xref: /freebsd/contrib/llvm-project/compiler-rt/lib/scudo/standalone/primary32.h (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
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 "bytemap.h"
13 #include "common.h"
14 #include "list.h"
15 #include "local_cache.h"
16 #include "options.h"
17 #include "release.h"
18 #include "report.h"
19 #include "stats.h"
20 #include "string_utils.h"
21 #include "thread_annotations.h"
22 
23 namespace scudo {
24 
25 // SizeClassAllocator32 is an allocator for 32 or 64-bit address space.
26 //
27 // It maps Regions of 2^RegionSizeLog bytes aligned on a 2^RegionSizeLog bytes
28 // boundary, and keeps a bytemap of the mappable address space to track the size
29 // class they are associated with.
30 //
31 // Mapped regions are split into equally sized Blocks according to the size
32 // class they belong to, and the associated pointers are shuffled to prevent any
33 // predictable address pattern (the predictability increases with the block
34 // size).
35 //
36 // Regions for size class 0 are special and used to hold TransferBatches, which
37 // allow to transfer arrays of pointers from the global size class freelist to
38 // the thread specific freelist for said class, and back.
39 //
40 // Memory used by this allocator is never unmapped but can be partially
41 // reclaimed if the platform allows for it.
42 
43 template <typename Config> class SizeClassAllocator32 {
44 public:
45   typedef typename Config::Primary::CompactPtrT CompactPtrT;
46   typedef typename Config::Primary::SizeClassMap SizeClassMap;
47   static const uptr GroupSizeLog = Config::Primary::GroupSizeLog;
48   // The bytemap can only track UINT8_MAX - 1 classes.
49   static_assert(SizeClassMap::LargestClassId <= (UINT8_MAX - 1), "");
50   // Regions should be large enough to hold the largest Block.
51   static_assert((1UL << Config::Primary::RegionSizeLog) >=
52                     SizeClassMap::MaxSize,
53                 "");
54   typedef SizeClassAllocator32<Config> ThisT;
55   typedef SizeClassAllocatorLocalCache<ThisT> CacheT;
56   typedef typename CacheT::TransferBatch TransferBatch;
57   typedef typename CacheT::BatchGroup BatchGroup;
58 
59   static uptr getSizeByClassId(uptr ClassId) {
60     return (ClassId == SizeClassMap::BatchClassId)
61                ? sizeof(TransferBatch)
62                : SizeClassMap::getSizeByClassId(ClassId);
63   }
64 
65   static bool canAllocate(uptr Size) { return Size <= SizeClassMap::MaxSize; }
66 
67   void init(s32 ReleaseToOsInterval) NO_THREAD_SAFETY_ANALYSIS {
68     if (SCUDO_FUCHSIA)
69       reportError("SizeClassAllocator32 is not supported on Fuchsia");
70 
71     if (SCUDO_TRUSTY)
72       reportError("SizeClassAllocator32 is not supported on Trusty");
73 
74     DCHECK(isAligned(reinterpret_cast<uptr>(this), alignof(ThisT)));
75     PossibleRegions.init();
76     u32 Seed;
77     const u64 Time = getMonotonicTimeFast();
78     if (!getRandom(reinterpret_cast<void *>(&Seed), sizeof(Seed)))
79       Seed = static_cast<u32>(
80           Time ^ (reinterpret_cast<uptr>(SizeClassInfoArray) >> 6));
81     for (uptr I = 0; I < NumClasses; I++) {
82       SizeClassInfo *Sci = getSizeClassInfo(I);
83       Sci->RandState = getRandomU32(&Seed);
84       // Sci->MaxRegionIndex is already initialized to 0.
85       Sci->MinRegionIndex = NumRegions;
86       Sci->ReleaseInfo.LastReleaseAtNs = Time;
87     }
88     setOption(Option::ReleaseInterval, static_cast<sptr>(ReleaseToOsInterval));
89   }
90 
91   void unmapTestOnly() {
92     {
93       ScopedLock L(RegionsStashMutex);
94       while (NumberOfStashedRegions > 0) {
95         unmap(reinterpret_cast<void *>(RegionsStash[--NumberOfStashedRegions]),
96               RegionSize);
97       }
98     }
99 
100     uptr MinRegionIndex = NumRegions, MaxRegionIndex = 0;
101     for (uptr I = 0; I < NumClasses; I++) {
102       SizeClassInfo *Sci = getSizeClassInfo(I);
103       ScopedLock L(Sci->Mutex);
104       if (Sci->MinRegionIndex < MinRegionIndex)
105         MinRegionIndex = Sci->MinRegionIndex;
106       if (Sci->MaxRegionIndex > MaxRegionIndex)
107         MaxRegionIndex = Sci->MaxRegionIndex;
108       *Sci = {};
109     }
110 
111     ScopedLock L(ByteMapMutex);
112     for (uptr I = MinRegionIndex; I <= MaxRegionIndex; I++)
113       if (PossibleRegions[I])
114         unmap(reinterpret_cast<void *>(I * RegionSize), RegionSize);
115     PossibleRegions.unmapTestOnly();
116   }
117 
118   // When all blocks are freed, it has to be the same size as `AllocatedUser`.
119   void verifyAllBlocksAreReleasedTestOnly() {
120     // `BatchGroup` and `TransferBatch` also use the blocks from BatchClass.
121     uptr BatchClassUsedInFreeLists = 0;
122     for (uptr I = 0; I < NumClasses; I++) {
123       // We have to count BatchClassUsedInFreeLists in other regions first.
124       if (I == SizeClassMap::BatchClassId)
125         continue;
126       SizeClassInfo *Sci = getSizeClassInfo(I);
127       ScopedLock L1(Sci->Mutex);
128       uptr TotalBlocks = 0;
129       for (BatchGroup &BG : Sci->FreeListInfo.BlockList) {
130         // `BG::Batches` are `TransferBatches`. +1 for `BatchGroup`.
