1 | //===-- primary32.h ---------------------------------------------*- C++ -*-===// |
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 | #ifndef SCUDO_PRIMARY32_H_ |
10 | #define SCUDO_PRIMARY32_H_ |
11 | |
12 | #include "allocator_common.h" |
13 | #include "bytemap.h" |
14 | #include "common.h" |
15 | #include "list.h" |
16 | #include "local_cache.h" |
17 | #include "options.h" |
18 | #include "release.h" |
19 | #include "report.h" |
20 | #include "stats.h" |
21 | #include "string_utils.h" |
22 | #include "thread_annotations.h" |
23 | |
24 | namespace scudo { |
25 | |
26 | // SizeClassAllocator32 is an allocator for 32 or 64-bit address space. |
27 | // |
28 | // It maps Regions of 2^RegionSizeLog bytes aligned on a 2^RegionSizeLog bytes |
29 | // boundary, and keeps a bytemap of the mappable address space to track the size |
30 | // class they are associated with. |
31 | // |
32 | // Mapped regions are split into equally sized Blocks according to the size |
33 | // class they belong to, and the associated pointers are shuffled to prevent any |
34 | // predictable address pattern (the predictability increases with the block |
35 | // size). |
36 | // |
37 | // Regions for size class 0 are special and used to hold TransferBatches, which |
38 | // allow to transfer arrays of pointers from the global size class freelist to |
39 | // the thread specific freelist for said class, and back. |
40 | // |
41 | // Memory used by this allocator is never unmapped but can be partially |
42 | // reclaimed if the platform allows for it. |
43 | |
44 | template <typename Config> class SizeClassAllocator32 { |
45 | public: |
46 | typedef typename Config::CompactPtrT CompactPtrT; |
47 | typedef typename Config::SizeClassMap SizeClassMap; |
48 | static const uptr GroupSizeLog = Config::getGroupSizeLog(); |
49 | // The bytemap can only track UINT8_MAX - 1 classes. |
50 | static_assert(SizeClassMap::LargestClassId <= (UINT8_MAX - 1), "" ); |
51 | // Regions should be large enough to hold the largest Block. |
52 | static_assert((1UL << Config::getRegionSizeLog()) >= SizeClassMap::MaxSize, |
53 | "" ); |
54 | typedef SizeClassAllocator32<Config> ThisT; |
55 | typedef SizeClassAllocatorLocalCache<ThisT> CacheT; |
56 | typedef TransferBatch<ThisT> TransferBatchT; |
57 | typedef BatchGroup<ThisT> BatchGroupT; |
58 | |
59 | static_assert(sizeof(BatchGroupT) <= sizeof(TransferBatchT), |
60 | "BatchGroupT uses the same class size as TransferBatchT" ); |
61 | |
62 | static uptr getSizeByClassId(uptr ClassId) { |
63 | return (ClassId == SizeClassMap::BatchClassId) |
64 | ? sizeof(TransferBatchT) |
65 | : SizeClassMap::getSizeByClassId(ClassId); |
66 | } |
67 | |
68 | static bool canAllocate(uptr Size) { return Size <= SizeClassMap::MaxSize; } |
69 | |
70 | void init(s32 ReleaseToOsInterval) NO_THREAD_SAFETY_ANALYSIS { |
71 | if (SCUDO_FUCHSIA) |
72 | reportError(Message: "SizeClassAllocator32 is not supported on Fuchsia" ); |
73 | |
74 | if (SCUDO_TRUSTY) |
75 | reportError(Message: "SizeClassAllocator32 is not supported on Trusty" ); |
76 | |
77 | DCHECK(isAligned(reinterpret_cast<uptr>(this), alignof(ThisT))); |
78 | PossibleRegions.init(); |
79 | u32 Seed; |
80 | const u64 Time = getMonotonicTimeFast(); |
81 | if (!getRandom(Buffer: reinterpret_cast<void *>(&Seed), Length: sizeof(Seed))) |
82 | Seed = static_cast<u32>( |
83 | Time ^ (reinterpret_cast<uptr>(SizeClassInfoArray) >> 6)); |
84 | for (uptr I = 0; I < NumClasses; I++) { |
85 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
86 | Sci->RandState = getRandomU32(State: &Seed); |
87 | // Sci->MaxRegionIndex is already initialized to 0. |
88 | Sci->MinRegionIndex = NumRegions; |
89 | Sci->ReleaseInfo.LastReleaseAtNs = Time; |
90 | } |
91 | setOption(O: Option::ReleaseInterval, Value: static_cast<sptr>(ReleaseToOsInterval)); |
92 | } |
93 | |
94 | void unmapTestOnly() { |
95 | { |
96 | ScopedLock L(RegionsStashMutex); |
97 | while (NumberOfStashedRegions > 0) { |
98 | unmap(Addr: reinterpret_cast<void *>(RegionsStash[--NumberOfStashedRegions]), |
99 | Size: RegionSize); |
100 | } |
101 | } |
102 | |
103 | uptr MinRegionIndex = NumRegions, MaxRegionIndex = 0; |
104 | for (uptr I = 0; I < NumClasses; I++) { |
105 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
106 | ScopedLock L(Sci->Mutex); |
107 | if (Sci->MinRegionIndex < MinRegionIndex) |
108 | MinRegionIndex = Sci->MinRegionIndex; |
109 | if (Sci->MaxRegionIndex > MaxRegionIndex) |
110 | MaxRegionIndex = Sci->MaxRegionIndex; |
111 | *Sci = {}; |
112 | } |
113 | |
114 | ScopedLock L(ByteMapMutex); |
115 | for (uptr I = MinRegionIndex; I <= MaxRegionIndex; I++) |
116 | if (PossibleRegions[I]) |
117 | unmap(Addr: reinterpret_cast<void *>(I * RegionSize), Size: RegionSize); |
118 | PossibleRegions.unmapTestOnly(); |
119 | } |
120 | |
121 | // When all blocks are freed, it has to be the same size as `AllocatedUser`. |
122 | void verifyAllBlocksAreReleasedTestOnly() { |
123 | // `BatchGroup` and `TransferBatch` also use the blocks from BatchClass. |
124 | uptr BatchClassUsedInFreeLists = 0; |
125 | for (uptr I = 0; I < NumClasses; I++) { |
126 | // We have to count BatchClassUsedInFreeLists in other regions first. |
127 | if (I == SizeClassMap::BatchClassId) |
128 | continue; |
129 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
130 | ScopedLock L1(Sci->Mutex); |
131 | uptr TotalBlocks = 0; |
132 | for (BatchGroupT &BG : Sci->FreeListInfo.BlockList) { |
133 | // `BG::Batches` are `TransferBatches`. +1 for `BatchGroup`. |
134 | BatchClassUsedInFreeLists += BG.Batches.size() + 1; |
135 | for (const auto &It : BG.Batches) |
136 | TotalBlocks += It.getCount(); |
137 | } |
138 | |
139 | const uptr BlockSize = getSizeByClassId(ClassId: I); |
140 | DCHECK_EQ(TotalBlocks, Sci->AllocatedUser / BlockSize); |
141 | DCHECK_EQ(Sci->FreeListInfo.PushedBlocks, Sci->FreeListInfo.PoppedBlocks); |
142 | } |
143 | |
144 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: SizeClassMap::BatchClassId); |
145 | ScopedLock L1(Sci->Mutex); |
146 | uptr TotalBlocks = 0; |
147 | for (BatchGroupT &BG : Sci->FreeListInfo.BlockList) { |
148 | if (LIKELY(!BG.Batches.empty())) { |
