OSDN Git Service

RISC-V: Add perf platform driver based on SBI PMU extension
authorAtish Patra <atish.patra@wdc.com>
Sat, 19 Feb 2022 00:46:57 +0000 (16:46 -0800)
committerPalmer Dabbelt <palmer@rivosinc.com>
Mon, 21 Mar 2022 21:58:33 +0000 (14:58 -0700)
RISC-V SBI specification added a PMU extension that allows to configure
start/stop any pmu counter. The RISC-V perf can use most of the generic
perf features except interrupt overflow and event filtering based on
privilege mode which will be added in future.

It also allows to monitor a handful of firmware counters that can provide
insights into firmware activity during a performance analysis.

Signed-off-by: Atish Patra <atish.patra@wdc.com>
Signed-off-by: Atish Patra <atishp@rivosinc.com>
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
drivers/perf/Kconfig
drivers/perf/Makefile
drivers/perf/riscv_pmu.c
drivers/perf/riscv_pmu_sbi.c [new file with mode: 0644]
include/linux/cpuhotplug.h
include/linux/perf/riscv_pmu.h

index 386162a..5645b56 100644 (file)
@@ -76,6 +76,16 @@ config RISCV_PMU_LEGACY
          of cycle/instruction counter and doesn't support counter overflow,
          or programmable counters. It will be removed in future.
 
+config RISCV_PMU_SBI
+       depends on RISCV_PMU && RISCV_SBI
+       bool "RISC-V PMU based on SBI PMU extension"
+       default y
+       help
+         Say y if you want to use the CPU performance monitor
+         using SBI PMU extension on RISC-V based systems. This option provides
+         full perf feature support i.e. counter overflow, privilege mode
+         filtering, counter configuration.
+
 config ARM_PMU_ACPI
        depends on ARM_PMU && ACPI
        def_bool y
index c3d3268..f149735 100644 (file)
@@ -12,6 +12,7 @@ obj-$(CONFIG_QCOM_L2_PMU)     += qcom_l2_pmu.o
 obj-$(CONFIG_QCOM_L3_PMU) += qcom_l3_pmu.o
 obj-$(CONFIG_RISCV_PMU) += riscv_pmu.o
 obj-$(CONFIG_RISCV_PMU_LEGACY) += riscv_pmu_legacy.o
+obj-$(CONFIG_RISCV_PMU_SBI) += riscv_pmu_sbi.o
 obj-$(CONFIG_THUNDERX2_PMU) += thunderx2_pmu.o
 obj-$(CONFIG_XGENE_PMU) += xgene_pmu.o
 obj-$(CONFIG_ARM_SPE_PMU) += arm_spe_pmu.o
index 590a578..b2b8d20 100644 (file)
@@ -15,6 +15,8 @@
 #include <linux/printk.h>
 #include <linux/smp.h>
 
+#include <asm/sbi.h>
+
 static unsigned long csr_read_num(int csr_num)
 {
 #define switchcase_csr_read(__csr_num, __val)          {\
diff --git a/drivers/perf/riscv_pmu_sbi.c b/drivers/perf/riscv_pmu_sbi.c
new file mode 100644 (file)
index 0000000..815d5c5
--- /dev/null
@@ -0,0 +1,578 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * RISC-V performance counter support.
+ *
+ * Copyright (C) 2021 Western Digital Corporation or its affiliates.
+ *
+ * This code is based on ARM perf event code which is in turn based on
+ * sparc64 and x86 code.