131         BatchClassUsedInFreeLists += BG.Batches.size() + 1;
132         for (const auto &It : BG.Batches)
133           TotalBlocks += It.getCount();
134       }
135 
136       const uptr BlockSize = getSizeByClassId(I);
137       DCHECK_EQ(TotalBlocks, Sci->AllocatedUser / BlockSize);
138       DCHECK_EQ(Sci->FreeListInfo.PushedBlocks, Sci->FreeListInfo.PoppedBlocks);
139     }
140 
141     SizeClassInfo *Sci = getSizeClassInfo(SizeClassMap::BatchClassId);
142     ScopedLock L1(Sci->Mutex);
143     uptr TotalBlocks = 0;
144     for (BatchGroup &BG : Sci->FreeListInfo.BlockList) {
145       if (LIKELY(!BG.Batches.empty())) {
146         for (const auto &It : BG.Batches)
147           TotalBlocks += It.getCount();
148       } else {
149         // `BatchGroup` with empty freelist doesn't have `TransferBatch` record
150         // itself.
151         ++TotalBlocks;
152       }
153     }
154 
155     const uptr BlockSize = getSizeByClassId(SizeClassMap::BatchClassId);
156     DCHECK_EQ(TotalBlocks + BatchClassUsedInFreeLists,
157               Sci->AllocatedUser / BlockSize);
158     const uptr BlocksInUse =
159         Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks;
160     DCHECK_EQ(BlocksInUse, BatchClassUsedInFreeLists);
161   }
162 
163   CompactPtrT compactPtr(UNUSED uptr ClassId, uptr Ptr) const {
164     return static_cast<CompactPtrT>(Ptr);
165   }
166 
167   void *decompactPtr(UNUSED uptr ClassId, CompactPtrT CompactPtr) const {
168     return reinterpret_cast<void *>(static_cast<uptr>(CompactPtr));
169   }
170 
171   uptr compactPtrGroupBase(CompactPtrT CompactPtr) {
172     const uptr Mask = (static_cast<uptr>(1) << GroupSizeLog) - 1;
173     return CompactPtr & ~Mask;
174   }
175 
176   uptr decompactGroupBase(uptr CompactPtrGroupBase) {
177     return CompactPtrGroupBase;
178   }
179 
180   ALWAYS_INLINE static bool isSmallBlock(uptr BlockSize) {
181     const uptr PageSize = getPageSizeCached();
182     return BlockSize < PageSize / 16U;
183   }
184 
185   ALWAYS_INLINE static bool isLargeBlock(uptr BlockSize) {
186     const uptr PageSize = getPageSizeCached();
187     return BlockSize > PageSize;
188   }
189 
190   TransferBatch *popBatch(CacheT *C, uptr ClassId) {
191     DCHECK_LT(ClassId, NumClasses);
192     SizeClassInfo *Sci = getSizeClassInfo(ClassId);
193     ScopedLock L(Sci->Mutex);
194     TransferBatch *B = popBatchImpl(C, ClassId, Sci);
195     if (UNLIKELY(!B)) {
196       if (UNLIKELY(!populateFreeList(C, ClassId, Sci)))
197         return nullptr;
198       B = popBatchImpl(C, ClassId, Sci);
199       // if `populateFreeList` succeeded, we are supposed to get free blocks.
200       DCHECK_NE(B, nullptr);
201     }
202     return B;
203   }
204 
205   // Push the array of free blocks to the designated batch group.
206   void pushBlocks(CacheT *C, uptr ClassId, CompactPtrT *Array, u32 Size) {
207     DCHECK_LT(ClassId, NumClasses);
208     DCHECK_GT(Size, 0);
209 
210     SizeClassInfo *Sci = getSizeClassInfo(ClassId);
211     if (ClassId == SizeClassMap::BatchClassId) {
212       ScopedLock L(Sci->Mutex);
213       pushBatchClassBlocks(Sci, Array, Size);
214       return;
215     }
216 
217     // TODO(chiahungduan): Consider not doing grouping if the group size is not
218     // greater than the block size with a certain scale.
219 
220     // Sort the blocks so that blocks belonging to the same group can be pushed
221     // together.
222     bool SameGroup = true;
223     for (u32 I = 1; I < Size; ++I) {
224       if (compactPtrGroupBase(Array[I - 1]) != compactPtrGroupBase(Array[I]))
225         SameGroup = false;
226       CompactPtrT Cur = Array[I];
227       u32 J = I;
228       while (J > 0 &&
229              compactPtrGroupBase(Cur) < compactPtrGroupBase(Array[J - 1])) {
230         Array[J] = Array[J - 1];
231         --J;
232       }
233       Array[J] = Cur;
234     }
235 
236     ScopedLock L(Sci->Mutex);
237     pushBlocksImpl(C, ClassId, Sci, Array, Size, SameGroup);
238   }
239 
240   void disable() NO_THREAD_SAFETY_ANALYSIS {
241     // The BatchClassId must be locked last since other classes can use it.
242     for (sptr I = static_cast<sptr>(NumClasses) - 1; I >= 0; I--) {
243       if (static_cast<uptr>(I) == SizeClassMap::BatchClassId)
244         continue;
245       getSizeClassInfo(static_cast<uptr>(I))->Mutex.lock();
246     }
247     getSizeClassInfo(SizeClassMap::BatchClassId)->Mutex.lock();
248     RegionsStashMutex.lock();
249     ByteMapMutex.lock();
250   }
251 
252   void enable() NO_THREAD_SAFETY_ANALYSIS {
253     ByteMapMutex.unlock();
254     RegionsStashMutex.unlock();
255     getSizeClassInfo(SizeClassMap::BatchClassId)->Mutex.unlock();
256     for (uptr I = 0; I < NumClasses; I++) {
257       if (I == SizeClassMap::BatchClassId)
258         continue;
259       getSizeClassInfo(I)->Mutex.unlock();
260     }
261   }
262 
263   template <typename F> void iterateOverBlocks(F Callback) {
264     uptr MinRegionIndex = NumRegions, MaxRegionIndex = 0;
265     for (uptr I = 0; I < NumClasses; I++) {
266       SizeClassInfo *Sci = getSizeClassInfo(I);
267       // TODO: The call of `iterateOverBlocks` requires disabling
268       // SizeClassAllocator32. We may consider locking each region on demand
269       // only.
