1 //===-- primary64.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_PRIMARY64_H_
10 #define SCUDO_PRIMARY64_H_
11 
12 #include "bytemap.h"
13 #include "common.h"
14 #include "list.h"
15 #include "local_cache.h"
16 #include "memtag.h"
17 #include "options.h"
18 #include "release.h"
19 #include "stats.h"
20 #include "string_utils.h"
21 
22 namespace scudo {
23 
24 // SizeClassAllocator64 is an allocator tuned for 64-bit address space.
25 //
26 // It starts by reserving NumClasses * 2^RegionSizeLog bytes, equally divided in
27 // Regions, specific to each size class. Note that the base of that mapping is
28 // random (based to the platform specific map() capabilities). If
29 // PrimaryEnableRandomOffset is set, each Region actually starts at a random
30 // offset from its base.
31 //
32 // Regions are mapped incrementally on demand to fulfill allocation requests,
33 // those mappings being split into equally sized Blocks based on the size class
34 // they belong to. The Blocks created are shuffled to prevent predictable
35 // address patterns (the predictability increases with the size of the Blocks).
36 //
37 // The 1st Region (for size class 0) holds the TransferBatches. This is a
38 // structure used to transfer arrays of available pointers from the class size
39 // freelist to the thread specific freelist, and back.
40 //
41 // The memory used by this allocator is never unmapped, but can be partially
42 // released if the platform allows for it.
43 
44 template <typename Config> class SizeClassAllocator64 {
45 public:
46   typedef typename Config::PrimaryCompactPtrT CompactPtrT;
47   static const uptr CompactPtrScale = Config::PrimaryCompactPtrScale;
48   static const uptr GroupSizeLog = Config::PrimaryGroupSizeLog;
49   typedef typename Config::SizeClassMap SizeClassMap;
50   typedef SizeClassAllocator64<Config> ThisT;
51   typedef SizeClassAllocatorLocalCache<ThisT> CacheT;
52   typedef typename CacheT::TransferBatch TransferBatch;
53   typedef typename CacheT::BatchGroup BatchGroup;
54 
55   static uptr getSizeByClassId(uptr ClassId) {
56     return (ClassId == SizeClassMap::BatchClassId)
57                ? roundUpTo(sizeof(TransferBatch), 1U << CompactPtrScale)
58                : SizeClassMap::getSizeByClassId(ClassId);
59   }
60 
61   static bool canAllocate(uptr Size) { return Size <= SizeClassMap::MaxSize; }
62 
63   void init(s32 ReleaseToOsInterval) {
64     DCHECK(isAligned(reinterpret_cast<uptr>(this), alignof(ThisT)));
65     DCHECK_EQ(PrimaryBase, 0U);
66     // Reserve the space required for the Primary.
67     PrimaryBase = reinterpret_cast<uptr>(
68         map(nullptr, PrimarySize, nullptr, MAP_NOACCESS, &Data));
69 
70     u32 Seed;
71     const u64 Time = getMonotonicTime();
72     if (!getRandom(reinterpret_cast<void *>(&Seed), sizeof(Seed)))
73       Seed = static_cast<u32>(Time ^ (PrimaryBase >> 12));
74     const uptr PageSize = getPageSizeCached();
75     for (uptr I = 0; I < NumClasses; I++) {
76       RegionInfo *Region = getRegionInfo(I);
77       // The actual start of a region is offset by a random number of pages
78       // when PrimaryEnableRandomOffset is set.
79       Region->RegionBeg = getRegionBaseByClassId(I) +
80                           (Config::PrimaryEnableRandomOffset
81                                ? ((getRandomModN(&Seed, 16) + 1) * PageSize)
82                                : 0);
83       Region->RandState = getRandomU32(&Seed);
84       Region->ReleaseInfo.LastReleaseAtNs = Time;
85     }
86     setOption(Option::ReleaseInterval, static_cast<sptr>(ReleaseToOsInterval));
87   }
88 
89   void unmapTestOnly() {
90     for (uptr I = 0; I < NumClasses; I++) {
91       RegionInfo *Region = getRegionInfo(I);
92       *Region = {};
93     }
94     if (PrimaryBase)
95       unmap(reinterpret_cast<void *>(PrimaryBase), PrimarySize, UNMAP_ALL,
96             &Data);
97     PrimaryBase = 0U;
98   }
99 
100   TransferBatch *popBatch(CacheT *C, uptr ClassId) {
101     DCHECK_LT(ClassId, NumClasses);
102     RegionInfo *Region = getRegionInfo(ClassId);
103     ScopedLock L(Region->Mutex);
104     TransferBatch *B = popBatchImpl(C, ClassId);
105     if (UNLIKELY(!B)) {
106       if (UNLIKELY(!populateFreeList(C, ClassId, Region)))
107         return nullptr;
108       B = popBatchImpl(C, ClassId);
109       // if `populateFreeList` succeeded, we are supposed to get free blocks.
