RISC-V: Remove the current perf implementation
authorAtish Patra <atish.patra@wdc.com>
Sat, 19 Feb 2022 00:46:52 +0000 (16:46 -0800)
committerPalmer Dabbelt <palmer@rivosinc.com>
Mon, 21 Mar 2022 21:58:12 +0000 (14:58 -0700)
The current perf implementation in RISC-V is not very useful as it can not
count any events other than cycle/instructions. Moreover, perf record
can not be used or the events can not be started or stopped.

Remove the implementation now for a better platform driver in future
that will implement most of the missing functionality.

Reviewed-by: Anup Patel <anup@brainfault.org>
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>
arch/riscv/Kconfig
arch/riscv/include/asm/perf_event.h
arch/riscv/kernel/Makefile
arch/riscv/kernel/perf_event.c [deleted file]

index 5adcbd9b5e88637f3e5d6fa562da6b15e7ec846a..00cc3d4c77d12cd4daa830e918c401d3d568333b 100644 (file)
@@ -333,19 +333,6 @@ config RISCV_ISA_C
 
           If you don't know what to do here, say Y.
 
-menu "supported PMU type"
-       depends on PERF_EVENTS
-
-config RISCV_BASE_PMU
-       bool "Base Performance Monitoring Unit"
-       def_bool y
-       help
-         A base PMU that serves as a reference implementation and has limited
-         feature of perf.  It can run on any RISC-V machines so serves as the
-         fallback, but this option can also be disable to reduce kernel size.
-
-endmenu
-
 config FPU
        bool "FPU support"
        default y
index 062efd3a1d5d1a6059d914e94a3d645daca81af2..d42c901f9a9770985b53da11098d69cf7934cc62 100644 (file)
@@ -9,77 +9,5 @@
 #define _ASM_RISCV_PERF_EVENT_H
 
 #include <linux/perf_event.h>
-#include <linux/ptrace.h>
-#include <linux/interrupt.h>
-
-#ifdef CONFIG_RISCV_BASE_PMU
-#define RISCV_BASE_COUNTERS    2
-
-/*
- * The RISCV_MAX_COUNTERS parameter should be specified.
- */
-
-#define RISCV_MAX_COUNTERS     2
-
-/*
- * These are the indexes of bits in counteren register *minus* 1,
- * except for cycle.  It would be coherent if it can directly mapped
- * to counteren bit definition, but there is a *time* register at
- * counteren[1].  Per-cpu structure is scarce resource here.
- *
- * According to the spec, an implementation can support counter up to
- * mhpmcounter31, but many high-end processors has at most 6 general
- * PMCs, we give the definition to MHPMCOUNTER8 here.
- */
-#define RISCV_PMU_CYCLE                0
-#define RISCV_PMU_INSTRET      1
-#define RISCV_PMU_MHPMCOUNTER3 2
-#define RISCV_PMU_MHPMCOUNTER4 3
-#define RISCV_PMU_MHPMCOUNTER5 4
-#define RISCV_PMU_MHPMCOUNTER6 5
-#define RISCV_PMU_MHPMCOUNTER7 6
-#define RISCV_PMU_MHPMCOUNTER8 7
-
-#define RISCV_OP_UNSUPP                (-EOPNOTSUPP)
-
-struct cpu_hw_events {
-       /* # currently enabled events*/
-       int                     n_events;
-       /* currently enabled events */
-       struct perf_event       *events[RISCV_MAX_COUNTERS];
-       /* vendor-defined PMU data */
-       void                    *platform;
-};
-
-struct riscv_pmu {
-       struct pmu      *pmu;
-
-       /* generic hw/cache events table */
-       const int       *hw_events;
-       const int       (*cache_events)[PERF_COUNT_HW_CACHE_MAX]
-                                      [PERF_COUNT_HW_CACHE_OP_MAX]
-                                      [PERF_COUNT_HW_CACHE_RESULT_MAX];
-       /* method used to map hw/cache events */
-       int             (*map_hw_event)(u64 config);
-       int             (*map_cache_event)(u64 config);
-
-       /* max generic hw events in map */
-       int             max_events;
-       /* number total counters, 2(base) + x(general) */
-       int             num_counters;
-       /* the width of the counter */
-       int             counter_width;
-
-       /* vendor-defined PMU features */
-       void            *platform;
-
-       irqreturn_t     (*handle_irq)(int irq_num, void *dev);
-       int             irq;
-};
-
-#endif
-#ifdef CONFIG_PERF_EVENTS
 #define perf_arch_bpf_user_pt_regs(regs) (struct user_regs_struct *)regs
-#endif
-
 #endif /* _ASM_RISCV_PERF_EVENT_H */
index 612556faa527f583cf6fc3bc9a737adcaa18a105..fb63b462ff85d2fb466b82fe77de435969e86132 100644 (file)
@@ -51,7 +51,6 @@ obj-$(CONFIG_MODULE_SECTIONS) += module-sections.o
 obj-$(CONFIG_FUNCTION_TRACER)  += mcount.o ftrace.o
 obj-$(CONFIG_DYNAMIC_FTRACE)   += mcount-dyn.o
 