149 | for (const auto &It : BG.Batches) |
150 | TotalBlocks += It.getCount(); |
151 | } else { |
152 | // `BatchGroup` with empty freelist doesn't have `TransferBatch` record |
153 | // itself. |
154 | ++TotalBlocks; |
155 | } |
156 | } |
157 | |
158 | const uptr BlockSize = getSizeByClassId(ClassId: SizeClassMap::BatchClassId); |
159 | DCHECK_EQ(TotalBlocks + BatchClassUsedInFreeLists, |
160 | Sci->AllocatedUser / BlockSize); |
161 | const uptr BlocksInUse = |
162 | Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks; |
163 | DCHECK_EQ(BlocksInUse, BatchClassUsedInFreeLists); |
164 | } |
165 | |
166 | CompactPtrT compactPtr(UNUSED uptr ClassId, uptr Ptr) const { |
167 | return static_cast<CompactPtrT>(Ptr); |
168 | } |
169 | |
170 | void *decompactPtr(UNUSED uptr ClassId, CompactPtrT CompactPtr) const { |
171 | return reinterpret_cast<void *>(static_cast<uptr>(CompactPtr)); |
172 | } |
173 | |
174 | uptr compactPtrGroupBase(CompactPtrT CompactPtr) { |
175 | const uptr Mask = (static_cast<uptr>(1) << GroupSizeLog) - 1; |
176 | return CompactPtr & ~Mask; |
177 | } |
178 | |
179 | uptr decompactGroupBase(uptr CompactPtrGroupBase) { |
180 | return CompactPtrGroupBase; |
181 | } |
182 | |
183 | ALWAYS_INLINE static bool isSmallBlock(uptr BlockSize) { |
184 | const uptr PageSize = getPageSizeCached(); |
185 | return BlockSize < PageSize / 16U; |
186 | } |
187 | |
188 | ALWAYS_INLINE static bool isLargeBlock(uptr BlockSize) { |
189 | const uptr PageSize = getPageSizeCached(); |
190 | return BlockSize > PageSize; |
191 | } |
192 | |
193 | u16 popBlocks(CacheT *C, uptr ClassId, CompactPtrT *ToArray, |
194 | const u16 MaxBlockCount) { |
195 | DCHECK_LT(ClassId, NumClasses); |
196 | SizeClassInfo *Sci = getSizeClassInfo(ClassId); |
197 | ScopedLock L(Sci->Mutex); |
198 | |
199 | u16 PopCount = popBlocksImpl(C, ClassId, Sci, ToArray, MaxBlockCount); |
200 | if (UNLIKELY(PopCount == 0)) { |
201 | if (UNLIKELY(!populateFreeList(C, ClassId, Sci))) |
202 | return 0U; |
203 | PopCount = popBlocksImpl(C, ClassId, Sci, ToArray, MaxBlockCount); |
204 | DCHECK_NE(PopCount, 0U); |
205 | } |
206 | |
207 | return PopCount; |
208 | } |
209 | |
210 | // Push the array of free blocks to the designated batch group. |
211 | void pushBlocks(CacheT *C, uptr ClassId, CompactPtrT *Array, u32 Size) { |
212 | DCHECK_LT(ClassId, NumClasses); |
213 | DCHECK_GT(Size, 0); |
214 | |
215 | SizeClassInfo *Sci = getSizeClassInfo(ClassId); |
216 | if (ClassId == SizeClassMap::BatchClassId) { |
217 | ScopedLock L(Sci->Mutex); |
218 | pushBatchClassBlocks(Sci, Array, Size); |
219 | return; |
220 | } |
221 | |
222 | // TODO(chiahungduan): Consider not doing grouping if the group size is not |
223 | // greater than the block size with a certain scale. |
224 | |
225 | // Sort the blocks so that blocks belonging to the same group can be pushed |
226 | // together. |
227 | bool SameGroup = true; |
228 | for (u32 I = 1; I < Size; ++I) { |
229 | if (compactPtrGroupBase(CompactPtr: Array[I - 1]) != compactPtrGroupBase(CompactPtr: Array[I])) |
230 | SameGroup = false; |
231 | CompactPtrT Cur = Array[I]; |
232 | u32 J = I; |
233 | while (J > 0 && |
234 | compactPtrGroupBase(CompactPtr: Cur) < compactPtrGroupBase(CompactPtr: Array[J - 1])) { |
235 | Array[J] = Array[J - 1]; |
236 | --J; |
237 | } |
238 | Array[J] = Cur; |
239 | } |
240 | |
241 | ScopedLock L(Sci->Mutex); |
242 | pushBlocksImpl(C, ClassId, Sci, Array, Size, SameGroup); |
243 | } |
244 | |
245 | void disable() NO_THREAD_SAFETY_ANALYSIS { |
246 | // The BatchClassId must be locked last since other classes can use it. |
247 | for (sptr I = static_cast<sptr>(NumClasses) - 1; I >= 0; I--) { |
248 | if (static_cast<uptr>(I) == SizeClassMap::BatchClassId) |
249 | continue; |
250 | getSizeClassInfo(ClassId: static_cast<uptr>(I))->Mutex.lock(); |
251 | } |
252 | getSizeClassInfo(ClassId: SizeClassMap::BatchClassId)->Mutex.lock(); |
253 | RegionsStashMutex.lock(); |
254 | ByteMapMutex.lock(); |
255 | } |
256 | |
257 | void enable() NO_THREAD_SAFETY_ANALYSIS { |
258 | ByteMapMutex.unlock(); |
259 | RegionsStashMutex.unlock(); |
260 | getSizeClassInfo(ClassId: SizeClassMap::BatchClassId)->Mutex.unlock(); |
261 | for (uptr I = 0; I < NumClasses; I++) { |
262 | if (I == SizeClassMap::BatchClassId) |
263 | continue; |
264 | getSizeClassInfo(ClassId: I)->Mutex.unlock(); |
265 | } |
266 | } |
267 | |
268 | template <typename F> void iterateOverBlocks(F Callback) { |
269 | uptr MinRegionIndex = NumRegions, MaxRegionIndex = 0; |
270 | for (uptr I = 0; I < NumClasses; I++) { |
271 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
272 | // TODO: The call of `iterateOverBlocks` requires disabling |
273 | // SizeClassAllocator32. We may consider locking each region on demand |
274 | // only. |
275 | Sci->Mutex.assertHeld(); |
276 | if (Sci->MinRegionIndex < MinRegionIndex) |
277 | MinRegionIndex = Sci->MinRegionIndex; |
278 | if (Sci->MaxRegionIndex > MaxRegionIndex) |
279 | MaxRegionIndex = Sci->MaxRegionIndex; |
280 | } |
281 | |
282 | // SizeClassAllocator32 is disabled, i.e., ByteMapMutex is held. |
283 | ByteMapMutex.assertHeld(); |
284 | |
285 | for (uptr I = MinRegionIndex; I <= MaxRegionIndex; I++) { |
286 | if (PossibleRegions[I] && |
287 | (PossibleRegions[I] - 1U) != SizeClassMap::BatchClassId) { |
288 | const uptr BlockSize = getSizeByClassId(ClassId: PossibleRegions[I] - 1U); |
289 | const uptr From = I * RegionSize; |
290 | const uptr To = From + (RegionSize / BlockSize) * BlockSize; |
291 | for (uptr Block = From; Block < To; Block += BlockSize) |
292 | Callback(Block); |
293 | } |
294 | } |
295 | } |
296 | |
297 | void getStats(ScopedString *Str) { |
298 | // TODO(kostyak): get the RSS per region. |
299 | uptr TotalMapped = 0; |
300 | uptr PoppedBlocks = 0; |
301 | uptr PushedBlocks = 0; |
302 | for (uptr I = 0; I < NumClasses; I++) { |
303 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
304 | ScopedLock L(Sci->Mutex); |
305 | TotalMapped += Sci->AllocatedUser; |
306 | PoppedBlocks += Sci->FreeListInfo.PoppedBlocks; |
307 | PushedBlocks += Sci->FreeListInfo.PushedBlocks; |
308 | } |
309 | Str->append(Format: "Stats: SizeClassAllocator32: %zuM mapped in %zu allocations; " |
