Quelle  apple_m1_cpu_pmu.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * CPU PMU driver for the Apple M1 and derivatives
 *
 * Copyright (C) 2021 Google LLC
 *
 * Author: Marc Zyngier <maz@kernel.org>
 *
 * Most of the information used in this driver was provided by the
 * Asahi Linux project. The rest was experimentally discovered.
 */

#include <linux/of.h>
#include <linux/perf/arm_pmu.h>
#include <linux/perf/arm_pmuv3.h>
#include <linux/platform_device.h>

#include <asm/apple_m1_pmu.h>
#include <asm/irq_regs.h>
#include <asm/perf_event.h>

#define M1_PMU_NR_COUNTERS  10

#define M1_PMU_CFG_EVENT  GENMASK(7, 0)

#define ANY_BUT_0_1   GENMASK(9, 2)
#define ONLY_2_TO_7   GENMASK(7, 2)
#define ONLY_2_4_6   (BIT(2) | BIT(4) | BIT(6))
#define ONLY_5_6_7   (BIT(5) | BIT(6) | BIT(7))

/*
 * Description of the events we actually know about, as well as those with
 * a specific counter affinity. Yes, this is a grand total of two known
 * counters, and the rest is anybody's guess.
 *
 * Not all counters can count all events. Counters #0 and #1 are wired to
 * count cycles and instructions respectively, and some events have
 * bizarre mappings (every other counter, or even *one* counter). These
 * restrictions equally apply to both P and E cores.
 *
 * It is worth noting that the PMUs attached to P and E cores are likely
 * to be different because the underlying uarches are different. At the
 * moment, we don't really need to distinguish between the two because we
 * know next to nothing about the events themselves, and we already have
 * per cpu-type PMU abstractions.
 *
 * If we eventually find out that the events are different across
 * implementations, we'll have to introduce per cpu-type tables.
 */
enum m1_pmu_events {
 M1_PMU_PERFCTR_RETIRE_UOP    = 0x1,
 M1_PMU_PERFCTR_CORE_ACTIVE_CYCLE   = 0x2,
 M1_PMU_PERFCTR_L1I_TLB_FILL    = 0x4,
 M1_PMU_PERFCTR_L1D_TLB_FILL    = 0x5,
 M1_PMU_PERFCTR_MMU_TABLE_WALK_INSTRUCTION  = 0x7,
 M1_PMU_PERFCTR_MMU_TABLE_WALK_DATA   = 0x8,
 M1_PMU_PERFCTR_L2_TLB_MISS_INSTRUCTION   = 0xa,
 M1_PMU_PERFCTR_L2_TLB_MISS_DATA    = 0xb,
 M1_PMU_PERFCTR_MMU_VIRTUAL_MEMORY_FAULT_NONSPEC  = 0xd,
 M1_PMU_PERFCTR_SCHEDULE_UOP    = 0x52,
 M1_PMU_PERFCTR_INTERRUPT_PENDING   = 0x6c,
 M1_PMU_PERFCTR_MAP_STALL_DISPATCH   = 0x70,
 M1_PMU_PERFCTR_MAP_REWIND    = 0x75,
 M1_PMU_PERFCTR_MAP_STALL    = 0x76,
 M1_PMU_PERFCTR_MAP_INT_UOP    = 0x7c,
 M1_PMU_PERFCTR_MAP_LDST_UOP    = 0x7d,