+ */
+
+#define pr_fmt(fmt) "riscv-pmu-sbi: " fmt
+
+#include <linux/mod_devicetable.h>
+#include <linux/perf/riscv_pmu.h>
+#include <linux/platform_device.h>
+
+#include <asm/sbi.h>
+
+union sbi_pmu_ctr_info {
+       unsigned long value;
+       struct {
+               unsigned long csr:12;
+               unsigned long width:6;
+#if __riscv_xlen == 32
+               unsigned long reserved:13;
+#else
+               unsigned long reserved:45;
+#endif
+               unsigned long type:1;
+       };
+};
+
+/**
+ * RISC-V doesn't have hetergenous harts yet. This need to be part of
+ * per_cpu in case of harts with different pmu counters
+ */
+static union sbi_pmu_ctr_info *pmu_ctr_list;
+
+struct sbi_pmu_event_data {
+       union {
+               union {
+                       struct hw_gen_event {
+                               uint32_t event_code:16;
+                               uint32_t event_type:4;
+                               uint32_t reserved:12;
+                       } hw_gen_event;
+                       struct hw_cache_event {
+                               uint32_t result_id:1;
+                               uint32_t op_id:2;
+                               uint32_t cache_id:13;
+                               uint32_t event_type:4;
+                               uint32_t reserved:12;
+                       } hw_cache_event;
+               };
+               uint32_t event_idx;
+       };
+};
+
+static const struct sbi_pmu_event_data pmu_hw_event_map[] = {
+       [PERF_COUNT_HW_CPU_CYCLES]              = {.hw_gen_event = {
+                                                       SBI_PMU_HW_CPU_CYCLES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_INSTRUCTIONS]            = {.hw_gen_event = {
+                                                       SBI_PMU_HW_INSTRUCTIONS,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_CACHE_REFERENCES]        = {.hw_gen_event = {
+                                                       SBI_PMU_HW_CACHE_REFERENCES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_CACHE_MISSES]            = {.hw_gen_event = {
+                                                       SBI_PMU_HW_CACHE_MISSES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]     = {.hw_gen_event = {
+                                                       SBI_PMU_HW_BRANCH_INSTRUCTIONS,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_BRANCH_MISSES]           = {.hw_gen_event = {
+                                                       SBI_PMU_HW_BRANCH_MISSES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_BUS_CYCLES]              = {.hw_gen_event = {
+                                                       SBI_PMU_HW_BUS_CYCLES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = {.hw_gen_event = {
+                                                       SBI_PMU_HW_STALLED_CYCLES_FRONTEND,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_STALLED_CYCLES_BACKEND]  = {.hw_gen_event = {
+                                                       SBI_PMU_HW_STALLED_CYCLES_BACKEND,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+       [PERF_COUNT_HW_REF_CPU_CYCLES]          = {.hw_gen_event = {
+                                                       SBI_PMU_HW_REF_CPU_CYCLES,
+                                                       SBI_PMU_EVENT_TYPE_HW, 0}},
+};
+
+#define C(x) PERF_COUNT_HW_CACHE_##x
+static const struct sbi_pmu_event_data pmu_cache_event_map[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)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(L1D), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(L1I)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS), C(OP_READ),
+                                       C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(L1I), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(LL)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(LL), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(DTLB)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(DTLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(ITLB)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(ITLB), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(BPU)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(BPU), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+       [C(NODE)] = {