270       Sci->Mutex.assertHeld();
271       if (Sci->MinRegionIndex < MinRegionIndex)
272         MinRegionIndex = Sci->MinRegionIndex;
273       if (Sci->MaxRegionIndex > MaxRegionIndex)
274         MaxRegionIndex = Sci->MaxRegionIndex;
275     }
276 
277     // SizeClassAllocator32 is disabled, i.e., ByteMapMutex is held.
278     ByteMapMutex.assertHeld();
279 
280     for (uptr I = MinRegionIndex; I <= MaxRegionIndex; I++) {
281       if (PossibleRegions[I] &&
282           (PossibleRegions[I] - 1U) != SizeClassMap::BatchClassId) {
283         const uptr BlockSize = getSizeByClassId(PossibleRegions[I] - 1U);
284         const uptr From = I * RegionSize;
285         const uptr To = From + (RegionSize / BlockSize) * BlockSize;
286         for (uptr Block = From; Block < To; Block += BlockSize)
287           Callback(Block);
288       }
289     }
290   }
291 
292   void getStats(ScopedString *Str) {
293     // TODO(kostyak): get the RSS per region.
294     uptr TotalMapped = 0;
295     uptr PoppedBlocks = 0;
296     uptr PushedBlocks = 0;
297     for (uptr I = 0; I < NumClasses; I++) {
298       SizeClassInfo *Sci = getSizeClassInfo(I);
299       ScopedLock L(Sci->Mutex);
300       TotalMapped += Sci->AllocatedUser;
301       PoppedBlocks += Sci->FreeListInfo.PoppedBlocks;
302       PushedBlocks += Sci->FreeListInfo.PushedBlocks;
303     }
304     Str->append("Stats: SizeClassAllocator32: %zuM mapped in %zu allocations; "
305                 "remains %zu\n",
306                 TotalMapped >> 20, PoppedBlocks, PoppedBlocks - PushedBlocks);
307     for (uptr I = 0; I < NumClasses; I++) {
308       SizeClassInfo *Sci = getSizeClassInfo(I);
309       ScopedLock L(Sci->Mutex);
310       getStats(Str, I, Sci);
311     }
312   }
313 
314   bool setOption(Option O, sptr Value) {
315     if (O == Option::ReleaseInterval) {
316       const s32 Interval = Max(Min(static_cast<s32>(Value),
317                                    Config::Primary::MaxReleaseToOsIntervalMs),
318                                Config::Primary::MinReleaseToOsIntervalMs);
319       atomic_store_relaxed(&ReleaseToOsIntervalMs, Interval);
320       return true;
321     }
322     // Not supported by the Primary, but not an error either.
323     return true;
324   }
325 
326   uptr tryReleaseToOS(uptr ClassId, ReleaseToOS ReleaseType) {
327     SizeClassInfo *Sci = getSizeClassInfo(ClassId);
328     // TODO: Once we have separate locks like primary64, we may consider using
329     // tryLock() as well.
330     ScopedLock L(Sci->Mutex);
331     return releaseToOSMaybe(Sci, ClassId, ReleaseType);
332   }
333 
334   uptr releaseToOS(ReleaseToOS ReleaseType) {
335     uptr TotalReleasedBytes = 0;
336     for (uptr I = 0; I < NumClasses; I++) {
337       if (I == SizeClassMap::BatchClassId)
338         continue;
339       SizeClassInfo *Sci = getSizeClassInfo(I);
340       ScopedLock L(Sci->Mutex);
341       TotalReleasedBytes += releaseToOSMaybe(Sci, I, ReleaseType);
342     }
343     return TotalReleasedBytes;
344   }
345 
346   const char *getRegionInfoArrayAddress() const { return nullptr; }
347   static uptr getRegionInfoArraySize() { return 0; }
348 
349   static BlockInfo findNearestBlock(UNUSED const char *RegionInfoData,
350                                     UNUSED uptr Ptr) {
351     return {};
352   }
353 
354   AtomicOptions Options;
355 
356 private:
357   static const uptr NumClasses = SizeClassMap::NumClasses;
358   static const uptr RegionSize = 1UL << Config::Primary::RegionSizeLog;
359   static const uptr NumRegions =
360       SCUDO_MMAP_RANGE_SIZE >> Config::Primary::RegionSizeLog;
361   static const u32 MaxNumBatches = SCUDO_ANDROID ? 4U : 8U;
362   typedef FlatByteMap<NumRegions> ByteMap;
363 
364   struct ReleaseToOsInfo {
365     uptr BytesInFreeListAtLastCheckpoint;
366     uptr RangesReleased;
367     uptr LastReleasedBytes;
368     u64 LastReleaseAtNs;
369   };
370 
371   struct BlocksInfo {
372     SinglyLinkedList<BatchGroup> BlockList = {};
373     uptr PoppedBlocks = 0;
374     uptr PushedBlocks = 0;
375   };
376 
377   struct alignas(SCUDO_CACHE_LINE_SIZE) SizeClassInfo {
378     HybridMutex Mutex;
379     BlocksInfo FreeListInfo GUARDED_BY(Mutex);
380     uptr CurrentRegion GUARDED_BY(Mutex);
381     uptr CurrentRegionAllocated GUARDED_BY(Mutex);
382     u32 RandState;
383     uptr AllocatedUser GUARDED_BY(Mutex);
384     // Lowest & highest region index allocated for this size class, to avoid
385     // looping through the whole NumRegions.