 M1_PMU_PERFCTR_MAP_SIMD_UOP    = 0x7e,
 M1_PMU_PERFCTR_FLUSH_RESTART_OTHER_NONSPEC  = 0x84,
 M1_PMU_PERFCTR_INST_ALL     = 0x8c,
 M1_PMU_PERFCTR_INST_BRANCH    = 0x8d,
 M1_PMU_PERFCTR_INST_BRANCH_CALL    = 0x8e,
 M1_PMU_PERFCTR_INST_BRANCH_RET    = 0x8f,
 M1_PMU_PERFCTR_INST_BRANCH_TAKEN   = 0x90,
 M1_PMU_PERFCTR_INST_BRANCH_INDIR   = 0x93,
 M1_PMU_PERFCTR_INST_BRANCH_COND    = 0x94,
 M1_PMU_PERFCTR_INST_INT_LD    = 0x95,
 M1_PMU_PERFCTR_INST_INT_ST    = 0x96,
 M1_PMU_PERFCTR_INST_INT_ALU    = 0x97,
 M1_PMU_PERFCTR_INST_SIMD_LD    = 0x98,
 M1_PMU_PERFCTR_INST_SIMD_ST    = 0x99,
 M1_PMU_PERFCTR_INST_SIMD_ALU    = 0x9a,
 M1_PMU_PERFCTR_INST_LDST    = 0x9b,
 M1_PMU_PERFCTR_INST_BARRIER    = 0x9c,
 M1_PMU_PERFCTR_UNKNOWN_9f    = 0x9f,
 M1_PMU_PERFCTR_L1D_TLB_ACCESS    = 0xa0,
 M1_PMU_PERFCTR_L1D_TLB_MISS    = 0xa1,
 M1_PMU_PERFCTR_L1D_CACHE_MISS_ST   = 0xa2,
 M1_PMU_PERFCTR_L1D_CACHE_MISS_LD   = 0xa3,
 M1_PMU_PERFCTR_LD_UNIT_UOP    = 0xa6,
 M1_PMU_PERFCTR_ST_UNIT_UOP    = 0xa7,
 M1_PMU_PERFCTR_L1D_CACHE_WRITEBACK   = 0xa8,
 M1_PMU_PERFCTR_LDST_X64_UOP    = 0xb1,
 M1_PMU_PERFCTR_LDST_XPG_UOP    = 0xb2,
 M1_PMU_PERFCTR_ATOMIC_OR_EXCLUSIVE_SUCC   = 0xb3,
 M1_PMU_PERFCTR_ATOMIC_OR_EXCLUSIVE_FAIL   = 0xb4,
 M1_PMU_PERFCTR_L1D_CACHE_MISS_LD_NONSPEC  = 0xbf,
 M1_PMU_PERFCTR_L1D_CACHE_MISS_ST_NONSPEC  = 0xc0,
 M1_PMU_PERFCTR_L1D_TLB_MISS_NONSPEC   = 0xc1,
 M1_PMU_PERFCTR_ST_MEMORY_ORDER_VIOLATION_NONSPEC = 0xc4,
 M1_PMU_PERFCTR_BRANCH_COND_MISPRED_NONSPEC  = 0xc5,
 M1_PMU_PERFCTR_BRANCH_INDIR_MISPRED_NONSPEC  = 0xc6,
 M1_PMU_PERFCTR_BRANCH_RET_INDIR_MISPRED_NONSPEC  = 0xc8,
 M1_PMU_PERFCTR_BRANCH_CALL_INDIR_MISPRED_NONSPEC = 0xca,
 M1_PMU_PERFCTR_BRANCH_MISPRED_NONSPEC   = 0xcb,
 M1_PMU_PERFCTR_L1I_TLB_MISS_DEMAND   = 0xd4,
 M1_PMU_PERFCTR_MAP_DISPATCH_BUBBLE   = 0xd6,
 M1_PMU_PERFCTR_L1I_CACHE_MISS_DEMAND   = 0xdb,
 M1_PMU_PERFCTR_FETCH_RESTART    = 0xde,
 M1_PMU_PERFCTR_ST_NT_UOP    = 0xe5,
 M1_PMU_PERFCTR_LD_NT_UOP    = 0xe6,
 M1_PMU_PERFCTR_UNKNOWN_f5    = 0xf5,
 M1_PMU_PERFCTR_UNKNOWN_f6    = 0xf6,
 M1_PMU_PERFCTR_UNKNOWN_f7    = 0xf7,
 M1_PMU_PERFCTR_UNKNOWN_f8    = 0xf8,
 M1_PMU_PERFCTR_UNKNOWN_fd    = 0xfd,
 M1_PMU_PERFCTR_LAST     = M1_PMU_CFG_EVENT,

 /*
 * From this point onwards, these are not actual HW events,
 * but attributes that get stored in hw->config_base.
 */
 M1_PMU_CFG_COUNT_USER     = BIT(8),
 M1_PMU_CFG_COUNT_KERNEL     = BIT(9),
 M1_PMU_CFG_COUNT_HOST     = BIT(10),
 M1_PMU_CFG_COUNT_GUEST     = BIT(11),
};

