1 | // Copyright (c) Microsoft Corporation. All rights reserved. |
2 | // Licensed under the MIT license. |
3 | |
4 | #pragma once |
5 | |
6 | #include <atomic> |
7 | #include <cassert> |
8 | #ifdef _DEBUG |
9 | #include <cstring> |
10 | #endif |
11 | |
12 | #include "status.h" |
13 | #include "thread.h" |
14 | |
15 | /// A fast allocator intended for mostly-FIFO workloads (e.g., allocating contexts for file-I/O |
16 | /// callbacks). Each thread allocates by bumping the tail of its current segment; when it fills a |
17 | /// segment, it malloc()s a new one. Any thread frees by decrementing the allocation's segment's |
18 | /// ref count; when a (filled) segment's ref count reaches 0, we free() it. So long as the workload |
19 | /// is mostly FIFO, we don't leak memory. |
20 | |
21 | namespace FASTER { |
22 | namespace core { |
23 | |
24 | /// Internal classes and structures. |
25 | namespace lss_memory { |
26 | |
27 | /// Size of each segment (in bytes). (In experiments, a segment size of 16,000 worked well for |
28 | /// on Windows, while 8,000 worked well on Linux.) |
29 | #ifdef _WIN32 |
30 | static constexpr uint32_t kSegmentSize = 16000; |
31 | #else |
32 | static constexpr uint32_t kSegmentSize = 8000; |
33 | #endif |
34 | |
35 | /// Preserving Windows malloc() behavior, all LSS allocations are aligned to 16 bytes. |
36 | static constexpr uint32_t kBaseAlignment = 16; |
37 | |
38 | /// Header, prepended to all allocated blocks; used to find the ref count variable, to decrement it |
39 | /// when the block is freed. (The allocation size isn't needed, since LSS allocations are |
40 | /// essentially stack allocations; but _DEBUG mode includes it for the benefit of the caller.) |
41 | #ifdef _DEBUG |
42 | struct alignas(8) Header { |
43 | Header(uint32_t size_, uint32_t offset_) |
44 | : offset{ offset_ } |
45 | , size{ size_ } { |
46 | } |
47 | |
48 | /// Offset from the head of the segment allocator's buffer to the memory block. |
49 | uint32_t offset; |
50 | |
51 | /// Size of the memory block. |
52 | uint32_t size; |
53 | }; |
54 | static_assert(sizeof(Header) == 8, "Header is not 8 bytes!" ); |
55 | #else |
56 | struct { |
57 | (uint16_t offset_) |
58 | : offset{ offset_ } { |
59 | } |
60 | |
61 | /// Offset from the head of the segment allocator's buffer to the memory block. |
62 | uint16_t ; |
63 | }; |
64 | static_assert(sizeof(Header) == 2, "Header is not 2 bytes!" ); |
65 | #endif |
66 | |
67 | class ThreadAllocator; |
68 | |
69 | class SegmentState { |
70 | public: |
71 | SegmentState() |
72 | : control{ 0 } { |
73 | } |
74 | |
75 | SegmentState(uint64_t control_) |
76 | : control{ control_ } { |
77 | } |
78 | |
79 | SegmentState(uint32_t allocations_, uint32_t frees_) |
80 | : frees{ frees_ } |
81 | , allocations{ allocations_ } { |
82 | } |
83 | |
84 | union { |
85 | struct { |
86 | /// Count of memory blocks freed inside this segment. Incremented on each free. Frees can |
87 | /// take place on any thread. |
88 | uint32_t frees; |
89 | /// If this segment is sealed, then the count of memory blocks allocated inside this |
90 | /// segment. Otherwise, zero. |
91 | uint32_t allocations; |
92 | }; |
93 | /// 64-bit control field, used so that threads can read the allocation count atomically at |
94 | /// the same time they increment the free count atomically. |
95 | std::atomic<uint64_t> control; |
96 | }; |
97 | }; |
98 | static_assert(sizeof(SegmentState) == 8, "sizeof(SegmentState) != 8" ); |
99 | static_assert(kSegmentSize < UINT16_MAX / 2, "kSegmentSize too large for offset size!" ); |
100 | |
101 | /// Allocation takes place inside segments. When a segment is no longer needed, we add it to the |
102 | /// garbage list. |
103 | class SegmentAllocator { |
104 | public: |
105 | /// Offset from the head of the class to the head of its buffer_ field. |
106 | #ifdef _DEBUG |
107 | static constexpr uint32_t kBufferOffset = 8; |
108 | #else |
109 | static constexpr uint32_t kBufferOffset = 14; |
110 | #endif |
111 | |
112 | /// Initialize the segment allocator and allocate the segment. |
113 | SegmentAllocator() |
114 | : state{} { |
115 | #ifdef _DEBUG |
116 | // Debug LSS memory codes: |
117 | // - 0xBA - initialized, not allocated. |
118 | std::memset(buffer, 0xBA, kSegmentSize); |
119 | #endif |
120 | } |
121 | |
122 | /// Free the specified memory block. The block must be inside this segment! Returns true if the |
123 | /// segment was freed; otherwise, returns false. |
124 | void Free(void* bytes); |
125 | |
126 | /// Seal the segment--no more blocks will be allocated inside this segment. Returns true if the |
127 | /// segment was freed; otherwise, returns false. |
128 | void Seal(uint32_t blocks_allocated); |
129 | |
130 | private: |
131 | /// Decrement the active references count, effectively freeing one allocation. Also frees the |
132 | /// segment if (1) it is sealed and (2) its active references count is now zero. Returns true if |
133 | /// the segment was freed; otherwise, returns false. |
134 | void Free(); |
135 | |
136 | public: |
137 | /// Segment allocator state (8 bytes). |
138 | SegmentState state; |
139 | |