110       DCHECK_NE(B, nullptr);
111     }
112     Region->Stats.PoppedBlocks += B->getCount();
113     return B;
114   }
115 
116   // Push the array of free blocks to the designated batch group.
117   void pushBlocks(CacheT *C, uptr ClassId, CompactPtrT *Array, u32 Size) {
118     DCHECK_LT(ClassId, NumClasses);
119     DCHECK_GT(Size, 0);
120 
121     RegionInfo *Region = getRegionInfo(ClassId);
122     if (ClassId == SizeClassMap::BatchClassId) {
123       ScopedLock L(Region->Mutex);
124       // Constructing a batch group in the free list will use two blocks in
125       // BatchClassId. If we are pushing BatchClassId blocks, we will use the
126       // blocks in the array directly (can't delegate local cache which will
127       // cause a recursive allocation). However, The number of free blocks may
128       // be less than two. Therefore, populate the free list before inserting
129       // the blocks.
130       if (Size == 1 && UNLIKELY(!populateFreeList(C, ClassId, Region)))
131         return;
132       pushBlocksImpl(C, ClassId, Array, Size);
133       Region->Stats.PushedBlocks += Size;
134       return;
135     }
136 
137     // TODO(chiahungduan): Consider not doing grouping if the group size is not
138     // greater than the block size with a certain scale.
139 
140     // Sort the blocks so that blocks belonging to the same group can be pushed
141     // together.
142     bool SameGroup = true;
143     for (u32 I = 1; I < Size; ++I) {
144       if (compactPtrGroup(Array[I - 1]) != compactPtrGroup(Array[I]))
145         SameGroup = false;
146       CompactPtrT Cur = Array[I];
147       u32 J = I;
148       while (J > 0 && compactPtrGroup(Cur) < compactPtrGroup(Array[J - 1])) {
149         Array[J] = Array[J - 1];
150         --J;
151       }
152       Array[J] = Cur;
153     }
154 
155     ScopedLock L(Region->Mutex);
156     pushBlocksImpl(C, ClassId, Array, Size, SameGroup);
157 
158     Region->Stats.PushedBlocks += Size;
159     if (ClassId != SizeClassMap::BatchClassId)
160       releaseToOSMaybe(Region, ClassId);
161   }
162 
163   void disable() {
164     // The BatchClassId must be locked last since other classes can use it.
165     for (sptr I = static_cast<sptr>(NumClasses) - 1; I >= 0; I--) {
166       if (static_cast<uptr>(I) == SizeClassMap::BatchClassId)
167         continue;
168       getRegionInfo(static_cast<uptr>(I))->Mutex.lock();
169     }
170     getRegionInfo(SizeClassMap::BatchClassId)->Mutex.lock();
171   }
172 
173   void enable() {
174     getRegionInfo(SizeClassMap::BatchClassId)->Mutex.unlock();
175     for (uptr I = 0; I < NumClasses; I++) {
176       if (I == SizeClassMap::BatchClassId)
177         continue;
178       getRegionInfo(I)->Mutex.unlock();
179     }
180   }
181 
182   template <typename F> void iterateOverBlocks(F Callback) {
183     for (uptr I = 0; I < NumClasses; I++) {
184       if (I == SizeClassMap::BatchClassId)
185         continue;
186       const RegionInfo *Region = getRegionInfo(I);
187       const uptr BlockSize = getSizeByClassId(I);
188       const uptr From = Region->RegionBeg;
189       const uptr To = From + Region->AllocatedUser;
190       for (uptr Block = From; Block < To; Block += BlockSize)
191         Callback(Block);
192     }
193   }
194 
195   void getStats(ScopedString *Str) {
196     // TODO(kostyak): get the RSS per region.