/*
 * Per-event affinity table. Most events can be installed on counter
 * 2-9, but there are a number of exceptions. Note that this table
 * has been created experimentally, and I wouldn't be surprised if more
 * counters had strange affinities.
 */
static const u16 m1_pmu_event_affinity[M1_PMU_PERFCTR_LAST + 1] = {
 [0 ... M1_PMU_PERFCTR_LAST]    = ANY_BUT_0_1,
 [M1_PMU_PERFCTR_RETIRE_UOP]    = BIT(7),
 [M1_PMU_PERFCTR_CORE_ACTIVE_CYCLE]   = ANY_BUT_0_1 | BIT(0),
 [M1_PMU_PERFCTR_INST_ALL]    = BIT(7) | BIT(1),
 [M1_PMU_PERFCTR_INST_BRANCH]    = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_BRANCH_CALL]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_BRANCH_RET]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_BRANCH_TAKEN]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_BRANCH_INDIR]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_BRANCH_COND]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_INT_LD]    = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_INT_ST]    = BIT(7),
 [M1_PMU_PERFCTR_INST_INT_ALU]    = BIT(7),
 [M1_PMU_PERFCTR_INST_SIMD_LD]    = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_SIMD_ST]    = ONLY_5_6_7,
 [M1_PMU_PERFCTR_INST_SIMD_ALU]    = BIT(7),
 [M1_PMU_PERFCTR_INST_LDST]    = BIT(7),
 [M1_PMU_PERFCTR_INST_BARRIER]    = ONLY_5_6_7,
 [M1_PMU_PERFCTR_UNKNOWN_9f]    = BIT(7),
 [M1_PMU_PERFCTR_L1D_CACHE_MISS_LD_NONSPEC]  = ONLY_5_6_7,
 [M1_PMU_PERFCTR_L1D_CACHE_MISS_ST_NONSPEC]  = ONLY_5_6_7,
 [M1_PMU_PERFCTR_L1D_TLB_MISS_NONSPEC]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_ST_MEMORY_ORDER_VIOLATION_NONSPEC] = ONLY_5_6_7,
 [M1_PMU_PERFCTR_BRANCH_COND_MISPRED_NONSPEC]  = ONLY_5_6_7,
 [M1_PMU_PERFCTR_BRANCH_INDIR_MISPRED_NONSPEC]  = ONLY_5_6_7,
 [M1_PMU_PERFCTR_BRANCH_RET_INDIR_MISPRED_NONSPEC] = ONLY_5_6_7,
 [M1_PMU_PERFCTR_BRANCH_CALL_INDIR_MISPRED_NONSPEC] = ONLY_5_6_7,
 [M1_PMU_PERFCTR_BRANCH_MISPRED_NONSPEC]   = ONLY_5_6_7,
 [M1_PMU_PERFCTR_UNKNOWN_f5]    = ONLY_2_4_6,
 [M1_PMU_PERFCTR_UNKNOWN_f6]    = ONLY_2_4_6,
 [M1_PMU_PERFCTR_UNKNOWN_f7]    = ONLY_2_4_6,
 [M1_PMU_PERFCTR_UNKNOWN_f8]    = ONLY_2_TO_7,
 [M1_PMU_PERFCTR_UNKNOWN_fd]    = ONLY_2_4_6,
};

static const unsigned m1_pmu_perf_map[PERF_COUNT_HW_MAX] = {
 PERF_MAP_ALL_UNSUPPORTED,
 [PERF_COUNT_HW_CPU_CYCLES]  = M1_PMU_PERFCTR_CORE_ACTIVE_CYCLE,
 [PERF_COUNT_HW_INSTRUCTIONS]  = M1_PMU_PERFCTR_INST_ALL,
 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = M1_PMU_PERFCTR_INST_BRANCH,
 [PERF_COUNT_HW_BRANCH_MISSES]  = M1_PMU_PERFCTR_BRANCH_MISPRED_NONSPEC,
};

#define M1_PMUV3_EVENT_MAP(pmuv3_event, m1_event)    \
 [ARMV8_PMUV3_PERFCTR_##pmuv3_event] = M1_PMU_PERFCTR_##m1_event

static const u16 m1_pmu_pmceid_map[ARMV8_PMUV3_MAX_COMMON_EVENTS] = {
 [0 ... ARMV8_PMUV3_MAX_COMMON_EVENTS - 1] = HW_OP_UNSUPPORTED,
 M1_PMUV3_EVENT_MAP(INST_RETIRED, INST_ALL),
 M1_PMUV3_EVENT_MAP(CPU_CYCLES,  CORE_ACTIVE_CYCLE),
 M1_PMUV3_EVENT_MAP(BR_RETIRED,  INST_BRANCH),
 M1_PMUV3_EVENT_MAP(BR_MIS_PRED_RETIRED, BRANCH_MISPRED_NONSPEC),
};