386     uptr MinRegionIndex GUARDED_BY(Mutex);
387     uptr MaxRegionIndex GUARDED_BY(Mutex);
388     ReleaseToOsInfo ReleaseInfo GUARDED_BY(Mutex);
389   };
390   static_assert(sizeof(SizeClassInfo) % SCUDO_CACHE_LINE_SIZE == 0, "");
391 
392   uptr computeRegionId(uptr Mem) {
393     const uptr Id = Mem >> Config::Primary::RegionSizeLog;
394     CHECK_LT(Id, NumRegions);
395     return Id;
396   }
397 
398   uptr allocateRegionSlow() {
399     uptr MapSize = 2 * RegionSize;
400     const uptr MapBase = reinterpret_cast<uptr>(
401         map(nullptr, MapSize, "scudo:primary", MAP_ALLOWNOMEM));
402     if (!MapBase)
403       return 0;
404     const uptr MapEnd = MapBase + MapSize;
405     uptr Region = MapBase;
406     if (isAligned(Region, RegionSize)) {
407       ScopedLock L(RegionsStashMutex);
408       if (NumberOfStashedRegions < MaxStashedRegions)
409         RegionsStash[NumberOfStashedRegions++] = MapBase + RegionSize;
410       else
411         MapSize = RegionSize;
412     } else {
413       Region = roundUp(MapBase, RegionSize);
414       unmap(reinterpret_cast<void *>(MapBase), Region - MapBase);
415       MapSize = RegionSize;
416     }
417     const uptr End = Region + MapSize;
418     if (End != MapEnd)
419       unmap(reinterpret_cast<void *>(End), MapEnd - End);
420 
421     DCHECK_EQ(Region % RegionSize, 0U);
422     static_assert(Config::Primary::RegionSizeLog == GroupSizeLog,
423                   "Memory group should be the same size as Region");
424 
425     return Region;
426   }
427 
428   uptr allocateRegion(SizeClassInfo *Sci, uptr ClassId) REQUIRES(Sci->Mutex) {
429     DCHECK_LT(ClassId, NumClasses);
430     uptr Region = 0;
431     {
432       ScopedLock L(RegionsStashMutex);
433       if (NumberOfStashedRegions > 0)
434         Region = RegionsStash[--NumberOfStashedRegions];
435     }
436     if (!Region)
437       Region = allocateRegionSlow();
438     if (LIKELY(Region)) {
439       // Sci->Mutex is held by the caller, updating the Min/Max is safe.
440       const uptr RegionIndex = computeRegionId(Region);
441       if (RegionIndex < Sci->MinRegionIndex)
442         Sci->MinRegionIndex = RegionIndex;
443       if (RegionIndex > Sci->MaxRegionIndex)
444         Sci->MaxRegionIndex = RegionIndex;
445       ScopedLock L(ByteMapMutex);
446       PossibleRegions.set(RegionIndex, static_cast<u8>(ClassId + 1U));
447     }
448     return Region;
449   }
450 
451   SizeClassInfo *getSizeClassInfo(uptr ClassId) {
452     DCHECK_LT(ClassId, NumClasses);
453     return &SizeClassInfoArray[ClassId];
454   }
455 
456   void pushBatchClassBlocks(SizeClassInfo *Sci, CompactPtrT *Array, u32 Size)
457       REQUIRES(Sci->Mutex) {
458     DCHECK_EQ(Sci, getSizeClassInfo(SizeClassMap::BatchClassId));
459 
460     // Free blocks are recorded by TransferBatch in freelist for all
461     // size-classes. In addition, TransferBatch is allocated from BatchClassId.
462     // In order not to use additional block to record the free blocks in
463     // BatchClassId, they are self-contained. I.e., A TransferBatch records the
464     // block address of itself. See the figure below:
465     //
466     // TransferBatch at 0xABCD
467     // +----------------------------+
468     // | Free blocks' addr          |
469     // | +------+------+------+     |
470     // | |0xABCD|...   |...   |     |
471     // | +------+------+------+     |
472     // +----------------------------+
473     //
474     // When we allocate all the free blocks in the TransferBatch, the block used
475     // by TransferBatch is also free for use. We don't need to recycle the
476     // TransferBatch. Note that the correctness is maintained by the invariant,
477     //
478     //   The unit of each popBatch() request is entire TransferBatch. Return
479     //   part of the blocks in a TransferBatch is invalid.
480     //
481     // This ensures that TransferBatch won't leak the address itself while it's
482     // still holding other valid data.
483     //
484     // Besides, BatchGroup is also allocated from BatchClassId and has its
485     // address recorded in the TransferBatch too. To maintain the correctness,
486     //
487     //   The address of BatchGroup is always recorded in the last TransferBatch
488     //   in the freelist (also imply that the freelist should only be
489     //   updated with push_front). Once the last TransferBatch is popped,
490     //   the block used by BatchGroup is also free for use.
491     //
492     // With this approach, the blocks used by BatchGroup and TransferBatch are
493     // reusable and don't need additional space for them.
494 
495     Sci->FreeListInfo.PushedBlocks += Size;
496     BatchGroup *BG = Sci->FreeListInfo.BlockList.front();
497 
498     if (BG == nullptr) {
499       // Construct `BatchGroup` on the last element.
500       BG = reinterpret_cast<BatchGroup *>(
501           decompactPtr(SizeClassMap::BatchClassId, Array[Size - 1]));
502       --Size;
503       BG->Batches.clear();
504       // BatchClass hasn't enabled memory group. Use `0` to indicate there's no
505       // memory group here.
506       BG->CompactPtrGroupBase = 0;
507       // `BG` is also the block of BatchClassId. Note that this is different
508       // from `CreateGroup` in `pushBlocksImpl`
509       BG->PushedBlocks = 1;
510       BG->BytesInBGAtLastCheckpoint = 0;
511       BG->MaxCachedPerBatch = TransferBatch::getMaxCached(
512           getSizeByClassId(SizeClassMap::BatchClassId));
513 
514       Sci->FreeListInfo.BlockList.push_front(BG);
515     }
516 
517     if (UNLIKELY(Size == 0))
518       return;
519 
520     // This happens under 2 cases.
521     //   1. just allocated a new `BatchGroup`.
522     //   2. Only 1 block is pushed when the freelist is empty.
523     if (BG->Batches.empty()) {
524       // Construct the `TransferBatch` on the last element.
525       TransferBatch *TB = reinterpret_cast<TransferBatch *>(
526           decompactPtr(SizeClassMap::BatchClassId, Array[Size - 1]));
527       TB->clear();
528       // As mentioned above, addresses of `TransferBatch` and `BatchGroup` are
529       // recorded in the TransferBatch.
530       TB->add(Array[Size - 1]);
531       TB->add(
532           compactPtr(SizeClassMap::BatchClassId, reinterpret_cast<uptr>(BG)));
533       --Size;
534       DCHECK_EQ(BG->PushedBlocks, 1U);
535       // `TB` is also the block of BatchClassId.