-obj-$(CONFIG_RISCV_BASE_PMU)   += perf_event.o
 obj-$(CONFIG_PERF_EVENTS)      += perf_callchain.o
 obj-$(CONFIG_HAVE_PERF_REGS)   += perf_regs.o
 obj-$(CONFIG_RISCV_SBI)                += sbi.o
diff --git a/arch/riscv/kernel/perf_event.c b/arch/riscv/kernel/perf_event.c
deleted file mode 100644 (file)
index c835f03..0000000
+++ /dev/null
@@ -1,485 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-/*
- * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
- * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
- * Copyright (C) 2009 Jaswinder Singh Rajput
- * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
- * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
- * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
- * Copyright (C) 2009 Google, Inc., Stephane Eranian
- * Copyright 2014 Tilera Corporation. All Rights Reserved.
- * Copyright (C) 2018 Andes Technology Corporation
- *
- * Perf_events support for RISC-V platforms.
- *
- * Since the spec. (as of now, Priv-Spec 1.10) does not provide enough
- * functionality for perf event to fully work, this file provides
- * the very basic framework only.
- *
- * For platform portings, please check Documentations/riscv/pmu.txt.
- *
- * The Copyright line includes x86 and tile ones.
- */
-
-#include <linux/kprobes.h>
-#include <linux/kernel.h>
-#include <linux/kdebug.h>
-#include <linux/mutex.h>
-#include <linux/bitmap.h>
-#include <linux/irq.h>
-#include <linux/perf_event.h>
-#include <linux/atomic.h>
-#include <linux/of.h>
-#include <asm/perf_event.h>
-
-static const struct riscv_pmu *riscv_pmu __read_mostly;
-static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
-
-/*
- * Hardware & cache maps and their methods
- */
-
-static const int riscv_hw_event_map[] = {
-       [PERF_COUNT_HW_CPU_CYCLES]              = RISCV_PMU_CYCLE,
-       [PERF_COUNT_HW_INSTRUCTIONS]            = RISCV_PMU_INSTRET,
-       [PERF_COUNT_HW_CACHE_REFERENCES]        = RISCV_OP_UNSUPP,
-       [PERF_COUNT_HW_CACHE_MISSES]            = RISCV_OP_UNSUPP,
-       [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]     = RISCV_OP_UNSUPP,
-       [PERF_COUNT_HW_BRANCH_MISSES]           = RISCV_OP_UNSUPP,
-       [PERF_COUNT_HW_BUS_CYCLES]              = RISCV_OP_UNSUPP,
-};
-
-#define C(x) PERF_COUNT_HW_CACHE_##x
-static const int riscv_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)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-       [C(L1I)] = {
-               [C(OP_READ)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-       [C(LL)] = {
-               [C(OP_READ)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-       [C(DTLB)] = {
-               [C(OP_READ)] = {
-                       [C(RESULT_ACCESS)] =  RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] =  RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-       [C(ITLB)] = {
-               [C(OP_READ)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-       [C(BPU)] = {
-               [C(OP_READ)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_WRITE)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-               [C(OP_PREFETCH)] = {
-                       [C(RESULT_ACCESS)] = RISCV_OP_UNSUPP,
-                       [C(RESULT_MISS)] = RISCV_OP_UNSUPP,
-               },
-       },
-};
-
-static int riscv_map_hw_event(u64 config)
-{
-       if (config >= riscv_pmu->max_events)
-               return -EINVAL;
-
-       return riscv_pmu->hw_events[config];
-}
-
-static int riscv_map_cache_decode(u64 config, unsigned int *type,
-                          unsigned int *op, unsigned int *result)
-{
-       return -ENOENT;
-}
-
-static int riscv_map_cache_event(u64 config)
-{
-       unsigned int type, op, result;
-       int err = -ENOENT;
-               int code;
-
-       err = riscv_map_cache_decode(config, &type, &op, &result);
-       if (!riscv_pmu->cache_events || err)
-               return err;
-
-       if (type >= PERF_COUNT_HW_CACHE_MAX ||
-           op >= PERF_COUNT_HW_CACHE_OP_MAX ||
-           result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
-               return -EINVAL;
-
-       code = (*riscv_pmu->cache_events)[type][op][result];
-       if (code == RISCV_OP_UNSUPP)
-               return -EINVAL;
-
-       return code;
-}
-
-/*
- * Low-level functions: reading/writing counters