/* sysfs definitions */
static ssize_t m1_pmu_events_sysfs_show(struct device *dev,
     struct device_attribute *attr,
     char *page)
{
 struct perf_pmu_events_attr *pmu_attr;

 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);

 return sprintf(page, "event=0x%04llx\n", pmu_attr->id);
}

#define M1_PMU_EVENT_ATTR(name, config)     \
 PMU_EVENT_ATTR_ID(name, m1_pmu_events_sysfs_show, config)

static struct attribute *m1_pmu_event_attrs[] = {
 M1_PMU_EVENT_ATTR(cycles, M1_PMU_PERFCTR_CORE_ACTIVE_CYCLE),
 M1_PMU_EVENT_ATTR(instructions, M1_PMU_PERFCTR_INST_ALL),
 NULL,
};

static const struct attribute_group m1_pmu_events_attr_group = {
 .name = "events",
 .attrs = m1_pmu_event_attrs,
};

PMU_FORMAT_ATTR(event, "config:0-7");

static struct attribute *m1_pmu_format_attrs[] = {
 &format_attr_event.attr,
 NULL,
};

static const struct attribute_group m1_pmu_format_attr_group = {
 .name = "format",
 .attrs = m1_pmu_format_attrs,
};

/* Low level accessors. No synchronisation. */
#define PMU_READ_COUNTER(_idx)      \
 case _idx: return read_sysreg_s(SYS_IMP_APL_PMC## _idx ##_EL1)

#define PMU_WRITE_COUNTER(_val, _idx)     \
 case _idx:       \
  write_sysreg_s(_val, SYS_IMP_APL_PMC## _idx ##_EL1); \
  return

static u64 m1_pmu_read_hw_counter(unsigned int index)
{
 switch (index) {
  PMU_READ_COUNTER(0);
  PMU_READ_COUNTER(1);
  PMU_READ_COUNTER(2);
  PMU_READ_COUNTER(3);
  PMU_READ_COUNTER(4);
  PMU_READ_COUNTER(5);
  PMU_READ_COUNTER(6);
  PMU_READ_COUNTER(7);
  PMU_READ_COUNTER(8);
  PMU_READ_COUNTER(9);
 }

 BUG();
}

static void m1_pmu_write_hw_counter(u64 val, unsigned int index)
{
 switch (index) {
  PMU_WRITE_COUNTER(val, 0);
  PMU_WRITE_COUNTER(val, 1);
  PMU_WRITE_COUNTER(val, 2);
  PMU_WRITE_COUNTER(val, 3);
  PMU_WRITE_COUNTER(val, 4);
  PMU_WRITE_COUNTER(val, 5);
  PMU_WRITE_COUNTER(val, 6);
  PMU_WRITE_COUNTER(val, 7);
  PMU_WRITE_COUNTER(val, 8);
  PMU_WRITE_COUNTER(val, 9);
 }

 BUG();
}

#define get_bit_offset(index, mask) (__ffs(mask) + (index))

static void __m1_pmu_enable_counter(unsigned int index, bool en)
{
 u64 val, bit;

 switch (index) {
 case 0 ... 7:
  bit = BIT(get_bit_offset(index, PMCR0_CNT_ENABLE_0_7));
  break;
 case 8 ... 9:
  bit = BIT(get_bit_offset(index - 8, PMCR0_CNT_ENABLE_8_9));
  break;
 default:
  BUG();
 }

 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);

 if (en)
  val |= bit;
 else
  val &= ~bit;

 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
}

static void m1_pmu_enable_counter(unsigned int index)
{
 __m1_pmu_enable_counter(index, true);
}

static void m1_pmu_disable_counter(unsigned int index)
{
 __m1_pmu_enable_counter(index, false);
}

static void __m1_pmu_enable_counter_interrupt(unsigned int index, bool en)
{
 u64 val, bit;

 switch (index) {
 case 0 ... 7:
  bit = BIT(get_bit_offset(index, PMCR0_PMI_ENABLE_0_7));
  break;
 case 8 ... 9:
  bit = BIT(get_bit_offset(index - 8, PMCR0_PMI_ENABLE_8_9));
  break;
 default:
  BUG();
 }

 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);

 if (en)
  val |= bit;
 else
  val &= ~bit;