536       BG->PushedBlocks += 1;
537       BG->Batches.push_front(TB);
538     }
539 
540     TransferBatch *CurBatch = BG->Batches.front();
541     DCHECK_NE(CurBatch, nullptr);
542 
543     for (u32 I = 0; I < Size;) {
544       u16 UnusedSlots =
545           static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount());
546       if (UnusedSlots == 0) {
547         CurBatch = reinterpret_cast<TransferBatch *>(
548             decompactPtr(SizeClassMap::BatchClassId, Array[I]));
549         CurBatch->clear();
550         // Self-contained
551         CurBatch->add(Array[I]);
552         ++I;
553         // TODO(chiahungduan): Avoid the use of push_back() in `Batches` of
554         // BatchClassId.
555         BG->Batches.push_front(CurBatch);
556         UnusedSlots = static_cast<u16>(BG->MaxCachedPerBatch - 1);
557       }
558       // `UnusedSlots` is u16 so the result will be also fit in u16.
559       const u16 AppendSize = static_cast<u16>(Min<u32>(UnusedSlots, Size - I));
560       CurBatch->appendFromArray(&Array[I], AppendSize);
561       I += AppendSize;
562     }
563 
564     BG->PushedBlocks += Size;
565   }
566   // Push the blocks to their batch group. The layout will be like,
567   //
568   // FreeListInfo.BlockList - > BG -> BG -> BG
569   //                            |     |     |
570   //                            v     v     v
571   //                            TB    TB    TB
572   //                            |
573   //                            v
574   //                            TB
575   //
576   // Each BlockGroup(BG) will associate with unique group id and the free blocks
577   // are managed by a list of TransferBatch(TB). To reduce the time of inserting
578   // blocks, BGs are sorted and the input `Array` are supposed to be sorted so
579   // that we can get better performance of maintaining sorted property.
580   // Use `SameGroup=true` to indicate that all blocks in the array are from the
581   // same group then we will skip checking the group id of each block.
582   //
583   // The region mutex needs to be held while calling this method.
584   void pushBlocksImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci,
585                       CompactPtrT *Array, u32 Size, bool SameGroup = false)
586       REQUIRES(Sci->Mutex) {
587     DCHECK_NE(ClassId, SizeClassMap::BatchClassId);
588     DCHECK_GT(Size, 0U);
589 
590     auto CreateGroup = [&](uptr CompactPtrGroupBase) {
591       BatchGroup *BG = C->createGroup();
592       BG->Batches.clear();
593       TransferBatch *TB = C->createBatch(ClassId, nullptr);
594       TB->clear();
595 
596       BG->CompactPtrGroupBase = CompactPtrGroupBase;
597       BG->Batches.push_front(TB);
598       BG->PushedBlocks = 0;
599       BG->BytesInBGAtLastCheckpoint = 0;
600       BG->MaxCachedPerBatch =
601           TransferBatch::getMaxCached(getSizeByClassId(ClassId));
602 
603       return BG;
604     };
605 
606     auto InsertBlocks = [&](BatchGroup *BG, CompactPtrT *Array, u32 Size) {
607       SinglyLinkedList<TransferBatch> &Batches = BG->Batches;
608       TransferBatch *CurBatch = Batches.front();
609       DCHECK_NE(CurBatch, nullptr);
610 
611       for (u32 I = 0; I < Size;) {
612         DCHECK_GE(BG->MaxCachedPerBatch, CurBatch->getCount());
613         u16 UnusedSlots =
614             static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount());
615         if (UnusedSlots == 0) {
616           CurBatch = C->createBatch(
617               ClassId,
618               reinterpret_cast<void *>(decompactPtr(ClassId, Array[I])));
619           CurBatch->clear();
620           Batches.push_front(CurBatch);
621           UnusedSlots = BG->MaxCachedPerBatch;
622         }
623         // `UnusedSlots` is u16 so the result will be also fit in u16.
624         u16 AppendSize = static_cast<u16>(Min<u32>(UnusedSlots, Size - I));
625         CurBatch->appendFromArray(&Array[I], AppendSize);
626         I += AppendSize;
627       }
628 
629       BG->PushedBlocks += Size;
630     };
631 
632     Sci->FreeListInfo.PushedBlocks += Size;
633     BatchGroup *Cur = Sci->FreeListInfo.BlockList.front();
634 
635     // In the following, `Cur` always points to the BatchGroup for blocks that
636     // will be pushed next. `Prev` is the element right before `Cur`.
637     BatchGroup *Prev = nullptr;
638 
639     while (Cur != nullptr &&
640            compactPtrGroupBase(Array[0]) > Cur->CompactPtrGroupBase) {
641       Prev = Cur;
642       Cur = Cur->Next;
643     }
644 
645     if (Cur == nullptr ||
646         compactPtrGroupBase(Array[0]) != Cur->CompactPtrGroupBase) {
647       Cur = CreateGroup(compactPtrGroupBase(Array[0]));
648       if (Prev == nullptr)
649         Sci->FreeListInfo.BlockList.push_front(Cur);
650       else
651         Sci->FreeListInfo.BlockList.insert(Prev, Cur);
652     }
653 
654     // All the blocks are from the same group, just push without checking group
655     // id.
656     if (SameGroup) {
657       for (u32 I = 0; I < Size; ++I)
658         DCHECK_EQ(compactPtrGroupBase(Array[I]), Cur->CompactPtrGroupBase);
659 
660       InsertBlocks(Cur, Array, Size);
661       return;
662     }
663 
664     // The blocks are sorted by group id. Determine the segment of group and
665     // push them to their group together.