140 | /// Padding, as needed, so that the first user allocation, at buffer_[sizeof(Header)] is 16-byte |
141 | /// aligned. |
142 | /// (In _DEBUG builds, sizeof(Header) == 8, so we require 0 bytes padding; in release builds, |
143 | /// sizeof(Header) == 2, so we require 6 bytes padding.) |
144 | private: |
145 | #ifdef _DEBUG |
146 | #else |
147 | uint8_t padding_[6]; |
148 | #endif |
149 | |
150 | public: |
151 | /// This segment's memory. (First allocation's 8-byte Header starts at 8 (mod 16), so the |
152 | /// allocation's contents will start at 0 (mod 16), as desired.) |
153 | uint8_t buffer[kSegmentSize]; |
154 | }; |
155 | |
156 | /// Allocator for a single thread. Allocates only; frees are directed by the global allocator |
157 | /// object directly to the relevant segment allocator. |
158 | class alignas(64) ThreadAllocator { |
159 | public: |
160 | static constexpr uint32_t kCacheLineSize = 64; |
161 | |
162 | /// Initialize the thread allocator. The real work happens lazily, when Allocate() is called for |
163 | /// the first time. |
164 | ThreadAllocator() |
165 | : segment_allocator_{ nullptr } |
166 | , segment_offset_{ 0 } |
167 | , allocations_{ 0 } { |
168 | } |
169 | |
170 | /// Allocate a memory block of the specified size < kSegmentSize. If allocation fails, returns |
171 | /// nullptr. |
172 | void* Allocate(uint32_t size); |
173 | void* AllocateAligned(uint32_t size, uint32_t offset); |
174 | |
175 | private: |
176 | inline uint32_t Reserve(uint32_t block_size) { |
177 | assert(block_size <= kSegmentSize); |
178 | ++allocations_; |
179 | uint32_t result = segment_offset_; |
180 | assert(result <= kSegmentSize); |
181 | segment_offset_ += block_size; |
182 | return result; |
183 | } |
184 | |
185 | /// Segment inside which each thread's new allocations occur (pointer, 8 bytes). |
186 | SegmentAllocator* segment_allocator_; |
187 | |
188 | /// Offset, into the active segment, of the next allocation. |
189 | uint32_t segment_offset_; |
190 | |
191 | /// Number of blocks allocated inside the active segment. |
192 | uint32_t allocations_; |
193 | }; |
194 | static_assert(sizeof(ThreadAllocator) == 64, "sizeof(ThreadAllocator) != 64." ); |
195 | |
196 | } // namespace lss_memory |
197 | |
198 | /// The LSS allocator allocates memory from a log-structured store, but does not perform garbage |
199 | /// collection. Memory is allocated from segments; each segment is freed only after all of its |
200 | /// allocations have been freed. This means that if a single allocation inside a segment is still |
201 | /// alive, the entire segment is still alive. |
202 | /// The LSS allocator works well in the case where memory usage is almost FIFO. In that case, all |
203 | /// of the segment's allocations will eventually be freed, so the segment will be freed. The LSS |
204 | /// allocator is intended to replace the (synchronous) function call stack, for asynchronous |
205 | /// continuations. |
206 | class LssAllocator { |
207 | public: |
208 | /// Maximum number of threads supported. For each possible thread, we reserve an 8-byte |
209 | /// ThreadAllocator; so the memory required is 8 * (kMaxThreadCount) bytes. For each actual |
210 | /// thread, we reserve a full SegmentAllocator, of size approximately kSegmentSize. |
211 | static constexpr size_t kMaxThreadCount = Thread::kMaxNumThreads; |
212 | |
213 | /// Size of each segment (in bytes). |
214 | static constexpr uint32_t kSegmentSize = lss_memory::kSegmentSize; |
215 | |
216 | /// Preserving Windows malloc() behavior, all LSS allocations are aligned to 16 bytes. |
217 | static constexpr uint32_t kBaseAlignment = lss_memory::kBaseAlignment; |
218 | |
219 | /// Initialize the LSS allocator. The real work happens lazily, when a thread calls Allocate() |
220 | /// for the first time. |
221 | LssAllocator() { |
222 | for(size_t idx = 0; idx < kMaxThreadCount; ++idx) { |
223 | thread_allocators_[idx] = lss_memory::ThreadAllocator{}; |
224 | } |
225 | } |
226 | |
227 | /// Allocate a memory block of the specified size. Note that size must be < kSegmentSize, since |
228 | /// the allocation will take place inside a segment. The Allocate() code is ultimately single- |
229 | /// threaded, since we maintain a separate ThreadAllocator per thread, each with its own |
230 | /// SegmentAllocator. If allocation fails, returns nullptr. |
231 | void* Allocate(uint32_t size); |
232 | void* AllocateAligned(uint32_t size, uint32_t alignment); |
233 | |
234 | /// Free the specified memory block. The Free() code is thread-safe, since the Free() request is |
235 | /// always directed to the SegmentAllocator() that originally allocated the code--regardless of |
236 | /// what thread it is issued from. |
237 | void Free(void* bytes); |
238 | |
239 | private: |
240 | /// To reduce contention (and avoid needing atomic primitives in the allocation path), we |
241 | /// maintain a unique allocator per thread. |
242 | lss_memory::ThreadAllocator thread_allocators_[kMaxThreadCount]; |
243 | }; |
244 | |
245 | /// The global LSS allocator instance. |
246 | extern LssAllocator lss_allocator; |
247 | |
248 | } |
249 | } // namespace FASTER::core |
250 | |