 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
}

static void m1_pmu_enable_counter_interrupt(unsigned int index)
{
 __m1_pmu_enable_counter_interrupt(index, true);
}

static void m1_pmu_disable_counter_interrupt(unsigned int index)
{
 __m1_pmu_enable_counter_interrupt(index, false);
}

static void __m1_pmu_configure_event_filter(unsigned int index, bool user,
         bool kernel, bool host)
{
 u64 clear, set, user_bit, kernel_bit;

 switch (index) {
 case 0 ... 7:
  user_bit = BIT(get_bit_offset(index, PMCR1_COUNT_A64_EL0_0_7));
  kernel_bit = BIT(get_bit_offset(index, PMCR1_COUNT_A64_EL1_0_7));
  break;
 case 8 ... 9:
  user_bit = BIT(get_bit_offset(index - 8, PMCR1_COUNT_A64_EL0_8_9));
  kernel_bit = BIT(get_bit_offset(index - 8, PMCR1_COUNT_A64_EL1_8_9));
  break;
 default:
  BUG();
 }

 clear = set = 0;
 if (user)
  set |= user_bit;
 else
  clear |= user_bit;

 if (kernel)
  set |= kernel_bit;
 else
  clear |= kernel_bit;

 if (host)
  sysreg_clear_set_s(SYS_IMP_APL_PMCR1_EL1, clear, set);
 else if (is_kernel_in_hyp_mode())
  sysreg_clear_set_s(SYS_IMP_APL_PMCR1_EL12, clear, set);
}

static void __m1_pmu_configure_eventsel(unsigned int index, u8 event)
{
 u64 clear = 0, set = 0;
 int shift;

 /*
 * Counters 0 and 1 have fixed events. For anything else,
 * place the event at the expected location in the relevant
 * register (PMESR0 holds the event configuration for counters
 * 2-5, resp. PMESR1 for counters 6-9).
 */
 switch (index) {
 case 0 ... 1:
  break;
 case 2 ... 5:
  shift = (index - 2) * 8;
  clear |= (u64)0xff << shift;
  set |= (u64)event << shift;
  sysreg_clear_set_s(SYS_IMP_APL_PMESR0_EL1, clear, set);
  break;
 case 6 ... 9:
  shift = (index - 6) * 8;
  clear |= (u64)0xff << shift;
  set |= (u64)event << shift;
  sysreg_clear_set_s(SYS_IMP_APL_PMESR1_EL1, clear, set);
  break;
 }
}

static void m1_pmu_configure_counter(unsigned int index, unsigned long config_base)
{
 bool kernel = config_base & M1_PMU_CFG_COUNT_KERNEL;
 bool guest = config_base & M1_PMU_CFG_COUNT_GUEST;
 bool host = config_base & M1_PMU_CFG_COUNT_HOST;
 bool user = config_base & M1_PMU_CFG_COUNT_USER;
 u8 evt = config_base & M1_PMU_CFG_EVENT;

 __m1_pmu_configure_event_filter(index, user && host, kernel && host, true);
 __m1_pmu_configure_event_filter(index, user && guest, kernel && guest, false);
 __m1_pmu_configure_eventsel(index, evt);
}

/* arm_pmu backend */
static void m1_pmu_enable_event(struct perf_event *event)
{
 bool user, kernel;
 u8 evt;

 evt = event->hw.config_base & M1_PMU_CFG_EVENT;
 user = event->hw.config_base & M1_PMU_CFG_COUNT_USER;
 kernel = event->hw.config_base & M1_PMU_CFG_COUNT_KERNEL;

 m1_pmu_configure_counter(event->hw.idx, event->hw.config_base);
 m1_pmu_enable_counter(event->hw.idx);
 m1_pmu_enable_counter_interrupt(event->hw.idx);
 isb();
}

static void m1_pmu_disable_event(struct perf_event *event)
{
 m1_pmu_disable_counter_interrupt(event->hw.idx);
 m1_pmu_disable_counter(event->hw.idx);
 isb();
}

static irqreturn_t m1_pmu_handle_irq(struct arm_pmu *cpu_pmu)
{
 struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
 struct pt_regs *regs;
 u64 overflow, state;
 int idx;

 overflow = read_sysreg_s(SYS_IMP_APL_PMSR_EL1);
 if (!overflow) {
  /* Spurious interrupt? */
  state = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
  state &= ~PMCR0_IACT;
  write_sysreg_s(state, SYS_IMP_APL_PMCR0_EL1);
  isb();
  return IRQ_NONE;
 }