+               [C(OP_READ)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_READ), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_WRITE)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_WRITE), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+               [C(OP_PREFETCH)] = {
+                       [C(RESULT_ACCESS)] = {.hw_cache_event = {C(RESULT_ACCESS),
+                                       C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+                       [C(RESULT_MISS)] = {.hw_cache_event = {C(RESULT_MISS),
+                                       C(OP_PREFETCH), C(NODE), SBI_PMU_EVENT_TYPE_CACHE, 0}},
+               },
+       },
+};
+
+static int pmu_sbi_ctr_get_width(int idx)
+{
+       return pmu_ctr_list[idx].width;
+}
+
+static bool pmu_sbi_ctr_is_fw(int cidx)
+{
+       union sbi_pmu_ctr_info *info;
+
+       info = &pmu_ctr_list[cidx];
+       if (!info)
+               return false;
+
+       return (info->type == SBI_PMU_CTR_TYPE_FW) ? true : false;
+}
+
+static int pmu_sbi_ctr_get_idx(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
+       struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
+       struct sbiret ret;
+       int idx;
+       uint64_t cbase = 0;
+       uint64_t cmask = GENMASK_ULL(rvpmu->num_counters - 1, 0);
+       unsigned long cflags = 0;
+
+       if (event->attr.exclude_kernel)
+               cflags |= SBI_PMU_CFG_FLAG_SET_SINH;
+       if (event->attr.exclude_user)
+               cflags |= SBI_PMU_CFG_FLAG_SET_UINH;
+
+       /* retrieve the available counter index */
+       ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_CFG_MATCH, cbase, cmask,
+                       cflags, hwc->event_base, hwc->config, 0);
+       if (ret.error) {
+               pr_debug("Not able to find a counter for event %lx config %llx\n",
+                       hwc->event_base, hwc->config);
+               return sbi_err_map_linux_errno(ret.error);
+       }
+
+       idx = ret.value;
+       if (idx >= rvpmu->num_counters || !pmu_ctr_list[idx].value)
+               return -ENOENT;
+
+       /* Additional sanity check for the counter id */
+       if (pmu_sbi_ctr_is_fw(idx)) {
+               if (!test_and_set_bit(idx, cpuc->used_fw_ctrs))
+                       return idx;
+       } else {
+               if (!test_and_set_bit(idx, cpuc->used_hw_ctrs))
+                       return idx;
+       }
+
+       return -ENOENT;
+}
+
+static void pmu_sbi_ctr_clear_idx(struct perf_event *event)
+{
+
+       struct hw_perf_event *hwc = &event->hw;
+       struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
+       struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
+       int idx = hwc->idx;
+
+       if (pmu_sbi_ctr_is_fw(idx))
+               clear_bit(idx, cpuc->used_fw_ctrs);
+       else
+               clear_bit(idx, cpuc->used_hw_ctrs);
+}
+
+static int pmu_event_find_cache(u64 config)
+{
+       unsigned int cache_type, cache_op, cache_result, ret;
+
+       cache_type = (config >>  0) & 0xff;
+       if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+               return -EINVAL;
+
+       cache_op = (config >>  8) & 0xff;
+       if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
+               return -EINVAL;
+
+       cache_result = (config >> 16) & 0xff;
+       if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+               return -EINVAL;
+
+       ret = pmu_cache_event_map[cache_type][cache_op][cache_result].event_idx;
+
+       return ret;
+}
+
+static bool pmu_sbi_is_fw_event(struct perf_event *event)
+{
+       u32 type = event->attr.type;
+       u64 config = event->attr.config;
+
+       if ((type == PERF_TYPE_RAW) && ((config >> 63) == 1))
+               return true;
+       else
+               return false;
+}
+
+static int pmu_sbi_event_map(struct perf_event *event, u64 *econfig)
+{
+       u32 type = event->attr.type;
+       u64 config = event->attr.config;
+       int bSoftware;
+       u64 raw_config_val;
+       int ret;
+
+       switch (type) {
+       case PERF_TYPE_HARDWARE:
+               if (config >= PERF_COUNT_HW_MAX)
+                       return -EINVAL;
+               ret = pmu_hw_event_map[event->attr.config].event_idx;
+               break;
+       case PERF_TYPE_HW_CACHE:
+               ret = pmu_event_find_cache(config);
+               break;
+       case PERF_TYPE_RAW:
+               /*
+                * As per SBI specification, the upper 16 bits must be unused for
+                * a raw event. Use the MSB (63b) to distinguish between hardware
+                * raw event and firmware events.
