| 1 | //===-- xray_fdr_log_writer.h ---------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file is a part of XRay, a function call tracing system. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | #ifndef COMPILER_RT_LIB_XRAY_XRAY_FDR_LOG_WRITER_H_ |
| 13 | #define COMPILER_RT_LIB_XRAY_XRAY_FDR_LOG_WRITER_H_ |
| 14 | |
| 15 | #include "xray_buffer_queue.h" |
| 16 | #include "xray_fdr_log_records.h" |
| 17 | #include <functional> |
| 18 | #include <tuple> |
| 19 | #include <type_traits> |
| 20 | #include <utility> |
| 21 | |
| 22 | namespace __xray { |
| 23 | |
| 24 | template <size_t Index> struct SerializerImpl { |
| 25 | template <class Tuple, |
| 26 | typename std::enable_if< |
| 27 | Index<std::tuple_size< |
| 28 | typename std::remove_reference<Tuple>::type>::value, |
| 29 | int>::type = 0> static void serializeTo(char *Buffer, |
| 30 | Tuple &&T) { |
| 31 | auto P = reinterpret_cast<const char *>(&std::get<Index>(T)); |
| 32 | constexpr auto Size = sizeof(std::get<Index>(T)); |
| 33 | internal_memcpy(dest: Buffer, src: P, n: Size); |
| 34 | SerializerImpl<Index + 1>::serializeTo(Buffer + Size, |
| 35 | std::forward<Tuple>(T)); |
| 36 | } |
| 37 | |
| 38 | template <class Tuple, |
| 39 | typename std::enable_if< |
| 40 | Index >= std::tuple_size<typename std::remove_reference< |
| 41 | Tuple>::type>::value, |
| 42 | int>::type = 0> |
| 43 | static void serializeTo(char *, Tuple &&) {} |
| 44 | }; |
| 45 | |
| 46 | using Serializer = SerializerImpl<0>; |
| 47 | |
| 48 | template <class Tuple, size_t Index> struct AggregateSizesImpl { |
| 49 | static constexpr size_t value = |
| 50 | sizeof(typename std::tuple_element<Index, Tuple>::type) + |
| 51 | AggregateSizesImpl<Tuple, Index - 1>::value; |
| 52 | }; |
| 53 | |
| 54 | template <class Tuple> struct AggregateSizesImpl<Tuple, 0> { |
| 55 | static constexpr size_t value = |
| 56 | sizeof(typename std::tuple_element<0, Tuple>::type); |
| 57 | }; |
| 58 | |
| 59 | template <class Tuple> struct AggregateSizes { |
| 60 | static constexpr size_t value = |
| 61 | AggregateSizesImpl<Tuple, std::tuple_size<Tuple>::value - 1>::value; |
| 62 | }; |
| 63 | |
| 64 | template <MetadataRecord::RecordKinds Kind, class... DataTypes> |
| 65 | MetadataRecord createMetadataRecord(DataTypes &&... Ds) { |
| 66 | static_assert(AggregateSizes<std::tuple<DataTypes...>>::value <= |
| 67 | sizeof(MetadataRecord) - 1, |
| 68 | "Metadata payload longer than metadata buffer!" ); |
| 69 | MetadataRecord R; |
| 70 | R.Type = 1; |
| 71 | R.RecordKind = static_cast<uint8_t>(Kind); |
| 72 | Serializer::serializeTo(R.Data, |
| 73 | std::make_tuple(std::forward<DataTypes>(Ds)...)); |
| 74 | return R; |
| 75 | } |
| 76 | |
| 77 | class FDRLogWriter { |
| 78 | BufferQueue::Buffer &Buffer; |
| 79 | char *NextRecord = nullptr; |
| 80 | |
| 81 | template <class T> void writeRecord(const T &R) { |
| 82 | internal_memcpy(dest: NextRecord, src: reinterpret_cast<const char *>(&R), n: sizeof(T)); |
| 83 | NextRecord += sizeof(T); |
| 84 | // We need this atomic fence here to ensure that other threads attempting to |
| 85 | // read the bytes in the buffer will see the writes committed before the |
| 86 | // extents are updated. |
| 87 | atomic_thread_fence(mo: memory_order_release); |