666     u32 Count = 1;
667     for (u32 I = 1; I < Size; ++I) {
668       if (compactPtrGroupBase(Array[I - 1]) != compactPtrGroupBase(Array[I])) {
669         DCHECK_EQ(compactPtrGroupBase(Array[I - 1]), Cur->CompactPtrGroupBase);
670         InsertBlocks(Cur, Array + I - Count, Count);
671 
672         while (Cur != nullptr &&
673                compactPtrGroupBase(Array[I]) > Cur->CompactPtrGroupBase) {
674           Prev = Cur;
675           Cur = Cur->Next;
676         }
677 
678         if (Cur == nullptr ||
679             compactPtrGroupBase(Array[I]) != Cur->CompactPtrGroupBase) {
680           Cur = CreateGroup(compactPtrGroupBase(Array[I]));
681           DCHECK_NE(Prev, nullptr);
682           Sci->FreeListInfo.BlockList.insert(Prev, Cur);
683         }
684 
685         Count = 1;
686       } else {
687         ++Count;
688       }
689     }
690 
691     InsertBlocks(Cur, Array + Size - Count, Count);
692   }
693 
694   // Pop one TransferBatch from a BatchGroup. The BatchGroup with the smallest
695   // group id will be considered first.
696   //
697   // The region mutex needs to be held while calling this method.
698   TransferBatch *popBatchImpl(CacheT *C, uptr ClassId, SizeClassInfo *Sci)
699       REQUIRES(Sci->Mutex) {
700     if (Sci->FreeListInfo.BlockList.empty())
701       return nullptr;
702 
703     SinglyLinkedList<TransferBatch> &Batches =
704         Sci->FreeListInfo.BlockList.front()->Batches;
705 
706     if (Batches.empty()) {
707       DCHECK_EQ(ClassId, SizeClassMap::BatchClassId);
708       BatchGroup *BG = Sci->FreeListInfo.BlockList.front();
709       Sci->FreeListInfo.BlockList.pop_front();
710 
711       // Block used by `BatchGroup` is from BatchClassId. Turn the block into
712       // `TransferBatch` with single block.
713       TransferBatch *TB = reinterpret_cast<TransferBatch *>(BG);
714       TB->clear();
715       TB->add(
716           compactPtr(SizeClassMap::BatchClassId, reinterpret_cast<uptr>(TB)));
717       Sci->FreeListInfo.PoppedBlocks += 1;
718       return TB;
719     }
720 
721     TransferBatch *B = Batches.front();
722     Batches.pop_front();
723     DCHECK_NE(B, nullptr);
724     DCHECK_GT(B->getCount(), 0U);
725 
726     if (Batches.empty()) {
727       BatchGroup *BG = Sci->FreeListInfo.BlockList.front();
728       Sci->FreeListInfo.BlockList.pop_front();
729 
730       // We don't keep BatchGroup with zero blocks to avoid empty-checking while
731       // allocating. Note that block used by constructing BatchGroup is recorded
732       // as free blocks in the last element of BatchGroup::Batches. Which means,
733       // once we pop the last TransferBatch, the block is implicitly
734       // deallocated.
735       if (ClassId != SizeClassMap::BatchClassId)
736         C->deallocate(SizeClassMap::BatchClassId, BG);
737     }
738 
739     Sci->FreeListInfo.PoppedBlocks += B->getCount();
740     return B;
741   }
742 
743   NOINLINE bool populateFreeList(CacheT *C, uptr ClassId, SizeClassInfo *Sci)
744       REQUIRES(Sci->Mutex) {
745     uptr Region;
746     uptr Offset;
747     // If the size-class currently has a region associated to it, use it. The
748     // newly created blocks will be located after the currently allocated memory
749     // for that region (up to RegionSize). Otherwise, create a new region, where
750     // the new blocks will be carved from the beginning.
751     if (Sci->CurrentRegion) {
752       Region = Sci->CurrentRegion;
753       DCHECK_GT(Sci->CurrentRegionAllocated, 0U);
754       Offset = Sci->CurrentRegionAllocated;
755     } else {
756       DCHECK_EQ(Sci->CurrentRegionAllocated, 0U);
757       Region = allocateRegion(Sci, ClassId);
758       if (UNLIKELY(!Region))
759         return false;
760       C->getStats().add(StatMapped, RegionSize);
761       Sci->CurrentRegion = Region;
762       Offset = 0;
763     }
764 
765     const uptr Size = getSizeByClassId(ClassId);
766     const u16 MaxCount = TransferBatch::getMaxCached(Size);
767     DCHECK_GT(MaxCount, 0U);
768     // The maximum number of blocks we should carve in the region is dictated
769     // by the maximum number of batches we want to fill, and the amount of
770     // memory left in the current region (we use the lowest of the two). This
771     // will not be 0 as we ensure that a region can at least hold one block (via
772     // static_assert and at the end of this function).
773     const u32 NumberOfBlocks =
774         Min(MaxNumBatches * MaxCount,
775             static_cast<u32>((RegionSize - Offset) / Size));
776     DCHECK_GT(NumberOfBlocks, 0U);
777 
778     constexpr u32 ShuffleArraySize =
779         MaxNumBatches * TransferBatch::MaxNumCached;
780     // Fill the transfer batches and put them in the size-class freelist. We
781     // need to randomize the blocks for security purposes, so we first fill a
782     // local array that we then shuffle before populating the batches.
783     CompactPtrT ShuffleArray[ShuffleArraySize];
784     DCHECK_LE(NumberOfBlocks, ShuffleArraySize);
785 
786     uptr P = Region + Offset;
787     for (u32 I = 0; I < NumberOfBlocks; I++, P += Size)
788       ShuffleArray[I] = reinterpret_cast<CompactPtrT>(P);
789 
790     if (ClassId != SizeClassMap::BatchClassId) {
791       u32 N = 1;
792       uptr CurGroup = compactPtrGroupBase(ShuffleArray[0]);
793       for (u32 I = 1; I < NumberOfBlocks; I++) {
794         if (UNLIKELY(compactPtrGroupBase(ShuffleArray[I]) != CurGroup)) {
795           shuffle(ShuffleArray + I - N, N, &Sci->RandState);
796           pushBlocksImpl(C, ClassId, Sci, ShuffleArray + I - N, N,
797                          /*SameGroup=*/true);
798           N = 1;
799           CurGroup = compactPtrGroupBase(ShuffleArray[I]);
800         } else {
801           ++N;
802         }
803       }
804 
805       shuffle(ShuffleArray + NumberOfBlocks - N, N, &Sci->RandState);
806       pushBlocksImpl(C, ClassId, Sci, &ShuffleArray[NumberOfBlocks - N], N,
807                      /*SameGroup=*/true);
808     } else {
809       pushBatchClassBlocks(Sci, ShuffleArray, NumberOfBlocks);
810     }
811 
812     // Note that `PushedBlocks` and `PoppedBlocks` are supposed to only record
813     // the requests from `PushBlocks` and `PopBatch` which are external
814     // interfaces. `populateFreeList` is the internal interface so we should set
815     // the values back to avoid incorrectly setting the stats.