+                */
+               bSoftware = config >> 63;
+               raw_config_val = config & RISCV_PMU_RAW_EVENT_MASK;
+               if (bSoftware) {
+                       if (raw_config_val < SBI_PMU_FW_MAX)
+                               ret = (raw_config_val & 0xFFFF) |
+                                     (SBI_PMU_EVENT_TYPE_FW << 16);
+                       else
+                               return -EINVAL;
+               } else {
+                       ret = RISCV_PMU_RAW_EVENT_IDX;
+                       *econfig = raw_config_val;
+               }
+               break;
+       default:
+               ret = -EINVAL;
+               break;
+       }
+
+       return ret;
+}
+
+static u64 pmu_sbi_ctr_read(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       int idx = hwc->idx;
+       struct sbiret ret;
+       union sbi_pmu_ctr_info info;
+       u64 val = 0;
+
+       if (pmu_sbi_is_fw_event(event)) {
+               ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_FW_READ,
+                               hwc->idx, 0, 0, 0, 0, 0);
+               if (!ret.error)
+                       val = ret.value;
+       } else {
+               info = pmu_ctr_list[idx];
+               val = riscv_pmu_ctr_read_csr(info.csr);
+               if (IS_ENABLED(CONFIG_32BIT))
+                       val = ((u64)riscv_pmu_ctr_read_csr(info.csr + 0x80)) << 31 | val;
+       }
+
+       return val;
+}
+
+static void pmu_sbi_ctr_start(struct perf_event *event, u64 ival)
+{
+       struct sbiret ret;
+       struct hw_perf_event *hwc = &event->hw;
+       unsigned long flag = SBI_PMU_START_FLAG_SET_INIT_VALUE;
+
+       ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_START, hwc->idx,
+                       1, flag, ival, ival >> 32, 0);
+       if (ret.error && (ret.error != SBI_ERR_ALREADY_STARTED))
+               pr_err("Starting counter idx %d failed with error %d\n",
+                       hwc->idx, sbi_err_map_linux_errno(ret.error));
+}
+
+static void pmu_sbi_ctr_stop(struct perf_event *event, unsigned long flag)
+{
+       struct sbiret ret;
+       struct hw_perf_event *hwc = &event->hw;
+
+       ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP, hwc->idx, 1, flag, 0, 0, 0);
+       if (ret.error && (ret.error != SBI_ERR_ALREADY_STOPPED) &&
+               flag != SBI_PMU_STOP_FLAG_RESET)
+               pr_err("Stopping counter idx %d failed with error %d\n",
+                       hwc->idx, sbi_err_map_linux_errno(ret.error));
+}
+
+static int pmu_sbi_find_num_ctrs(void)
+{
+       struct sbiret ret;
+
+       ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_NUM_COUNTERS, 0, 0, 0, 0, 0, 0);
+       if (!ret.error)
+               return ret.value;
+       else
+               return sbi_err_map_linux_errno(ret.error);
+}
+
+static int pmu_sbi_get_ctrinfo(int nctr)
+{
+       struct sbiret ret;
+       int i, num_hw_ctr = 0, num_fw_ctr = 0;
+       union sbi_pmu_ctr_info cinfo;
+
+       pmu_ctr_list = kcalloc(nctr, sizeof(*pmu_ctr_list), GFP_KERNEL);
+       if (!pmu_ctr_list)
+               return -ENOMEM;
+
+       for (i = 0; i <= nctr; i++) {
+               ret = sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_GET_INFO, i, 0, 0, 0, 0, 0);
+               if (ret.error)
+                       /* The logical counter ids are not expected to be contiguous */
+                       continue;
+               cinfo.value = ret.value;
+               if (cinfo.type == SBI_PMU_CTR_TYPE_FW)
+                       num_fw_ctr++;
+               else
+                       num_hw_ctr++;
+               pmu_ctr_list[i].value = cinfo.value;
+       }
+
+       pr_info("%d firmware and %d hardware counters\n", num_fw_ctr, num_hw_ctr);
+
+       return 0;
+}
+
+static int pmu_sbi_starting_cpu(unsigned int cpu, struct hlist_node *node)
+{
+       struct riscv_pmu *pmu = hlist_entry_safe(node, struct riscv_pmu, node);
+
+       /* Enable the access for TIME csr only from the user mode now */
+       csr_write(CSR_SCOUNTEREN, 0x2);