| 88 | atomic_fetch_add(a: Buffer.Extents, v: sizeof(T), mo: memory_order_acq_rel); |
| 89 | } |
| 90 | |
| 91 | public: |
| 92 | explicit FDRLogWriter(BufferQueue::Buffer &B, char *P) |
| 93 | : Buffer(B), NextRecord(P) { |
| 94 | DCHECK_NE(Buffer.Data, nullptr); |
| 95 | DCHECK_NE(NextRecord, nullptr); |
| 96 | } |
| 97 | |
| 98 | explicit FDRLogWriter(BufferQueue::Buffer &B) |
| 99 | : FDRLogWriter(B, static_cast<char *>(B.Data)) {} |
| 100 | |
| 101 | template <MetadataRecord::RecordKinds Kind, class... Data> |
| 102 | bool writeMetadata(Data &&... Ds) { |
| 103 | // TODO: Check boundary conditions: |
| 104 | // 1) Buffer is full, and cannot handle one metadata record. |
| 105 | // 2) Buffer queue is finalising. |
| 106 | writeRecord(createMetadataRecord<Kind>(std::forward<Data>(Ds)...)); |
| 107 | return true; |
| 108 | } |
| 109 | |
| 110 | template <size_t N> size_t writeMetadataRecords(MetadataRecord (&Recs)[N]) { |
| 111 | constexpr auto Size = sizeof(MetadataRecord) * N; |
| 112 | internal_memcpy(dest: NextRecord, src: reinterpret_cast<const char *>(Recs), n: Size); |
| 113 | NextRecord += Size; |
| 114 | // We need this atomic fence here to ensure that other threads attempting to |
| 115 | // read the bytes in the buffer will see the writes committed before the |
| 116 | // extents are updated. |
| 117 | atomic_thread_fence(mo: memory_order_release); |
| 118 | atomic_fetch_add(a: Buffer.Extents, v: Size, mo: memory_order_acq_rel); |
| 119 | return Size; |
| 120 | } |
| 121 | |
| 122 | enum class FunctionRecordKind : uint8_t { |
| 123 | Enter = 0x00, |
| 124 | Exit = 0x01, |
| 125 | TailExit = 0x02, |
| 126 | EnterArg = 0x03, |
| 127 | }; |
| 128 | |
| 129 | bool writeFunction(FunctionRecordKind Kind, int32_t FuncId, int32_t Delta) { |
| 130 | FunctionRecord R; |
| 131 | R.Type = 0; |
| 132 | R.RecordKind = uint8_t(Kind); |
| 133 | R.FuncId = FuncId; |
| 134 | R.TSCDelta = Delta; |
| 135 | writeRecord(R); |
| 136 | return true; |
| 137 | } |
| 138 | |
| 139 | bool writeFunctionWithArg(FunctionRecordKind Kind, int32_t FuncId, |
| 140 | int32_t Delta, uint64_t Arg) { |
| 141 | // We need to write the function with arg into the buffer, and then |
| 142 | // atomically update the buffer extents. This ensures that any reads |
| 143 | // synchronised on the buffer extents record will always see the writes |
| 144 | // that happen before the atomic update. |
| 145 | FunctionRecord R; |
| 146 | R.Type = 0; |
| 147 | R.RecordKind = uint8_t(Kind); |
| 148 | R.FuncId = FuncId; |
| 149 | R.TSCDelta = Delta; |
| 150 | MetadataRecord A = |
| 151 | createMetadataRecord<MetadataRecord::RecordKinds::CallArgument>(Ds&: Arg); |
| 152 | NextRecord = reinterpret_cast<char *>(internal_memcpy( |
| 153 | dest: NextRecord, src: reinterpret_cast<char *>(&R), n: sizeof(R))) + |
| 154 | sizeof(R); |
| 155 | NextRecord = reinterpret_cast<char *>(internal_memcpy( |
| 156 | dest: NextRecord, src: reinterpret_cast<char *>(&A), n: sizeof(A))) + |
| 157 | sizeof(A); |
| 158 | // We need this atomic fence here to ensure that other threads attempting to |
| 159 | // read the bytes in the buffer will see the writes committed before the |
| 160 | // extents are updated. |
| 161 | atomic_thread_fence(mo: memory_order_release); |