816     Sci->FreeListInfo.PushedBlocks -= NumberOfBlocks;
817 
818     const uptr AllocatedUser = Size * NumberOfBlocks;
819     C->getStats().add(StatFree, AllocatedUser);
820     DCHECK_LE(Sci->CurrentRegionAllocated + AllocatedUser, RegionSize);
821     // If there is not enough room in the region currently associated to fit
822     // more blocks, we deassociate the region by resetting CurrentRegion and
823     // CurrentRegionAllocated. Otherwise, update the allocated amount.
824     if (RegionSize - (Sci->CurrentRegionAllocated + AllocatedUser) < Size) {
825       Sci->CurrentRegion = 0;
826       Sci->CurrentRegionAllocated = 0;
827     } else {
828       Sci->CurrentRegionAllocated += AllocatedUser;
829     }
830     Sci->AllocatedUser += AllocatedUser;
831 
832     return true;
833   }
834 
835   void getStats(ScopedString *Str, uptr ClassId, SizeClassInfo *Sci)
836       REQUIRES(Sci->Mutex) {
837     if (Sci->AllocatedUser == 0)
838       return;
839     const uptr BlockSize = getSizeByClassId(ClassId);
840     const uptr InUse =
841         Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks;
842     const uptr BytesInFreeList = Sci->AllocatedUser - InUse * BlockSize;
843     uptr PushedBytesDelta = 0;
844     if (BytesInFreeList >= Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint) {
845       PushedBytesDelta =
846           BytesInFreeList - Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint;
847     }
848     const uptr AvailableChunks = Sci->AllocatedUser / BlockSize;
849     Str->append("  %02zu (%6zu): mapped: %6zuK popped: %7zu pushed: %7zu "
850                 "inuse: %6zu avail: %6zu releases: %6zu last released: %6zuK "
851                 "latest pushed bytes: %6zuK\n",
852                 ClassId, getSizeByClassId(ClassId), Sci->AllocatedUser >> 10,
853                 Sci->FreeListInfo.PoppedBlocks, Sci->FreeListInfo.PushedBlocks,
854                 InUse, AvailableChunks, Sci->ReleaseInfo.RangesReleased,
855                 Sci->ReleaseInfo.LastReleasedBytes >> 10,
856                 PushedBytesDelta >> 10);
857   }
858 
859   NOINLINE uptr releaseToOSMaybe(SizeClassInfo *Sci, uptr ClassId,
860                                  ReleaseToOS ReleaseType = ReleaseToOS::Normal)
861       REQUIRES(Sci->Mutex) {
862     const uptr BlockSize = getSizeByClassId(ClassId);
863     const uptr PageSize = getPageSizeCached();
864 
865     DCHECK_GE(Sci->FreeListInfo.PoppedBlocks, Sci->FreeListInfo.PushedBlocks);
866     const uptr BytesInFreeList =
867         Sci->AllocatedUser -
868         (Sci->FreeListInfo.PoppedBlocks - Sci->FreeListInfo.PushedBlocks) *
869             BlockSize;
870 
871     if (UNLIKELY(BytesInFreeList == 0))
872       return 0;
873 
874     if (BytesInFreeList <= Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint)
875       Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint = BytesInFreeList;
876 
877     // Always update `BytesInFreeListAtLastCheckpoint` with the smallest value
878     // so that we won't underestimate the releasable pages. For example, the
879     // following is the region usage,
880     //
881     //  BytesInFreeListAtLastCheckpoint   AllocatedUser
882     //                v                         v
883     //  |--------------------------------------->
884     //         ^                   ^
885     //  BytesInFreeList     ReleaseThreshold
886     //
887     // In general, if we have collected enough bytes and the amount of free
888     // bytes meets the ReleaseThreshold, we will try to do page release. If we
889     // don't update `BytesInFreeListAtLastCheckpoint` when the current
890     // `BytesInFreeList` is smaller, we may take longer time to wait for enough
891     // freed blocks because we miss the bytes between
892     // (BytesInFreeListAtLastCheckpoint - BytesInFreeList).
893     const uptr PushedBytesDelta =
894         BytesInFreeList - Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint;
895     if (PushedBytesDelta < PageSize && ReleaseType != ReleaseToOS::ForceAll)
896       return 0;
897 
898     const bool CheckDensity =
899         isSmallBlock(BlockSize) && ReleaseType != ReleaseToOS::ForceAll;
900     // Releasing smaller blocks is expensive, so we want to make sure that a
901     // significant amount of bytes are free, and that there has been a good
902     // amount of batches pushed to the freelist before attempting to release.
903     if (CheckDensity && ReleaseType == ReleaseToOS::Normal)
904       if (PushedBytesDelta < Sci->AllocatedUser / 16U)
905         return 0;
906 
907     if (ReleaseType == ReleaseToOS::Normal) {
908       const s32 IntervalMs = atomic_load_relaxed(&ReleaseToOsIntervalMs);
909       if (IntervalMs < 0)
910         return 0;
911 
912       // The constant 8 here is selected from profiling some apps and the number
913       // of unreleased pages in the large size classes is around 16 pages or
914       // more. Choose half of it as a heuristic and which also avoids page
915       // release every time for every pushBlocks() attempt by large blocks.
916       const bool ByPassReleaseInterval =
917           isLargeBlock(BlockSize) && PushedBytesDelta > 8 * PageSize;
918       if (!ByPassReleaseInterval) {
919         if (Sci->ReleaseInfo.LastReleaseAtNs +
920                 static_cast<u64>(IntervalMs) * 1000000 >
921             getMonotonicTimeFast()) {
922           // Memory was returned recently.