- */
-
-static inline u64 read_counter(int idx)
-{
-       u64 val = 0;
-
-       switch (idx) {
-       case RISCV_PMU_CYCLE:
-               val = csr_read(CSR_CYCLE);
-               break;
-       case RISCV_PMU_INSTRET:
-               val = csr_read(CSR_INSTRET);
-               break;
-       default:
-               WARN_ON_ONCE(idx < 0 || idx > RISCV_MAX_COUNTERS);
-               return -EINVAL;
-       }
-
-       return val;
-}
-
-static inline void write_counter(int idx, u64 value)
-{
-       /* currently not supported */
-       WARN_ON_ONCE(1);
-}
-
-/*
- * pmu->read: read and update the counter
- *
- * Other architectures' implementation often have a xxx_perf_event_update
- * routine, which can return counter values when called in the IRQ, but
- * return void when being called by the pmu->read method.
- */
-static void riscv_pmu_read(struct perf_event *event)
-{
-       struct hw_perf_event *hwc = &event->hw;
-       u64 prev_raw_count, new_raw_count;
-       u64 oldval;
-       int idx = hwc->idx;
-       u64 delta;
-
-       do {
-               prev_raw_count = local64_read(&hwc->prev_count);
-               new_raw_count = read_counter(idx);
-
-               oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count,
-                                        new_raw_count);
-       } while (oldval != prev_raw_count);
-
-       /*
-        * delta is the value to update the counter we maintain in the kernel.
-        */
-       delta = (new_raw_count - prev_raw_count) &
-               ((1ULL << riscv_pmu->counter_width) - 1);
-       local64_add(delta, &event->count);
-       /*
-        * Something like local64_sub(delta, &hwc->period_left) here is
-        * needed if there is an interrupt for perf.
-        */
-}
-
-/*
- * State transition functions:
- *
- * stop()/start() & add()/del()
- */
-
-/*
- * pmu->stop: stop the counter
- */
-static void riscv_pmu_stop(struct perf_event *event, int flags)
-{
-       struct hw_perf_event *hwc = &event->hw;
-
-       WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
-       hwc->state |= PERF_HES_STOPPED;
-
-       if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
-               riscv_pmu->pmu->read(event);
-               hwc->state |= PERF_HES_UPTODATE;
-       }
-}
-
-/*
- * pmu->start: start the event.
- */
-static void riscv_pmu_start(struct perf_event *event, int flags)
-{
-       struct hw_perf_event *hwc = &event->hw;
-
-       if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
-               return;
-
-       if (flags & PERF_EF_RELOAD) {
-               WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
-
-               /*
-                * Set the counter to the period to the next interrupt here,
-                * if you have any.
-                */
-       }
-
-       hwc->state = 0;
-       perf_event_update_userpage(event);
-
-       /*
-        * Since we cannot write to counters, this serves as an initialization
-        * to the delta-mechanism in pmu->read(); otherwise, the delta would be
-        * wrong when pmu->read is called for the first time.
-        */
-       local64_set(&hwc->prev_count, read_counter(hwc->idx));
-}
-
-/*
- * pmu->add: add the event to PMU.
- */
-static int riscv_pmu_add(struct perf_event *event, int flags)
-{
-       struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
-       struct hw_perf_event *hwc = &event->hw;
-
-       if (cpuc->n_events == riscv_pmu->num_counters)
-               return -ENOSPC;
-
-       /*
-        * We don't have general conunters, so no binding-event-to-counter
-        * process here.
-        *
-        * Indexing using hwc->config generally not works, since config may
-        * contain extra information, but here the only info we have in
-        * hwc->config is the event index.
-        */
-       hwc->idx = hwc->config;
-       cpuc->events[hwc->idx] = event;
-       cpuc->n_events++;
-
-       hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
-
-       if (flags & PERF_EF_START)
-               riscv_pmu->pmu->start(event, PERF_EF_RELOAD);
-
-       return 0;
-}
-
-/*
- * pmu->del: delete the event from PMU.
- */
-static void riscv_pmu_del(struct perf_event *event, int flags)
-{
-       struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