310 | "remains %zu\n" , |
311 | TotalMapped >> 20, PoppedBlocks, PoppedBlocks - PushedBlocks); |
312 | for (uptr I = 0; I < NumClasses; I++) { |
313 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
314 | ScopedLock L(Sci->Mutex); |
315 | getStats(Str, I, Sci); |
316 | } |
317 | } |
318 | |
319 | void getFragmentationInfo(ScopedString *Str) { |
320 | Str->append( |
321 | Format: "Fragmentation Stats: SizeClassAllocator32: page size = %zu bytes\n" , |
322 | getPageSizeCached()); |
323 | |
324 | for (uptr I = 1; I < NumClasses; I++) { |
325 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
326 | ScopedLock L(Sci->Mutex); |
327 | getSizeClassFragmentationInfo(Sci, ClassId: I, Str); |
328 | } |
329 | } |
330 | |
331 | bool setOption(Option O, sptr Value) { |
332 | if (O == Option::ReleaseInterval) { |
333 | const s32 Interval = Max( |
334 | Min(static_cast<s32>(Value), Config::getMaxReleaseToOsIntervalMs()), |
335 | Config::getMinReleaseToOsIntervalMs()); |
336 | atomic_store_relaxed(A: &ReleaseToOsIntervalMs, V: Interval); |
337 | return true; |
338 | } |
339 | // Not supported by the Primary, but not an error either. |
340 | return true; |
341 | } |
342 | |
343 | uptr tryReleaseToOS(uptr ClassId, ReleaseToOS ReleaseType) { |
344 | SizeClassInfo *Sci = getSizeClassInfo(ClassId); |
345 | // TODO: Once we have separate locks like primary64, we may consider using |
346 | // tryLock() as well. |
347 | ScopedLock L(Sci->Mutex); |
348 | return releaseToOSMaybe(Sci, ClassId, ReleaseType); |
349 | } |
350 | |
351 | uptr releaseToOS(ReleaseToOS ReleaseType) { |
352 | uptr TotalReleasedBytes = 0; |
353 | for (uptr I = 0; I < NumClasses; I++) { |
354 | if (I == SizeClassMap::BatchClassId) |
355 | continue; |
356 | SizeClassInfo *Sci = getSizeClassInfo(ClassId: I); |
357 | ScopedLock L(Sci->Mutex); |
358 | TotalReleasedBytes += releaseToOSMaybe(Sci, ClassId: I, ReleaseType); |
359 | } |
360 | return TotalReleasedBytes; |
361 | } |
362 | |
363 | const char *getRegionInfoArrayAddress() const { return nullptr; } |
364 | static uptr getRegionInfoArraySize() { return 0; } |
365 | |
366 | static BlockInfo findNearestBlock(UNUSED const char *RegionInfoData, |
367 | UNUSED uptr Ptr) { |
368 | return {}; |
369 | } |
370 | |
371 | AtomicOptions Options; |
372 | |
373 | private: |
374 | static const uptr NumClasses = SizeClassMap::NumClasses; |
375 | static const uptr RegionSize = 1UL << Config::getRegionSizeLog(); |
376 | static const uptr NumRegions = SCUDO_MMAP_RANGE_SIZE >> |
377 | Config::getRegionSizeLog(); |
378 | static const u32 MaxNumBatches = SCUDO_ANDROID ? 4U : 8U; |
379 | typedef FlatByteMap<NumRegions> ByteMap; |
380 | |
381 | struct ReleaseToOsInfo { |
382 | uptr BytesInFreeListAtLastCheckpoint; |
383 | uptr RangesReleased; |
384 | uptr LastReleasedBytes; |
385 | u64 LastReleaseAtNs; |
386 | }; |
387 | |
388 | struct BlocksInfo { |
389 | SinglyLinkedList<BatchGroupT> BlockList = {}; |
390 | uptr PoppedBlocks = 0; |
391 | uptr PushedBlocks = 0; |
392 | }; |
393 | |
394 | struct alignas(SCUDO_CACHE_LINE_SIZE) SizeClassInfo { |
395 | HybridMutex Mutex; |
396 | BlocksInfo FreeListInfo GUARDED_BY(Mutex); |
397 | uptr CurrentRegion GUARDED_BY(Mutex); |
398 | uptr CurrentRegionAllocated GUARDED_BY(Mutex); |
399 | u32 RandState; |
400 | uptr AllocatedUser GUARDED_BY(Mutex); |
401 | // Lowest & highest region index allocated for this size class, to avoid |
402 | // looping through the whole NumRegions. |
403 | uptr MinRegionIndex GUARDED_BY(Mutex); |
404 | uptr MaxRegionIndex GUARDED_BY(Mutex); |
405 | ReleaseToOsInfo ReleaseInfo GUARDED_BY(Mutex); |
406 | }; |
407 | static_assert(sizeof(SizeClassInfo) % SCUDO_CACHE_LINE_SIZE == 0, "" ); |
408 | |
409 | uptr computeRegionId(uptr Mem) { |
410 | const uptr Id = Mem >> Config::getRegionSizeLog(); |
411 | CHECK_LT(Id, NumRegions); |
412 | return Id; |
413 | } |
414 | |
415 | uptr allocateRegionSlow() { |
416 | uptr MapSize = 2 * RegionSize; |
417 | const uptr MapBase = reinterpret_cast<uptr>( |
418 | map(Addr: nullptr, Size: MapSize, Name: "scudo:primary" , MAP_ALLOWNOMEM)); |
419 | if (!MapBase) |
420 | return 0; |
421 | const uptr MapEnd = MapBase + MapSize; |
422 | uptr Region = MapBase; |
423 | if (isAligned(X: Region, Alignment: RegionSize)) { |
424 | ScopedLock L(RegionsStashMutex); |
425 | if (NumberOfStashedRegions < MaxStashedRegions) |
426 | RegionsStash[NumberOfStashedRegions++] = MapBase + RegionSize; |
427 | else |
428 | MapSize = RegionSize; |
429 | } else { |
430 | Region = roundUp(X: MapBase, Boundary: RegionSize); |
431 | unmap(Addr: reinterpret_cast<void *>(MapBase), Size: Region - MapBase); |
432 | MapSize = RegionSize; |
433 | } |
434 | const uptr End = Region + MapSize; |
435 | if (End != MapEnd) |
436 | unmap(Addr: reinterpret_cast<void *>(End), Size: MapEnd - End); |
437 | |
438 | DCHECK_EQ(Region % RegionSize, 0U); |
439 | static_assert(Config::getRegionSizeLog() == GroupSizeLog, |
440 | "Memory group should be the same size as Region" ); |
441 | |
442 | return Region; |
443 | } |
444 | |
445 | uptr allocateRegion(SizeClassInfo *Sci, uptr ClassId) REQUIRES(Sci->Mutex) { |
446 | DCHECK_LT(ClassId, NumClasses); |
447 | uptr Region = 0; |
448 | { |
449 | ScopedLock L(RegionsStashMutex); |
450 | if (NumberOfStashedRegions > 0) |
451 | Region = RegionsStash[--NumberOfStashedRegions]; |
452 | } |
453 | if (!Region) |
454 | Region = allocateRegionSlow(); |
455 | if (LIKELY(Region)) { |
456 | // Sci->Mutex is held by the caller, updating the Min/Max is safe. |
457 | const uptr RegionIndex = computeRegionId(Mem: Region); |
458 | if (RegionIndex < Sci->MinRegionIndex) |
459 | Sci->MinRegionIndex = RegionIndex; |
460 | if (RegionIndex > Sci->MaxRegionIndex) |
461 | Sci->MaxRegionIndex = RegionIndex; |
462 | ScopedLock L(ByteMapMutex); |
463 | PossibleRegions.set(RegionIndex, static_cast<u8>(ClassId + 1U)); |
464 | } |
465 | return Region; |
466 | } |
467 | |
468 | SizeClassInfo *getSizeClassInfo(uptr ClassId) { |
469 | DCHECK_LT(ClassId, NumClasses); |
470 | return &SizeClassInfoArray[ClassId]; |
471 | } |
472 | |
473 | void pushBatchClassBlocks(SizeClassInfo *Sci, CompactPtrT *Array, u32 Size) |