| 162 | atomic_fetch_add(a: Buffer.Extents, v: sizeof(R) + sizeof(A), |
| 163 | mo: memory_order_acq_rel); |
| 164 | return true; |
| 165 | } |
| 166 | |
| 167 | bool writeCustomEvent(int32_t Delta, const void *Event, int32_t EventSize) { |
| 168 | // We write the metadata record and the custom event data into the buffer |
| 169 | // first, before we atomically update the extents for the buffer. This |
| 170 | // allows us to ensure that any threads reading the extents of the buffer |
| 171 | // will only ever see the full metadata and custom event payload accounted |
| 172 | // (no partial writes accounted). |
| 173 | MetadataRecord R = |
| 174 | createMetadataRecord<MetadataRecord::RecordKinds::CustomEventMarker>( |
| 175 | Ds&: EventSize, Ds&: Delta); |
| 176 | NextRecord = reinterpret_cast<char *>(internal_memcpy( |
| 177 | dest: NextRecord, src: reinterpret_cast<char *>(&R), n: sizeof(R))) + |
| 178 | sizeof(R); |
| 179 | NextRecord = reinterpret_cast<char *>( |
| 180 | internal_memcpy(dest: NextRecord, src: Event, n: EventSize)) + |
| 181 | EventSize; |
| 182 | |
| 183 | // We need this atomic fence here to ensure that other threads attempting to |
| 184 | // read the bytes in the buffer will see the writes committed before the |
| 185 | // extents are updated. |
| 186 | atomic_thread_fence(mo: memory_order_release); |
| 187 | atomic_fetch_add(a: Buffer.Extents, v: sizeof(R) + EventSize, |
| 188 | mo: memory_order_acq_rel); |
| 189 | return true; |
| 190 | } |
| 191 | |
| 192 | bool writeTypedEvent(int32_t Delta, uint16_t EventType, const void *Event, |
| 193 | int32_t EventSize) { |
| 194 | // We do something similar when writing out typed events, see |
| 195 | // writeCustomEvent(...) above for details. |
| 196 | MetadataRecord R = |
| 197 | createMetadataRecord<MetadataRecord::RecordKinds::TypedEventMarker>( |
| 198 | Ds&: EventSize, Ds&: Delta, Ds&: EventType); |
| 199 | NextRecord = reinterpret_cast<char *>(internal_memcpy( |
| 200 | dest: NextRecord, src: reinterpret_cast<char *>(&R), n: sizeof(R))) + |
| 201 | sizeof(R); |
| 202 | NextRecord = reinterpret_cast<char *>( |
| 203 | internal_memcpy(dest: NextRecord, src: Event, n: EventSize)) + |
| 204 | EventSize; |
| 205 | |
| 206 | // We need this atomic fence here to ensure that other threads attempting to |
| 207 | // read the bytes in the buffer will see the writes committed before the |
| 208 | // extents are updated. |
| 209 | atomic_thread_fence(mo: memory_order_release); |
| 210 | atomic_fetch_add(a: Buffer.Extents, v: EventSize, mo: memory_order_acq_rel); |
| 211 | return true; |
| 212 | } |
| 213 | |
| 214 | char *getNextRecord() const { return NextRecord; } |
| 215 | |
| 216 | void resetRecord() { |
| 217 | NextRecord = reinterpret_cast<char *>(Buffer.Data); |
| 218 | atomic_store(a: Buffer.Extents, v: 0, mo: memory_order_release); |
| 219 | } |
| 220 | |
| 221 | void undoWrites(size_t B) { |
| 222 | DCHECK_GE(NextRecord - B, reinterpret_cast<char *>(Buffer.Data)); |
| 223 | NextRecord -= B; |
| 224 | atomic_fetch_sub(a: Buffer.Extents, v: B, mo: memory_order_acq_rel); |
| 225 | } |
| 226 | |
| 227 | }; // namespace __xray |
| 228 | |
| 229 | } // namespace __xray |
| 230 | |
| 231 | #endif // COMPILER-RT_LIB_XRAY_XRAY_FDR_LOG_WRITER_H_ |
| 232 | |