;      }      return 0;