474 | REQUIRES(Sci->Mutex) { |
475 | DCHECK_EQ(Sci, getSizeClassInfo(SizeClassMap::BatchClassId)); |
476 | |
477 | // Free blocks are recorded by TransferBatch in freelist for all |
478 | // size-classes. In addition, TransferBatch is allocated from BatchClassId. |
479 | // In order not to use additional block to record the free blocks in |
480 | // BatchClassId, they are self-contained. I.e., A TransferBatch records the |
481 | // block address of itself. See the figure below: |
482 | // |
483 | // TransferBatch at 0xABCD |
484 | // +----------------------------+ |
485 | // | Free blocks' addr | |
486 | // | +------+------+------+ | |
487 | // | |0xABCD|... |... | | |
488 | // | +------+------+------+ | |
489 | // +----------------------------+ |
490 | // |
491 | // When we allocate all the free blocks in the TransferBatch, the block used |
492 | // by TransferBatch is also free for use. We don't need to recycle the |
493 | // TransferBatch. Note that the correctness is maintained by the invariant, |
494 | // |
495 | // Each popBlocks() request returns the entire TransferBatch. Returning |
496 | // part of the blocks in a TransferBatch is invalid. |
497 | // |
498 | // This ensures that TransferBatch won't leak the address itself while it's |
499 | // still holding other valid data. |
500 | // |
501 | // Besides, BatchGroup is also allocated from BatchClassId and has its |
502 | // address recorded in the TransferBatch too. To maintain the correctness, |
503 | // |
504 | // The address of BatchGroup is always recorded in the last TransferBatch |
505 | // in the freelist (also imply that the freelist should only be |
506 | // updated with push_front). Once the last TransferBatch is popped, |
507 | // the block used by BatchGroup is also free for use. |
508 | // |
509 | // With this approach, the blocks used by BatchGroup and TransferBatch are |
510 | // reusable and don't need additional space for them. |
511 | |
512 | Sci->FreeListInfo.PushedBlocks += Size; |
513 | BatchGroupT *BG = Sci->FreeListInfo.BlockList.front(); |
514 | |
515 | if (BG == nullptr) { |
516 | // Construct `BatchGroup` on the last element. |
517 | BG = reinterpret_cast<BatchGroupT *>( |
518 | decompactPtr(ClassId: SizeClassMap::BatchClassId, CompactPtr: Array[Size - 1])); |
519 | --Size; |
520 | BG->Batches.clear(); |
521 | // BatchClass hasn't enabled memory group. Use `0` to indicate there's no |
522 | // memory group here. |
523 | BG->CompactPtrGroupBase = 0; |
524 | // `BG` is also the block of BatchClassId. Note that this is different |
525 | // from `CreateGroup` in `pushBlocksImpl` |
526 | BG->PushedBlocks = 1; |
527 | BG->BytesInBGAtLastCheckpoint = 0; |
528 | BG->MaxCachedPerBatch = |
529 | CacheT::getMaxCached(getSizeByClassId(ClassId: SizeClassMap::BatchClassId)); |
530 | |
531 | Sci->FreeListInfo.BlockList.push_front(BG); |
532 | } |
533 | |
534 | if (UNLIKELY(Size == 0)) |
535 | return; |
536 | |
537 | // This happens under 2 cases. |
538 | // 1. just allocated a new `BatchGroup`. |
539 | // 2. Only 1 block is pushed when the freelist is empty. |
540 | if (BG->Batches.empty()) { |
541 | // Construct the `TransferBatch` on the last element. |
542 | TransferBatchT *TB = reinterpret_cast<TransferBatchT *>( |
543 | decompactPtr(ClassId: SizeClassMap::BatchClassId, CompactPtr: Array[Size - 1])); |
544 | TB->clear(); |
545 | // As mentioned above, addresses of `TransferBatch` and `BatchGroup` are |
546 | // recorded in the TransferBatch. |
547 | TB->add(Array[Size - 1]); |
548 | TB->add( |
549 | compactPtr(ClassId: SizeClassMap::BatchClassId, Ptr: reinterpret_cast<uptr>(BG))); |
550 | --Size; |
551 | DCHECK_EQ(BG->PushedBlocks, 1U); |
552 | // `TB` is also the block of BatchClassId. |
553 | BG->PushedBlocks += 1; |
554 | BG->Batches.push_front(TB); |
555 | } |
556 | |
557 | TransferBatchT *CurBatch = BG->Batches.front(); |
558 | DCHECK_NE(CurBatch, nullptr); |
559 | |
560 | for (u32 I = 0; I < Size;) { |
561 | u16 UnusedSlots = |
562 | static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount()); |
563 | if (UnusedSlots == 0) { |
564 | CurBatch = reinterpret_cast<TransferBatchT *>( |
565 | decompactPtr(ClassId: SizeClassMap::BatchClassId, CompactPtr: Array[I])); |
566 | CurBatch->clear(); |
567 | // Self-contained |
568 | CurBatch->add(Array[I]); |
569 | ++I; |
570 | // TODO(chiahungduan): Avoid the use of push_back() in `Batches` of |
571 | // BatchClassId. |
572 | BG->Batches.push_front(CurBatch); |
573 | UnusedSlots = static_cast<u16>(BG->MaxCachedPerBatch - 1); |
574 | } |
575 | // `UnusedSlots` is u16 so the result will be also fit in u16. |
576 | const u16 AppendSize = static_cast<u16>(Min<u32>(A: UnusedSlots, B: Size - I)); |
577 | CurBatch->appendFromArray(&Array[I], AppendSize); |
578 | I += AppendSize; |
579 | } |
580 | |
581 | BG->PushedBlocks += Size; |
582 | } |
583 | // Push the blocks to their batch group. The layout will be like, |
584 | // |
585 | // FreeListInfo.BlockList - > BG -> BG -> BG |
586 | // | | | |
587 | // v v v |
588 | // TB TB TB |
589 | // | |
590 | // v |
591 | // TB |
592 | // |
593 | // Each BlockGroup(BG) will associate with unique group id and the free blocks |
594 | // are managed by a list of TransferBatch(TB). To reduce the time of inserting |
595 | // blocks, BGs are sorted and the input `Array` are supposed to be sorted so |
596 | // that we can get better performance of maintaining sorted property. |
597 | // Use `SameGroup=true` to indicate that all blocks in the array are from the |
598 | // same group then we will skip checking the group id of each block. |
599 | // |
600 | // The region mutex needs to be held while calling this method. |
601 | void pushBlocksImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci, |
602 | CompactPtrT *Array, u32 Size, bool SameGroup = false) |
603 | REQUIRES(Sci->Mutex) { |
604 | DCHECK_NE(ClassId, SizeClassMap::BatchClassId); |
605 | DCHECK_GT(Size, 0U); |
606 | |
607 | auto CreateGroup = [&](uptr CompactPtrGroupBase) { |
608 | BatchGroupT *BG = |
609 | reinterpret_cast<BatchGroupT *>(C->getBatchClassBlock()); |
610 | BG->Batches.clear(); |
611 | TransferBatchT *TB = |