+
+       /* Stop all the counters so that they can be enabled from perf */
+       sbi_ecall(SBI_EXT_PMU, SBI_EXT_PMU_COUNTER_STOP,
+                 0, GENMASK_ULL(pmu->num_counters - 1, 0), 0, 0, 0, 0);
+
+       return 0;
+}
+
+static int pmu_sbi_dying_cpu(unsigned int cpu, struct hlist_node *node)
+{
+       /* Disable all counters access for user mode now */
+       csr_write(CSR_SCOUNTEREN, 0x0);
+
+       return 0;
+}
+
+static int pmu_sbi_device_probe(struct platform_device *pdev)
+{
+       struct riscv_pmu *pmu = NULL;
+       int num_counters;
+       int ret;
+
+       pr_info("SBI PMU extension is available\n");
+       pmu = riscv_pmu_alloc();
+       if (!pmu)
+               return -ENOMEM;
+
+       num_counters = pmu_sbi_find_num_ctrs();
+       if (num_counters < 0) {
+               pr_err("SBI PMU extension doesn't provide any counters\n");
+               return -ENODEV;
+       }
+
+       /* cache all the information about counters now */
+       if (pmu_sbi_get_ctrinfo(num_counters))
+               return -ENODEV;
+
+       pmu->num_counters = num_counters;
+       pmu->ctr_start = pmu_sbi_ctr_start;
+       pmu->ctr_stop = pmu_sbi_ctr_stop;
+       pmu->event_map = pmu_sbi_event_map;
+       pmu->ctr_get_idx = pmu_sbi_ctr_get_idx;
+       pmu->ctr_get_width = pmu_sbi_ctr_get_width;
+       pmu->ctr_clear_idx = pmu_sbi_ctr_clear_idx;
+       pmu->ctr_read = pmu_sbi_ctr_read;
+
+       ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
+       if (ret)
+               return ret;
+
+       ret = perf_pmu_register(&pmu->pmu, "cpu", PERF_TYPE_RAW);
+       if (ret) {
+               cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_STARTING, &pmu->node);
+               return ret;
+       }
+
+       return 0;
+}
+
+static struct platform_driver pmu_sbi_driver = {
+       .probe          = pmu_sbi_device_probe,
+       .driver         = {
+               .name   = RISCV_PMU_PDEV_NAME,
+       },
+};
+
+static int __init pmu_sbi_devinit(void)
+{
+       int ret;
+       struct platform_device *pdev;
+
+       if (sbi_spec_version < sbi_mk_version(0, 3) ||
+           sbi_probe_extension(SBI_EXT_PMU) <= 0) {
+               return 0;
+       }
+
+       ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_RISCV_STARTING,
+                                     "perf/riscv/pmu:starting",
+                                     pmu_sbi_starting_cpu, pmu_sbi_dying_cpu);
+       if (ret) {
+               pr_err("CPU hotplug notifier could not be registered: %d\n",
+                      ret);
+               return ret;
+       }
+
+       ret = platform_driver_register(&pmu_sbi_driver);
+       if (ret)
+               return ret;
+
+       pdev = platform_device_register_simple(RISCV_PMU_PDEV_NAME, -1, NULL, 0);
+       if (IS_ERR(pdev)) {
+               platform_driver_unregister(&pmu_sbi_driver);
+               return PTR_ERR(pdev);
+       }
+
+       /* Notify legacy implementation that SBI pmu is available*/
+       riscv_pmu_legacy_skip_init();
+
+       return ret;
+}
+device_initcall(pmu_sbi_devinit)
index 411a428..51c7e49 100644 (file)
@@ -165,6 +165,7 @@ enum cpuhp_state {
        CPUHP_AP_PERF_ARM_HW_BREAKPOINT_STARTING,
        CPUHP_AP_PERF_ARM_ACPI_STARTING,
        CPUHP_AP_PERF_ARM_STARTING,
+       CPUHP_AP_PERF_RISCV_STARTING,
        CPUHP_AP_ARM_L2X0_STARTING,
        CPUHP_AP_EXYNOS4_MCT_TIMER_STARTING,
        CPUHP_AP_ARM_ARCH_TIMER_STARTING,
index 9140c49..0f22694 100644 (file)
@@ -31,8 +31,10 @@ struct cpu_hw_events {
        int                     n_events;
        /* currently enabled events */
        struct perf_event       *events[RISCV_MAX_COUNTERS];
-       /* currently enabled counters */
-       DECLARE_BITMAP(used_event_ctrs, RISCV_MAX_COUNTERS);
+       /* currently enabled hardware counters */
+       DECLARE_BITMAP(used_hw_ctrs, RISCV_MAX_COUNTERS);
+       /* currently enabled firmware counters */
+       DECLARE_BITMAP(used_fw_ctrs, RISCV_MAX_COUNTERS);
 };
 
 struct riscv_pmu {