| 1 | /* |
| 2 | * SPDX-License-Identifier: MIT |
| 3 | * |
| 4 | * Copyright © 2017-2018 Intel Corporation |
| 5 | */ |
| 6 | |
| 7 | #ifndef __I915_PMU_H__ |
| 8 | #define __I915_PMU_H__ |
| 9 | |
| 10 | #include <linux/hrtimer.h> |
| 11 | #include <linux/perf_event.h> |
| 12 | #include <linux/spinlock_types.h> |
| 13 | #include <uapi/drm/i915_drm.h> |
| 14 | |
| 15 | struct drm_i915_private; |
| 16 | struct intel_gt; |
| 17 | |
| 18 | /* |
| 19 | * Non-engine events that we need to track enabled-disabled transition and |
| 20 | * current state. |
| 21 | */ |
| 22 | enum i915_pmu_tracked_events { |
| 23 | __I915_PMU_ACTUAL_FREQUENCY_ENABLED = 0, |
| 24 | __I915_PMU_REQUESTED_FREQUENCY_ENABLED, |
| 25 | __I915_PMU_RC6_RESIDENCY_ENABLED, |
| 26 | __I915_PMU_TRACKED_EVENT_COUNT, /* count marker */ |
| 27 | }; |
| 28 | |
| 29 | /* |
| 30 | * Slots used from the sampling timer (non-engine events) with some extras for |
| 31 | * convenience. |
| 32 | */ |
| 33 | enum { |
| 34 | __I915_SAMPLE_FREQ_ACT = 0, |
| 35 | __I915_SAMPLE_FREQ_REQ, |
| 36 | __I915_SAMPLE_RC6, |
| 37 | __I915_SAMPLE_RC6_LAST_REPORTED, |
| 38 | __I915_NUM_PMU_SAMPLERS |
| 39 | }; |
| 40 | |
| 41 | #define I915_PMU_MAX_GT 2 |
| 42 | |
| 43 | /* |
| 44 | * How many different events we track in the global PMU mask. |
| 45 | * |
| 46 | * It is also used to know to needed number of event reference counters. |
| 47 | */ |
| 48 | #define I915_PMU_MASK_BITS \ |
| 49 | (I915_ENGINE_SAMPLE_COUNT + \ |
| 50 | I915_PMU_MAX_GT * __I915_PMU_TRACKED_EVENT_COUNT) |
| 51 | |
| 52 | #define I915_ENGINE_SAMPLE_COUNT (I915_SAMPLE_SEMA + 1) |
| 53 | |
| 54 | struct i915_pmu_sample { |
| 55 | u64 cur; |
| 56 | }; |
| 57 | |
| 58 | struct i915_pmu { |
| 59 | /** |
| 60 | * @base: PMU base. |
| 61 | */ |
| 62 | struct pmu base; |
| 63 | /** |
| 64 | * @registered: PMU is registered and not in the unregistering process. |
| 65 | */ |
| 66 | bool registered; |
| 67 | /** |
| 68 | * @name: Name as registered with perf core. |
| 69 | */ |
| 70 | const char *name; |
| 71 | /** |
| 72 | * @lock: Lock protecting enable mask and ref count handling. |
| 73 | */ |
| 74 | spinlock_t lock; |
| 75 | /** |
| 76 | * @unparked: GT unparked mask. |
| 77 | */ |
| 78 | unsigned int unparked; |
| 79 | /** |
| 80 | * @timer: Timer for internal i915 PMU sampling. |
| 81 | */ |
| 82 | struct hrtimer timer; |
| 83 | /** |
| 84 | * @enable: Bitmask of specific enabled events. |
| 85 | * |
| 86 | * For some events we need to track their state and do some internal |
| 87 | * house keeping. |
| 88 | * |
| 89 | * Each engine event sampler type and event listed in enum |
| 90 | * i915_pmu_tracked_events gets a bit in this field. |
| 91 | * |
| 92 | * Low bits are engine samplers and other events continue from there. |
| 93 | */ |
| 94 | u32 enable; |
| 95 | |
| 96 | /** |
| 97 | * @timer_last: |
| 98 | * |
| 99 | * Timestamp of the previous timer invocation. |
| 100 | */ |
| 101 | ktime_t timer_last; |
| 102 | |
| 103 | /** |
| 104 | * @enable_count: Reference counts for the enabled events. |
| 105 | * |
| 106 | * Array indices are mapped in the same way as bits in the @enable field |
| 107 | * and they are used to control sampling on/off when multiple clients |
| 108 | * are using the PMU API. |
| 109 | */ |
| 110 | unsigned int enable_count[I915_PMU_MASK_BITS]; |
| 111 | /** |
| 112 | * @timer_enabled: Should the internal sampling timer be running. |
| 113 | */ |
| 114 | bool timer_enabled; |
| 115 | /** |
| 116 | * @sample: Current and previous (raw) counters for sampling events. |
| 117 | * |
| 118 | * These counters are updated from the i915 PMU sampling timer. |
| 119 | * |
| 120 | * Only global counters are held here, while the per-engine ones are in |
| 121 | * struct intel_engine_cs. |
| 122 | */ |
| 123 | struct i915_pmu_sample sample[I915_PMU_MAX_GT][__I915_NUM_PMU_SAMPLERS]; |
| 124 | /** |
| 125 | * @sleep_last: Last time GT parked for RC6 estimation. |
| 126 | */ |
| 127 | ktime_t sleep_last[I915_PMU_MAX_GT]; |
| 128 | /** |
| 129 | * @irq_count: Number of interrupts |
| 130 | * |
| 131 | * Intentionally unsigned long to avoid atomics or heuristics on 32bit. |
| 132 | * 4e9 interrupts are a lot and postprocessing can really deal with an |
| 133 | * occasional wraparound easily. It's 32bit after all. |
| 134 | */ |
| 135 | unsigned long irq_count; |
| 136 | /** |
| 137 | * @events_attr_group: Device events attribute group. |
| 138 | */ |
| 139 | struct attribute_group events_attr_group; |
| 140 | /** |
| 141 | * @i915_attr: Memory block holding device attributes. |
| 142 | */ |
| 143 | void *i915_attr; |
| 144 | /** |
| 145 | * @pmu_attr: Memory block holding device attributes. |
| 146 | */ |
| 147 | void *pmu_attr; |
| 148 | }; |
| 149 | |
| 150 | #ifdef CONFIG_PERF_EVENTS |
| 151 | void i915_pmu_register(struct drm_i915_private *i915); |
| 152 | void i915_pmu_unregister(struct drm_i915_private *i915); |
| 153 | void i915_pmu_gt_parked(struct intel_gt *gt); |
| 154 | void i915_pmu_gt_unparked(struct intel_gt *gt); |
| 155 | #else |
| 156 | static inline void i915_pmu_register(struct drm_i915_private *i915) {} |
| 157 | static inline void i915_pmu_unregister(struct drm_i915_private *i915) {} |
| 158 | static inline void i915_pmu_gt_parked(struct intel_gt *gt) {} |
| 159 | static inline void i915_pmu_gt_unparked(struct intel_gt *gt) {} |
| 160 | #endif |
| 161 | |
| 162 | #endif |
| 163 | |