612 | reinterpret_cast<TransferBatchT *>(C->getBatchClassBlock()); |
613 | TB->clear(); |
614 | |
615 | BG->CompactPtrGroupBase = CompactPtrGroupBase; |
616 | BG->Batches.push_front(TB); |
617 | BG->PushedBlocks = 0; |
618 | BG->BytesInBGAtLastCheckpoint = 0; |
619 | BG->MaxCachedPerBatch = TransferBatchT::MaxNumCached; |
620 | |
621 | return BG; |
622 | }; |
623 | |
624 | auto InsertBlocks = [&](BatchGroupT *BG, CompactPtrT *Array, u32 Size) { |
625 | SinglyLinkedList<TransferBatchT> &Batches = BG->Batches; |
626 | TransferBatchT *CurBatch = Batches.front(); |
627 | DCHECK_NE(CurBatch, nullptr); |
628 | |
629 | for (u32 I = 0; I < Size;) { |
630 | DCHECK_GE(BG->MaxCachedPerBatch, CurBatch->getCount()); |
631 | u16 UnusedSlots = |
632 | static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount()); |
633 | if (UnusedSlots == 0) { |
634 | CurBatch = |
635 | reinterpret_cast<TransferBatchT *>(C->getBatchClassBlock()); |
636 | CurBatch->clear(); |
637 | Batches.push_front(CurBatch); |
638 | UnusedSlots = BG->MaxCachedPerBatch; |
639 | } |
640 | // `UnusedSlots` is u16 so the result will be also fit in u16. |
641 | u16 AppendSize = static_cast<u16>(Min<u32>(A: UnusedSlots, B: Size - I)); |
642 | CurBatch->appendFromArray(&Array[I], AppendSize); |
643 | I += AppendSize; |
644 | } |
645 | |
646 | BG->PushedBlocks += Size; |
647 | }; |
648 | |
649 | Sci->FreeListInfo.PushedBlocks += Size; |
650 | BatchGroupT *Cur = Sci->FreeListInfo.BlockList.front(); |
651 | |
652 | // In the following, `Cur` always points to the BatchGroup for blocks that |
653 | // will be pushed next. `Prev` is the element right before `Cur`. |
654 | BatchGroupT *Prev = nullptr; |
655 | |
656 | while (Cur != nullptr && |
657 | compactPtrGroupBase(CompactPtr: Array[0]) > Cur->CompactPtrGroupBase) { |
658 | Prev = Cur; |
659 | Cur = Cur->Next; |
660 | } |
661 | |
662 | if (Cur == nullptr || |
663 | compactPtrGroupBase(CompactPtr: Array[0]) != Cur->CompactPtrGroupBase) { |
664 | Cur = CreateGroup(compactPtrGroupBase(CompactPtr: Array[0])); |
665 | if (Prev == nullptr) |
666 | Sci->FreeListInfo.BlockList.push_front(Cur); |
667 | else |
668 | Sci->FreeListInfo.BlockList.insert(Prev, Cur); |
669 | } |
670 | |
671 | // All the blocks are from the same group, just push without checking group |
672 | // id. |
673 | if (SameGroup) { |
674 | for (u32 I = 0; I < Size; ++I) |
675 | DCHECK_EQ(compactPtrGroupBase(Array[I]), Cur->CompactPtrGroupBase); |
676 | |
677 | InsertBlocks(Cur, Array, Size); |
678 | return; |
679 | } |
680 | |
681 | // The blocks are sorted by group id. Determine the segment of group and |
682 | // push them to their group together. |
683 | u32 Count = 1; |
684 | for (u32 I = 1; I < Size; ++I) { |
685 | if (compactPtrGroupBase(CompactPtr: Array[I - 1]) != compactPtrGroupBase(CompactPtr: Array[I])) { |
686 | DCHECK_EQ(compactPtrGroupBase(Array[I - 1]), Cur->CompactPtrGroupBase); |
687 | InsertBlocks(Cur, Array + I - Count, Count); |
688 | |
689 | while (Cur != nullptr && |
690 | compactPtrGroupBase(CompactPtr: Array[I]) > Cur->CompactPtrGroupBase) { |
691 | Prev = Cur; |
692 | Cur = Cur->Next; |
693 | } |
694 | |
695 | if (Cur == nullptr || |
696 | compactPtrGroupBase(CompactPtr: Array[I]) != Cur->CompactPtrGroupBase) { |
697 | Cur = CreateGroup(compactPtrGroupBase(CompactPtr: Array[I])); |
698 | DCHECK_NE(Prev, nullptr); |
699 | Sci->FreeListInfo.BlockList.insert(Prev, Cur); |
700 | } |
701 | |
702 | Count = 1; |
703 | } else { |
704 | ++Count; |
705 | } |
706 | } |
707 | |
708 | InsertBlocks(Cur, Array + Size - Count, Count); |
709 | } |
710 | |
711 | u16 popBlocksImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci, |
712 | CompactPtrT *ToArray, const u16 MaxBlockCount) |
713 | REQUIRES(Sci->Mutex) { |
714 | if (Sci->FreeListInfo.BlockList.empty()) |
715 | return 0U; |
716 | |
717 | SinglyLinkedList<TransferBatchT> &Batches = |
718 | Sci->FreeListInfo.BlockList.front()->Batches; |
719 | |
720 | if (Batches.empty()) { |
721 | DCHECK_EQ(ClassId, SizeClassMap::BatchClassId); |
722 | BatchGroupT *BG = Sci->FreeListInfo.BlockList.front(); |
723 | Sci->FreeListInfo.BlockList.pop_front(); |
724 | |
725 | // Block used by `BatchGroup` is from BatchClassId. Turn the block into |
726 | // `TransferBatch` with single block. |
727 | TransferBatchT *TB = reinterpret_cast<TransferBatchT *>(BG); |
728 | ToArray[0] = |
729 | compactPtr(ClassId: SizeClassMap::BatchClassId, Ptr: reinterpret_cast<uptr>(TB)); |
730 | Sci->FreeListInfo.PoppedBlocks += 1; |
731 | return 1U; |
732 | } |
733 | |
734 | // So far, instead of always filling the blocks to `MaxBlockCount`, we only |
735 | // examine single `TransferBatch` to minimize the time spent on the primary |
736 | // allocator. Besides, the sizes of `TransferBatch` and |
737 | // `CacheT::getMaxCached()` may also impact the time spent on accessing the |
738 | // primary allocator. |
739 | // TODO(chiahungduan): Evaluate if we want to always prepare `MaxBlockCount` |
740 | // blocks and/or adjust the size of `TransferBatch` according to |
741 | // `CacheT::getMaxCached()`. |
742 | TransferBatchT *B = Batches.front(); |
743 | DCHECK_NE(B, nullptr); |
744 | DCHECK_GT(B->getCount(), 0U); |
745 | |
746 | // BachClassId should always take all blocks in the TransferBatch. Read the |
747 | // comment in `pushBatchClassBlocks()` for more details. |
748 | const u16 PopCount = ClassId == SizeClassMap::BatchClassId |
749 | ? B->getCount() |
750 | : Min(MaxBlockCount, B->getCount()); |
751 | B->moveNToArray(ToArray, PopCount); |
752 | |
753 | // TODO(chiahungduan): The deallocation of unused BatchClassId blocks can be |
754 | // done without holding `Mutex`. |
755 | if (B->empty()) { |
756 | Batches.pop_front(); |
757 | // `TransferBatch` of BatchClassId is self-contained, no need to |
758 | // deallocate. Read the comment in `pushBatchClassBlocks()` for more |
759 | // details. |
760 | if (ClassId != SizeClassMap::BatchClassId) |
761 | C->deallocate(SizeClassMap::BatchClassId, B); |
762 | |
763 | if (Batches.empty()) { |
764 | BatchGroupT *BG = Sci->FreeListInfo.BlockList.front(); |
765 | Sci->FreeListInfo.BlockList.pop_front(); |
766 | |