-       struct hw_perf_event *hwc = &event->hw;
-
-       cpuc->events[hwc->idx] = NULL;
-       cpuc->n_events--;
-       riscv_pmu->pmu->stop(event, PERF_EF_UPDATE);
-       perf_event_update_userpage(event);
-}
-
-/*
- * Interrupt: a skeletion for reference.
- */
-
-static DEFINE_MUTEX(pmc_reserve_mutex);
-
-static irqreturn_t riscv_base_pmu_handle_irq(int irq_num, void *dev)
-{
-       return IRQ_NONE;
-}
-
-static int reserve_pmc_hardware(void)
-{
-       int err = 0;
-
-       mutex_lock(&pmc_reserve_mutex);
-       if (riscv_pmu->irq >= 0 && riscv_pmu->handle_irq) {
-               err = request_irq(riscv_pmu->irq, riscv_pmu->handle_irq,
-                                 IRQF_PERCPU, "riscv-base-perf", NULL);
-       }
-       mutex_unlock(&pmc_reserve_mutex);
-
-       return err;
-}
-
-static void release_pmc_hardware(void)
-{
-       mutex_lock(&pmc_reserve_mutex);
-       if (riscv_pmu->irq >= 0)
-               free_irq(riscv_pmu->irq, NULL);
-       mutex_unlock(&pmc_reserve_mutex);
-}
-
-/*
- * Event Initialization/Finalization
- */
-
-static atomic_t riscv_active_events = ATOMIC_INIT(0);
-
-static void riscv_event_destroy(struct perf_event *event)
-{
-       if (atomic_dec_return(&riscv_active_events) == 0)
-               release_pmc_hardware();
-}
-
-static int riscv_event_init(struct perf_event *event)
-{
-       struct perf_event_attr *attr = &event->attr;
-       struct hw_perf_event *hwc = &event->hw;
-       int err;
-       int code;
-
-       if (atomic_inc_return(&riscv_active_events) == 1) {
-               err = reserve_pmc_hardware();
-
-               if (err) {
-                       pr_warn("PMC hardware not available\n");
-                       atomic_dec(&riscv_active_events);
-                       return -EBUSY;
-               }
-       }
-
-       switch (event->attr.type) {
-       case PERF_TYPE_HARDWARE:
-               code = riscv_pmu->map_hw_event(attr->config);
-               break;
-       case PERF_TYPE_HW_CACHE:
-               code = riscv_pmu->map_cache_event(attr->config);
-               break;
-       case PERF_TYPE_RAW:
-               return -EOPNOTSUPP;
-       default:
-               return -ENOENT;
-       }
-
-       event->destroy = riscv_event_destroy;
-       if (code < 0) {
-               event->destroy(event);
-               return code;
-       }
-
-       /*
-        * idx is set to -1 because the index of a general event should not be
-        * decided until binding to some counter in pmu->add().
-        *
-        * But since we don't have such support, later in pmu->add(), we just
-        * use hwc->config as the index instead.
-        */
-       hwc->config = code;
-       hwc->idx = -1;
-
-       return 0;
-}
-
-/*
- * Initialization
- */
-
-static struct pmu min_pmu = {
-       .name           = "riscv-base",
-       .event_init     = riscv_event_init,
-       .add            = riscv_pmu_add,
-       .del            = riscv_pmu_del,
-       .start          = riscv_pmu_start,
-       .stop           = riscv_pmu_stop,
-       .read           = riscv_pmu_read,
-};
-
-static const struct riscv_pmu riscv_base_pmu = {
-       .pmu = &min_pmu,
-       .max_events = ARRAY_SIZE(riscv_hw_event_map),
-       .map_hw_event = riscv_map_hw_event,
-       .hw_events = riscv_hw_event_map,
-       .map_cache_event = riscv_map_cache_event,
-       .cache_events = &riscv_cache_event_map,
-       .counter_width = 63,
-       .num_counters = RISCV_BASE_COUNTERS + 0,
-       .handle_irq = &riscv_base_pmu_handle_irq,
-
-       /* This means this PMU has no IRQ. */
-       .irq = -1,
-};
-
-static const struct of_device_id riscv_pmu_of_ids[] = {
-       {.compatible = "riscv,base-pmu",        .data = &riscv_base_pmu},
-       { /* sentinel value */ }
-};
-
-static int __init init_hw_perf_events(void)
-{
-       struct device_node *node = of_find_node_by_type(NULL, "pmu");
-       const struct of_device_id *of_id;
-
-       riscv_pmu = &riscv_base_pmu;
-
-       if (node) {
-               of_id = of_match_node(riscv_pmu_of_ids, node);
-
-               if (of_id)
-                       riscv_pmu = of_id->data;
-               of_node_put(node);
-       }
-
-       perf_pmu_register(riscv_pmu->pmu, "cpu", PERF_TYPE_RAW);
-       return 0;
-}
-arch_initcall(init_hw_perf_events);