 cpu_pmu->stop(cpu_pmu);

 regs = get_irq_regs();

 for_each_set_bit(idx, cpu_pmu->cntr_mask, M1_PMU_NR_COUNTERS) {
  struct perf_event *event = cpuc->events[idx];
  struct perf_sample_data data;

  if (!event)
   continue;

  armpmu_event_update(event);
  perf_sample_data_init(&data, 0, event->hw.last_period);
  if (!armpmu_event_set_period(event))
   continue;

  perf_event_overflow(event, &data, regs);
 }

 cpu_pmu->start(cpu_pmu);

 return IRQ_HANDLED;
}

static u64 m1_pmu_read_counter(struct perf_event *event)
{
 return m1_pmu_read_hw_counter(event->hw.idx);
}

static void m1_pmu_write_counter(struct perf_event *event, u64 value)
{
 m1_pmu_write_hw_counter(value, event->hw.idx);
 isb();
}

static int m1_pmu_get_event_idx(struct pmu_hw_events *cpuc,
    struct perf_event *event)
{
 unsigned long evtype = event->hw.config_base & M1_PMU_CFG_EVENT;
 unsigned long affinity = m1_pmu_event_affinity[evtype];
 int idx;

 /*
 * Place the event on the first free counter that can count
 * this event.
 *
 * We could do a better job if we had a view of all the events
 * counting on the PMU at any given time, and by placing the
 * most constraining events first.
 */
 for_each_set_bit(idx, &affinity, M1_PMU_NR_COUNTERS) {
  if (!test_and_set_bit(idx, cpuc->used_mask))
   return idx;
 }

 return -EAGAIN;
}

static void m1_pmu_clear_event_idx(struct pmu_hw_events *cpuc,
       struct perf_event *event)
{
 clear_bit(event->hw.idx, cpuc->used_mask);
}

static void __m1_pmu_set_mode(u8 mode)
{
 u64 val;

 val = read_sysreg_s(SYS_IMP_APL_PMCR0_EL1);
 val &= ~(PMCR0_IMODE | PMCR0_IACT);
 val |= FIELD_PREP(PMCR0_IMODE, mode);
 write_sysreg_s(val, SYS_IMP_APL_PMCR0_EL1);
 isb();
}

static void m1_pmu_start(struct arm_pmu *cpu_pmu)
{
 __m1_pmu_set_mode(PMCR0_IMODE_FIQ);
}

static void m1_pmu_stop(struct arm_pmu *cpu_pmu)
{
 __m1_pmu_set_mode(PMCR0_IMODE_OFF);
}

static int m1_pmu_map_event(struct perf_event *event)
{
 /*
 * Although the counters are 48bit wide, bit 47 is what
 * triggers the overflow interrupt. Advertise the counters
 * being 47bit wide to mimick the behaviour of the ARM PMU.
 */
 event->hw.flags |= ARMPMU_EVT_47BIT;
 return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
}

static int m2_pmu_map_event(struct perf_event *event)
{
 /*
 * Same deal as the above, except that M2 has 64bit counters.
 * Which, as far as we're concerned, actually means 63 bits.
 * Yes, this is getting awkward.
 */
 event->hw.flags |= ARMPMU_EVT_63BIT;
 return armpmu_map_event(event, &m1_pmu_perf_map, NULL, M1_PMU_CFG_EVENT);
}

static int m1_pmu_map_pmuv3_event(unsigned int eventsel)
{
 u16 m1_event = HW_OP_UNSUPPORTED;

 if (eventsel < ARMV8_PMUV3_MAX_COMMON_EVENTS)
  m1_event = m1_pmu_pmceid_map[eventsel];

 return m1_event == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : m1_event;
}

static void m1_pmu_init_pmceid(struct arm_pmu *pmu)
{
 unsigned int event;

 for (event = 0; event < ARMV8_PMUV3_MAX_COMMON_EVENTS; event++) {
  if (m1_pmu_map_pmuv3_event(event) >= 0)
   set_bit(event, pmu->pmceid_bitmap);
 }
}

static void m1_pmu_reset(void *info)
{
 int i;