767 | // We don't keep BatchGroup with zero blocks to avoid empty-checking |
768 | // while allocating. Note that block used for constructing BatchGroup is |
769 | // recorded as free blocks in the last element of BatchGroup::Batches. |
770 | // Which means, once we pop the last TransferBatch, the block is |
771 | // implicitly deallocated. |
772 | if (ClassId != SizeClassMap::BatchClassId) |
773 | C->deallocate(SizeClassMap::BatchClassId, BG); |
774 | } |
775 | } |
776 | |
777 | Sci->FreeListInfo.PoppedBlocks += PopCount; |
778 | return PopCount; |
779 | } |
780 | |
781 | NOINLINE bool populateFreeList(CacheT *C, uptr ClassId, SizeClassInfo *Sci) |
782 | REQUIRES(Sci->Mutex) { |
783 | uptr Region; |
784 | uptr Offset; |
785 | // If the size-class currently has a region associated to it, use it. The |
786 | // newly created blocks will be located after the currently allocated memory |
787 | // for that region (up to RegionSize). Otherwise, create a new region, where |
788 | // the new blocks will be carved from the beginning. |
789 | if (Sci->CurrentRegion) { |
790 | Region = Sci->CurrentRegion; |
791 | DCHECK_GT(Sci->CurrentRegionAllocated, 0U); |
792 | Offset = Sci->CurrentRegionAllocated; |
793 | } else { |
794 | DCHECK_EQ(Sci->CurrentRegionAllocated, 0U); |
795 | Region = allocateRegion(Sci, ClassId); |
796 | if (UNLIKELY(!Region)) |
797 | return false; |
798 | C->getStats().add(StatMapped, RegionSize); |
799 | Sci->CurrentRegion = Region; |
800 | Offset = 0; |
801 | } |
802 | |
803 | const uptr Size = getSizeByClassId(ClassId); |
804 | const u16 MaxCount = CacheT::getMaxCached(Size); |
805 | DCHECK_GT(MaxCount, 0U); |
806 | // The maximum number of blocks we should carve in the region is dictated |
807 | // by the maximum number of batches we want to fill, and the amount of |
808 | // memory left in the current region (we use the lowest of the two). This |
809 | // will not be 0 as we ensure that a region can at least hold one block (via |
810 | // static_assert and at the end of this function). |
811 | const u32 NumberOfBlocks = |
812 | Min(A: MaxNumBatches * MaxCount, |
813 | B: static_cast<u32>((RegionSize - Offset) / Size)); |
814 | DCHECK_GT(NumberOfBlocks, 0U); |
815 | |
816 | constexpr u32 ShuffleArraySize = |
817 | MaxNumBatches * TransferBatchT::MaxNumCached; |
818 | // Fill the transfer batches and put them in the size-class freelist. We |
819 | // need to randomize the blocks for security purposes, so we first fill a |
820 | // local array that we then shuffle before populating the batches. |
821 | CompactPtrT ShuffleArray[ShuffleArraySize]; |
822 | DCHECK_LE(NumberOfBlocks, ShuffleArraySize); |
823 | |
824 | uptr P = Region + Offset; |
825 | for (u32 I = 0; I < NumberOfBlocks; I++, P += Size) |
826 | ShuffleArray[I] = reinterpret_cast<CompactPtrT>(P); |
827 | |
828 | if (ClassId != SizeClassMap::BatchClassId) { |
829 | u32 N = 1; |
830 | uptr CurGroup = compactPtrGroupBase(CompactPtr: ShuffleArray[0]); |
831 | for (u32 I = 1; I < NumberOfBlocks; I++) { |
832 | if (UNLIKELY(compactPtrGroupBase(ShuffleArray[I]) != CurGroup)) { |
833 | shuffle(ShuffleArray + I - N, N, &Sci->RandState); |
834 | pushBlocksImpl(C, ClassId, Sci, Array: ShuffleArray + I - N, Size: N, |
835 | /*SameGroup=*/SameGroup: true); |
836 | N = 1; |
837 | CurGroup = compactPtrGroupBase(CompactPtr: ShuffleArray[I]); |
838 | } else { |
839 | ++N; |
840 | } |
841 | } |
842 | |
843 | shuffle(ShuffleArray + NumberOfBlocks - N, N, &Sci->RandState); |
844 | pushBlocksImpl(C, ClassId, Sci, Array: &ShuffleArray[NumberOfBlocks - N], Size: N, |
845 | /*SameGroup=*/SameGroup: true); |
846 | } else { |
847 | pushBatchClassBlocks(Sci, Array: ShuffleArray, Size: NumberOfBlocks); |
848 | } |
849 | |
850 | // Note that `PushedBlocks` and `PoppedBlocks` are supposed to only record |
851 | // the requests from `PushBlocks` and `PopBatch` which are external |
852 | // interfaces. `populateFreeList` is the internal interface so we should set |
853 | // the values back to avoid incorrectly setting the stats. |
854 | Sci->FreeListInfo.PushedBlocks -= NumberOfBlocks; |
855 | |
856 | const uptr AllocatedUser = Size * NumberOfBlocks; |
857 | C->getStats().add(StatFree, AllocatedUser); |
858 | DCHECK_LE(Sci->CurrentRegionAllocated + AllocatedUser, RegionSize); |
859 | // If there is not enough room in the region currently associated to fit |
860 | // more blocks, we deassociate the region by resetting CurrentRegion and |
861 | // CurrentRegionAllocated. Otherwise, update the allocated amount. |
862 | if (RegionSize - (Sci->CurrentRegionAllocated + AllocatedUser) < Size) { |
863 | Sci->CurrentRegion = 0; |
864 | Sci->CurrentRegionAllocated = 0; |
865 | } else { |
866 | Sci->CurrentRegionAllocated += AllocatedUser; |
867 | } |
868 | Sci->AllocatedUser += AllocatedUser; |
869 | |
870 | return true; |
871 | } |
872 | |
873 | void getStats(ScopedString *Str, uptr ClassId, SizeClassInfo *Sci) |
874 | REQUIRES(Sci->Mutex) { |
875 | if (Sci->AllocatedUser == 0) |
876 | return; |
877 | const uptr BlockSize = getSizeByClassId(ClassId); |
878 | const uptr InUse = |
879 | Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks; |
880 | const uptr BytesInFreeList = Sci->AllocatedUser - InUse * BlockSize; |
881 | uptr PushedBytesDelta = 0; |
882 | if (BytesInFreeList >= Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint) { |
883 | PushedBytesDelta = |
884 | BytesInFreeList - Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint; |
885 | } |
886 | const uptr AvailableChunks = Sci->AllocatedUser / BlockSize; |
887 | Str->append(Format: " %02zu (%6zu): mapped: %6zuK popped: %7zu pushed: %7zu " |
888 | "inuse: %6zu avail: %6zu releases: %6zu last released: %6zuK " |
889 | "latest pushed bytes: %6zuK\n" , |
890 | ClassId, getSizeByClassId(ClassId), Sci->AllocatedUser >> 10, |
891 | Sci->FreeListInfo.PoppedBlocks, Sci->FreeListInfo.PushedBlocks, |
892 | InUse, AvailableChunks, Sci->ReleaseInfo.RangesReleased, |
893 | Sci->ReleaseInfo.LastReleasedBytes >> 10, |
894 | PushedBytesDelta >> 10); |
895 | } |
896 | |
897 | void getSizeClassFragmentationInfo(SizeClassInfo *Sci, uptr ClassId, |
898 | ScopedString *Str) REQUIRES(Sci->Mutex) { |