923           return 0;
924         }
925       }
926     } // if (ReleaseType == ReleaseToOS::Normal)
927 
928     const uptr First = Sci->MinRegionIndex;
929     const uptr Last = Sci->MaxRegionIndex;
930     DCHECK_NE(Last, 0U);
931     DCHECK_LE(First, Last);
932     uptr TotalReleasedBytes = 0;
933     const uptr Base = First * RegionSize;
934     const uptr NumberOfRegions = Last - First + 1U;
935     const uptr GroupSize = (1U << GroupSizeLog);
936     const uptr CurGroupBase =
937         compactPtrGroupBase(compactPtr(ClassId, Sci->CurrentRegion));
938 
939     ReleaseRecorder Recorder(Base);
940     PageReleaseContext Context(BlockSize, NumberOfRegions,
941                                /*ReleaseSize=*/RegionSize);
942 
943     auto DecompactPtr = [](CompactPtrT CompactPtr) {
944       return reinterpret_cast<uptr>(CompactPtr);
945     };
946     for (BatchGroup &BG : Sci->FreeListInfo.BlockList) {
947       const uptr GroupBase = decompactGroupBase(BG.CompactPtrGroupBase);
948       // The `GroupSize` may not be divided by `BlockSize`, which means there is
949       // an unused space at the end of Region. Exclude that space to avoid
950       // unused page map entry.
951       uptr AllocatedGroupSize = GroupBase == CurGroupBase
952                                     ? Sci->CurrentRegionAllocated
953                                     : roundDownSlow(GroupSize, BlockSize);
954       if (AllocatedGroupSize == 0)
955         continue;
956 
957       // TransferBatches are pushed in front of BG.Batches. The first one may
958       // not have all caches used.
959       const uptr NumBlocks = (BG.Batches.size() - 1) * BG.MaxCachedPerBatch +
960                              BG.Batches.front()->getCount();
961       const uptr BytesInBG = NumBlocks * BlockSize;
962 
963       if (ReleaseType != ReleaseToOS::ForceAll &&
964           BytesInBG <= BG.BytesInBGAtLastCheckpoint) {
965         BG.BytesInBGAtLastCheckpoint = BytesInBG;
966         continue;
967       }
968       const uptr PushedBytesDelta = BytesInBG - BG.BytesInBGAtLastCheckpoint;
969       if (ReleaseType != ReleaseToOS::ForceAll && PushedBytesDelta < PageSize)
970         continue;
971 
972       // Given the randomness property, we try to release the pages only if the
973       // bytes used by free blocks exceed certain proportion of allocated
974       // spaces.
975       if (CheckDensity && (BytesInBG * 100U) / AllocatedGroupSize <
976                               (100U - 1U - BlockSize / 16U)) {
977         continue;
978       }
979 
980       // TODO: Consider updating this after page release if `ReleaseRecorder`
981       // can tell the releasd bytes in each group.
982       BG.BytesInBGAtLastCheckpoint = BytesInBG;
983 
984       const uptr MaxContainedBlocks = AllocatedGroupSize / BlockSize;
985       const uptr RegionIndex = (GroupBase - Base) / RegionSize;
986 
987       if (NumBlocks == MaxContainedBlocks) {
988         for (const auto &It : BG.Batches)
989           for (u16 I = 0; I < It.getCount(); ++I)
990             DCHECK_EQ(compactPtrGroupBase(It.get(I)), BG.CompactPtrGroupBase);
991 
992         const uptr To = GroupBase + AllocatedGroupSize;
993         Context.markRangeAsAllCounted(GroupBase, To, GroupBase, RegionIndex,
994                                       AllocatedGroupSize);
995       } else {
996         DCHECK_LT(NumBlocks, MaxContainedBlocks);
997 
998         // Note that we don't always visit blocks in each BatchGroup so that we
999         // may miss the chance of releasing certain pages that cross
1000         // BatchGroups.
1001         Context.markFreeBlocksInRegion(BG.Batches, DecompactPtr, GroupBase,
1002                                        RegionIndex, AllocatedGroupSize,
1003                                        /*MayContainLastBlockInRegion=*/true);
1004       }
1005 
1006       // We may not be able to do the page release In a rare case that we may
1007       // fail on PageMap allocation.
1008       if (UNLIKELY(!Context.hasBlockMarked()))
1009         return 0;
1010     }
1011 
1012     if (!Context.hasBlockMarked())
1013       return 0;
1014 
1015     auto SkipRegion = [this, First, ClassId](uptr RegionIndex) {
1016       ScopedLock L(ByteMapMutex);
1017       return (PossibleRegions[First + RegionIndex] - 1U) != ClassId;
1018     };
1019     releaseFreeMemoryToOS(Context, Recorder, SkipRegion);
1020 
1021     if (Recorder.getReleasedRangesCount() > 0) {
1022       Sci->ReleaseInfo.BytesInFreeListAtLastCheckpoint = BytesInFreeList;
1023       Sci->ReleaseInfo.RangesReleased += Recorder.getReleasedRangesCount();
1024       Sci->ReleaseInfo.LastReleasedBytes = Recorder.getReleasedBytes();
1025       TotalReleasedBytes += Sci->ReleaseInfo.LastReleasedBytes;
1026     }
1027     Sci->ReleaseInfo.LastReleaseAtNs = getMonotonicTimeFast();
1028 
1029     return TotalReleasedBytes;
1030   }
1031 
1032   SizeClassInfo SizeClassInfoArray[NumClasses] = {};
1033 
1034   HybridMutex ByteMapMutex;
1035   // Track the regions in use, 0 is unused, otherwise store ClassId + 1.
1036   ByteMap PossibleRegions GUARDED_BY(ByteMapMutex) = {};
1037   atomic_s32 ReleaseToOsIntervalMs = {};
1038   // Unless several threads request regions simultaneously from different size
1039   // classes, the stash rarely contains more than 1 entry.
1040   static constexpr uptr MaxStashedRegions = 4;
1041   HybridMutex RegionsStashMutex;
1042   uptr NumberOfStashedRegions GUARDED_BY(RegionsStashMutex) = 0;
1043   uptr RegionsStash[MaxStashedRegions] GUARDED_BY(RegionsStashMutex) = {};
1044 };
1045 
1046 } // namespace scudo
1047 
1048 #endif // SCUDO_PRIMARY32_H_
1049