 __m1_pmu_set_mode(PMCR0_IMODE_OFF);

 for (i = 0; i < M1_PMU_NR_COUNTERS; i++) {
  m1_pmu_disable_counter(i);
  m1_pmu_disable_counter_interrupt(i);
  m1_pmu_write_hw_counter(0, i);
 }

 isb();
}

static int m1_pmu_set_event_filter(struct hw_perf_event *event,
       struct perf_event_attr *attr)
{
 unsigned long config_base = 0;

 if (!attr->exclude_guest && !is_kernel_in_hyp_mode()) {
  pr_debug("ARM performance counters do not support mode exclusion\n");
  return -EOPNOTSUPP;
 }
 if (!attr->exclude_kernel)
  config_base |= M1_PMU_CFG_COUNT_KERNEL;
 if (!attr->exclude_user)
  config_base |= M1_PMU_CFG_COUNT_USER;
 if (!attr->exclude_host)
  config_base |= M1_PMU_CFG_COUNT_HOST;
 if (!attr->exclude_guest)
  config_base |= M1_PMU_CFG_COUNT_GUEST;

 event->config_base = config_base;

 return 0;
}

static int m1_pmu_init(struct arm_pmu *cpu_pmu, u32 flags)
{
 cpu_pmu->handle_irq   = m1_pmu_handle_irq;
 cpu_pmu->enable    = m1_pmu_enable_event;
 cpu_pmu->disable   = m1_pmu_disable_event;
 cpu_pmu->read_counter   = m1_pmu_read_counter;
 cpu_pmu->write_counter   = m1_pmu_write_counter;
 cpu_pmu->get_event_idx   = m1_pmu_get_event_idx;
 cpu_pmu->clear_event_idx  = m1_pmu_clear_event_idx;
 cpu_pmu->start    = m1_pmu_start;
 cpu_pmu->stop    = m1_pmu_stop;

 if (flags & ARMPMU_EVT_47BIT)
  cpu_pmu->map_event = m1_pmu_map_event;
 else if (flags & ARMPMU_EVT_63BIT)
  cpu_pmu->map_event = m2_pmu_map_event;
 else
  return WARN_ON(-EINVAL);

 cpu_pmu->reset    = m1_pmu_reset;
 cpu_pmu->set_event_filter = m1_pmu_set_event_filter;

 cpu_pmu->map_pmuv3_event  = m1_pmu_map_pmuv3_event;
 m1_pmu_init_pmceid(cpu_pmu);

 bitmap_set(cpu_pmu->cntr_mask, 0, M1_PMU_NR_COUNTERS);
 cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_EVENTS] = &m1_pmu_events_attr_group;
 cpu_pmu->attr_groups[ARMPMU_ATTR_GROUP_FORMATS] = &m1_pmu_format_attr_group;
 return 0;
}

/* Device driver gunk */
static int m1_pmu_ice_init(struct arm_pmu *cpu_pmu)
{
 cpu_pmu->name = "apple_icestorm_pmu";
 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_47BIT);
}

static int m1_pmu_fire_init(struct arm_pmu *cpu_pmu)
{
 cpu_pmu->name = "apple_firestorm_pmu";
 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_47BIT);
}

static int m2_pmu_avalanche_init(struct arm_pmu *cpu_pmu)
{
 cpu_pmu->name = "apple_avalanche_pmu";
 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_63BIT);
}

static int m2_pmu_blizzard_init(struct arm_pmu *cpu_pmu)
{
 cpu_pmu->name = "apple_blizzard_pmu";
 return m1_pmu_init(cpu_pmu, ARMPMU_EVT_63BIT);
}

static const struct of_device_id m1_pmu_of_device_ids[] = {
 { .compatible = "apple,avalanche-pmu", .data = m2_pmu_avalanche_init, },
 { .compatible = "apple,blizzard-pmu", .data = m2_pmu_blizzard_init, },
 { .compatible = "apple,icestorm-pmu", .data = m1_pmu_ice_init, },
 { .compatible = "apple,firestorm-pmu", .data = m1_pmu_fire_init, },
 { },
};
MODULE_DEVICE_TABLE(of, m1_pmu_of_device_ids);

static int m1_pmu_device_probe(struct platform_device *pdev)
{
 return arm_pmu_device_probe(pdev, m1_pmu_of_device_ids, NULL);
}

static struct platform_driver m1_pmu_driver = {
 .driver  = {
  .name   = "apple-m1-cpu-pmu",
  .of_match_table  = m1_pmu_of_device_ids,
  .suppress_bind_attrs = true,
 },
 .probe  = m1_pmu_device_probe,
};

module_platform_driver(m1_pmu_driver);