899 | const uptr BlockSize = getSizeByClassId(ClassId); |
900 | const uptr First = Sci->MinRegionIndex; |
901 | const uptr Last = Sci->MaxRegionIndex; |
902 | const uptr Base = First * RegionSize; |
903 | const uptr NumberOfRegions = Last - First + 1U; |
904 | auto SkipRegion = [this, First, ClassId](uptr RegionIndex) { |
905 | ScopedLock L(ByteMapMutex); |
906 | return (PossibleRegions[First + RegionIndex] - 1U) != ClassId; |
907 | }; |
908 | |
909 | FragmentationRecorder Recorder; |
910 | if (!Sci->FreeListInfo.BlockList.empty()) { |
911 | PageReleaseContext Context = |
912 | markFreeBlocks(Sci, ClassId, BlockSize, Base, NumberOfRegions, |
913 | ReleaseType: ReleaseToOS::ForceAll); |
914 | releaseFreeMemoryToOS(Context, Recorder, SkipRegion); |
915 | } |
916 | |
917 | const uptr PageSize = getPageSizeCached(); |
918 | const uptr TotalBlocks = Sci->AllocatedUser / BlockSize; |
919 | const uptr InUseBlocks = |
920 | Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks; |
921 | uptr AllocatedPagesCount = 0; |
922 | if (TotalBlocks != 0U) { |
923 | for (uptr I = 0; I < NumberOfRegions; ++I) { |
924 | if (SkipRegion(I)) |
925 | continue; |
926 | AllocatedPagesCount += RegionSize / PageSize; |
927 | } |
928 | |
929 | DCHECK_NE(AllocatedPagesCount, 0U); |
930 | } |
931 | |
932 | DCHECK_GE(AllocatedPagesCount, Recorder.getReleasedPagesCount()); |
933 | const uptr InUsePages = |
934 | AllocatedPagesCount - Recorder.getReleasedPagesCount(); |
935 | const uptr InUseBytes = InUsePages * PageSize; |
936 | |
937 | uptr Integral; |
938 | uptr Fractional; |
939 | computePercentage(Numerator: BlockSize * InUseBlocks, Denominator: InUsePages * PageSize, Integral: &Integral, |
940 | Fractional: &Fractional); |
941 | Str->append(Format: " %02zu (%6zu): inuse/total blocks: %6zu/%6zu inuse/total " |
942 | "pages: %6zu/%6zu inuse bytes: %6zuK util: %3zu.%02zu%%\n" , |
943 | ClassId, BlockSize, InUseBlocks, TotalBlocks, InUsePages, |
944 | AllocatedPagesCount, InUseBytes >> 10, Integral, Fractional); |
945 | } |
946 | |
947 | NOINLINE uptr releaseToOSMaybe(SizeClassInfo *Sci, uptr ClassId, |
948 | ReleaseToOS ReleaseType = ReleaseToOS::Normal) |
949 | REQUIRES(Sci->Mutex) { |
950 | const uptr BlockSize = getSizeByClassId(ClassId); |
951 | |
952 | DCHECK_GE(Sci->FreeListInfo.PoppedBlocks, Sci->FreeListInfo.PushedBlocks); |
953 | const uptr BytesInFreeList = |
954 | Sci->AllocatedUser - |
955 | (Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks) * |
956 | BlockSize; |
957 | |
958 | if (UNLIKELY(BytesInFreeList == 0)) |
959 | return 0; |
960 | |
961 | // ====================================================================== // |
962 | // 1. Check if we have enough free blocks and if it's worth doing a page |
963 | // release. |
964 | // ====================================================================== // |
965 | if (ReleaseType != ReleaseToOS::ForceAll && |
966 | !hasChanceToReleasePages(Sci, BlockSize, BytesInFreeList, |
967 | ReleaseType)) { |
968 | return 0; |
969 | } |
970 | |
971 | const uptr First = Sci->MinRegionIndex; |
972 | const uptr Last = Sci->MaxRegionIndex; |
973 | DCHECK_NE(Last, 0U); |
974 | DCHECK_LE(First, Last); |
975 | uptr TotalReleasedBytes = 0; |
976 | const uptr Base = First * RegionSize; |
977 | const uptr NumberOfRegions = Last - First + 1U; |
978 | |
979 | // ==================================================================== // |
980 | // 2. Mark the free blocks and we can tell which pages are in-use by |
981 | // querying `PageReleaseContext`. |
982 | // ==================================================================== // |
983 | PageReleaseContext Context = markFreeBlocks(Sci, ClassId, BlockSize, Base, |
984 | NumberOfRegions, ReleaseType); |
985 | if (!Context.hasBlockMarked()) |
986 | return 0; |
987 | |
988 | // ==================================================================== // |
989 | // 3. Release the unused physical pages back to the OS. |
990 | // ==================================================================== // |
991 | ReleaseRecorder Recorder(Base); |
992 | auto SkipRegion = [this, First, ClassId](uptr RegionIndex) { |
993 | ScopedLock L(ByteMapMutex); |
994 | return (PossibleRegions[First + RegionIndex] - 1U) != ClassId; |
995 | }; |
996 | releaseFreeMemoryToOS(Context, Recorder, SkipRegion); |
997 | |
998 | if (Recorder.getReleasedRangesCount() > 0) { |
999 | Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint = BytesInFreeList; |
1000 | Sci->ReleaseInfo.RangesReleased += Recorder.getReleasedRangesCount(); |
1001 | Sci->ReleaseInfo.LastReleasedBytes = Recorder.getReleasedBytes(); |
1002 | TotalReleasedBytes += Sci->ReleaseInfo.LastReleasedBytes; |
1003 | } |
1004 | Sci->ReleaseInfo.LastReleaseAtNs = getMonotonicTimeFast(); |
1005 | |
1006 | return TotalReleasedBytes; |
1007 | } |
1008 | |
1009 | bool hasChanceToReleasePages(SizeClassInfo *Sci, uptr BlockSize, |
1010 | uptr BytesInFreeList, ReleaseToOS ReleaseType) |
1011 | REQUIRES(Sci->Mutex) { |
1012 | DCHECK_GE(Sci->FreeListInfo.PoppedBlocks, Sci->FreeListInfo.PushedBlocks); |
1013 | const uptr PageSize = getPageSizeCached(); |
1014 | |
1015 | if (BytesInFreeList <= Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint) |
1016 | Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint = BytesInFreeList; |
1017 | |
1018 | // Always update `BytesInFreeListAtLastCheckpoint` with the smallest value |
1019 | // so that we won't underestimate the releasable pages. For example, the |
1020 | // following is the region usage, |
1021 | // |
1022 | // BytesInFreeListAtLastCheckpoint AllocatedUser |
1023 | // v v |
1024 | // |---------------------------------------> |
1025 | // ^ ^ |
1026 | // BytesInFreeList ReleaseThreshold |
1027 | // |
1028 | // In general, if we have collected enough bytes and the amount of free |
1029 | // bytes meets the ReleaseThreshold, we will try to do page release. If we |
1030 | // don't update `BytesInFreeListAtLastCheckpoint` when the current |
1031 | // `BytesInFreeList` is smaller, we may take longer time to wait for enough |
1032 | // freed blocks because we miss the bytes between |
1033 | // (BytesInFreeListAtLastCheckpoint - BytesInFreeList). |
1034 | const uptr PushedBytesDelta = |
1035 | BytesInFreeList - Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint; |
1036 | if (PushedBytesDelta < PageSize) |
1037 | return false; |
1038 | |
1039 | // Releasing smaller blocks is expensive, so we want to make sure that a |
1040 | // significant amount of bytes are free, and that there has been a good |
1041 | // amount of batches pushed to the freelist before attempting to release. |
1042 | if (isSmallBlock(BlockSize) && ReleaseType == ReleaseToOS::Normal) |
1043 | if (PushedBytesDelta < Sci->AllocatedUser / 16U) |
1044 | return false; |
1045 | |
1046 | if (ReleaseType == ReleaseToOS::Normal) { |
1047 | const s32 IntervalMs = atomic_load_relaxed(A: &ReleaseToOsIntervalMs); |
1048 | if (IntervalMs < 0) |
1049 | return false; |
1050 | |
1051 | // The constant 8 here is selected from profiling some apps and the number |
1052 | // of unreleased pages in the large size classes is around 16 pages or |
1053 | // more. Choose half of it as a heuristic and which also avoids page |
1054 | // release every time for every pushBlocks() attempt by large blocks. |
1055 | const bool ByPassReleaseInterval = |
1056 | isLargeBlock(BlockSize) && PushedBytesDelta > 8 * PageSize; |
1057 | if (!ByPassReleaseInterval) { |
1058 | if (Sci->ReleaseInfo.LastReleaseAtNs + |
1059 | static_cast<u64>(IntervalMs) * 1000000 > |
1060 | getMonotonicTimeFast()) { |
1061 | // Memory was returned recently. |
1062 | return false; |
1063 | } |
1064 | } |
1065 | } // if (ReleaseType == ReleaseToOS::Normal) |
1066 | |
1067 | return true; |
1068 | } |
1069 | |
1070 | PageReleaseContext markFreeBlocks(SizeClassInfo *Sci, const uptr ClassId, |
1071 | const uptr BlockSize, const uptr Base, |
1072 | const uptr NumberOfRegions, |
1073 | ReleaseToOS ReleaseType) |
1074 | REQUIRES(Sci->Mutex) { |
1075 | const uptr PageSize = getPageSizeCached(); |
1076 | const uptr GroupSize = (1UL << GroupSizeLog); |
1077 | const uptr CurGroupBase = |
1078 | compactPtrGroupBase(CompactPtr: compactPtr(ClassId, Ptr: Sci->CurrentRegion)); |
1079 | |
1080 | PageReleaseContext Context(BlockSize, NumberOfRegions, |
1081 | /*ReleaseSize=*/RegionSize); |
1082 | |
1083 | auto DecompactPtr = [](CompactPtrT CompactPtr) { |
1084 | return reinterpret_cast<uptr>(CompactPtr); |
1085 | }; |
1086 | for (BatchGroupT &BG : Sci->FreeListInfo.BlockList) { |
1087 | const uptr GroupBase = decompactGroupBase(CompactPtrGroupBase: BG.CompactPtrGroupBase); |
1088 | // The `GroupSize` may not be divided by `BlockSize`, which means there is |
1089 | // an unused space at the end of Region. Exclude that space to avoid |
1090 | // unused page map entry. |
1091 | uptr AllocatedGroupSize = GroupBase == CurGroupBase |
1092 | ? Sci->CurrentRegionAllocated |
1093 | : roundDownSlow(X: GroupSize, Boundary: BlockSize); |
1094 | if (AllocatedGroupSize == 0) |
1095 | continue; |
1096 | |
1097 | // TransferBatches are pushed in front of BG.Batches. The first one may |
1098 | // not have all caches used. |
1099 | const uptr NumBlocks = (BG.Batches.size() - 1) * BG.MaxCachedPerBatch + |
1100 | BG.Batches.front()->getCount(); |
1101 | const uptr BytesInBG = NumBlocks * BlockSize; |
1102 | |
1103 | if (ReleaseType != ReleaseToOS::ForceAll) { |
1104 | if (BytesInBG <= BG.BytesInBGAtLastCheckpoint) { |
1105 | BG.BytesInBGAtLastCheckpoint = BytesInBG; |
1106 | continue; |
1107 | } |
1108 | |
1109 | const uptr PushedBytesDelta = BytesInBG - BG.BytesInBGAtLastCheckpoint; |
1110 | if (PushedBytesDelta < PageSize) |
1111 | continue; |
1112 | |
1113 | // Given the randomness property, we try to release the pages only if |
1114 | // the bytes used by free blocks exceed certain proportion of allocated |
1115 | // spaces. |
1116 | if (isSmallBlock(BlockSize) && (BytesInBG * 100U) / AllocatedGroupSize < |
1117 | (100U - 1U - BlockSize / 16U)) { |
1118 | continue; |
1119 | } |
1120 | } |
1121 | |
1122 | // TODO: Consider updating this after page release if `ReleaseRecorder` |
1123 | // can tell the released bytes in each group. |
1124 | BG.BytesInBGAtLastCheckpoint = BytesInBG; |
1125 | |
1126 | const uptr MaxContainedBlocks = AllocatedGroupSize / BlockSize; |
1127 | const uptr RegionIndex = (GroupBase - Base) / RegionSize; |
1128 | |
1129 | if (NumBlocks == MaxContainedBlocks) { |
1130 | for (const auto &It : BG.Batches) |
1131 | for (u16 I = 0; I < It.getCount(); ++I) |
1132 | DCHECK_EQ(compactPtrGroupBase(It.get(I)), BG.CompactPtrGroupBase); |
1133 | |
1134 | const uptr To = GroupBase + AllocatedGroupSize; |
1135 | Context.markRangeAsAllCounted(From: GroupBase, To, Base: GroupBase, RegionIndex, |
1136 | RegionSize: AllocatedGroupSize); |
1137 | } else { |
1138 | DCHECK_LT(NumBlocks, MaxContainedBlocks); |
1139 | |
1140 | // Note that we don't always visit blocks in each BatchGroup so that we |
1141 | // may miss the chance of releasing certain pages that cross |
1142 | // BatchGroups. |
1143 | Context.markFreeBlocksInRegion(BG.Batches, DecompactPtr, GroupBase, |
1144 | RegionIndex, AllocatedGroupSize, |
1145 | /*MayContainLastBlockInRegion=*/true); |
1146 | } |
1147 | |
1148 | // We may not be able to do the page release In a rare case that we may |
1149 | // fail on PageMap allocation. |
1150 | if (UNLIKELY(!Context.hasBlockMarked())) |
1151 | break; |
1152 | } |
1153 | |
1154 | return Context; |
1155 | } |
1156 | |
1157 | SizeClassInfo SizeClassInfoArray[NumClasses] = {}; |
1158 | |
1159 | HybridMutex ByteMapMutex; |
1160 | // Track the regions in use, 0 is unused, otherwise store ClassId + 1. |
1161 | ByteMap PossibleRegions GUARDED_BY(ByteMapMutex) = {}; |
1162 | atomic_s32 ReleaseToOsIntervalMs = {}; |
1163 | // Unless several threads request regions simultaneously from different size |
1164 | // classes, the stash rarely contains more than 1 entry. |
1165 | static constexpr uptr MaxStashedRegions = 4; |
1166 | HybridMutex RegionsStashMutex; |
1167 | uptr NumberOfStashedRegions GUARDED_BY(RegionsStashMutex) = 0; |
1168 | uptr RegionsStash[MaxStashedRegions] GUARDED_BY(RegionsStashMutex) = {}; |
1169 | }; |
1170 | |
1171 | } // namespace scudo |
1172 | |
1173 | #endif // SCUDO_PRIMARY32_H_ |
1174 | |