197     uptr TotalMapped = 0;
198     uptr PoppedBlocks = 0;
199     uptr PushedBlocks = 0;
200     for (uptr I = 0; I < NumClasses; I++) {
201       RegionInfo *Region = getRegionInfo(I);
202       if (Region->MappedUser)
203         TotalMapped += Region->MappedUser;
204       PoppedBlocks += Region->Stats.PoppedBlocks;
205       PushedBlocks += Region->Stats.PushedBlocks;
206     }
207     Str->append("Stats: SizeClassAllocator64: %zuM mapped (%uM rss) in %zu "
208                 "allocations; remains %zu\n",
209                 TotalMapped >> 20, 0U, PoppedBlocks,
210                 PoppedBlocks - PushedBlocks);
211 
212     for (uptr I = 0; I < NumClasses; I++)
213       getStats(Str, I, 0);
214   }
215 
216   bool setOption(Option O, sptr Value) {
217     if (O == Option::ReleaseInterval) {
218       const s32 Interval = Max(
219           Min(static_cast<s32>(Value), Config::PrimaryMaxReleaseToOsIntervalMs),
220           Config::PrimaryMinReleaseToOsIntervalMs);
221       atomic_store_relaxed(&ReleaseToOsIntervalMs, Interval);
222       return true;
223     }
224     // Not supported by the Primary, but not an error either.
225     return true;
226   }
227 
228   uptr releaseToOS() {
229     uptr TotalReleasedBytes = 0;
230     for (uptr I = 0; I < NumClasses; I++) {
231       if (I == SizeClassMap::BatchClassId)
232         continue;
233       RegionInfo *Region = getRegionInfo(I);
234       ScopedLock L(Region->Mutex);
235       TotalReleasedBytes += releaseToOSMaybe(Region, I, /*Force=*/true);
236     }
237     return TotalReleasedBytes;
238   }
239 
240   const char *getRegionInfoArrayAddress() const {
241     return reinterpret_cast<const char *>(RegionInfoArray);
242   }
243 
244   static uptr getRegionInfoArraySize() { return sizeof(RegionInfoArray); }
245 
246   uptr getCompactPtrBaseByClassId(uptr ClassId) {
247     // If we are not compacting pointers, base everything off of 0.
248     if (sizeof(CompactPtrT) == sizeof(uptr) && CompactPtrScale == 0)
249       return 0;
250     return getRegionInfo(ClassId)->RegionBeg;
251   }
252 
253   CompactPtrT compactPtr(uptr ClassId, uptr Ptr) {
254     DCHECK_LE(ClassId, SizeClassMap::LargestClassId);
255     return compactPtrInternal(getCompactPtrBaseByClassId(ClassId), Ptr);
256   }
257 
258   void *decompactPtr(uptr ClassId, CompactPtrT CompactPtr) {
259     DCHECK_LE(ClassId, SizeClassMap::LargestClassId);
260     return reinterpret_cast<void *>(
261         decompactPtrInternal(getCompactPtrBaseByClassId(ClassId), CompactPtr));
262   }
263 
264   static BlockInfo findNearestBlock(const char *RegionInfoData, uptr Ptr) {
265     const RegionInfo *RegionInfoArray =
266         reinterpret_cast<const RegionInfo *>(RegionInfoData);
267     uptr ClassId;
268     uptr MinDistance = -1UL;
269     for (uptr I = 0; I != NumClasses; ++I) {
270       if (I == SizeClassMap::BatchClassId)
271         continue;
272       uptr Begin = RegionInfoArray[I].RegionBeg;
273       uptr End = Begin + RegionInfoArray[I].AllocatedUser;
274       if (Begin > End || End - Begin < SizeClassMap::getSizeByClassId(I))
275         continue;
276       uptr RegionDistance;
277       if (Begin <= Ptr) {
278         if (Ptr < End)
279           RegionDistance = 0;
280         else
281           RegionDistance = Ptr - End;
282       } else {
283         RegionDistance = Begin - Ptr;
284       }
285 
286       if (RegionDistance < MinDistance) {
287         MinDistance = RegionDistance;
288         ClassId = I;
289       }
290     }
291 
292     BlockInfo B = {};
293     if (MinDistance <= 8192) {
294       B.RegionBegin = RegionInfoArray[ClassId].RegionBeg;
295       B.RegionEnd = B.RegionBegin + RegionInfoArray[ClassId].AllocatedUser;
296       B.BlockSize = SizeClassMap::getSizeByClassId(ClassId);
297       B.BlockBegin =
298           B.RegionBegin + uptr(sptr(Ptr - B.RegionBegin) / sptr(B.BlockSize) *
299                                sptr(B.BlockSize));
300       while (B.BlockBegin < B.RegionBegin)
301         B.BlockBegin += B.BlockSize;
302       while (B.RegionEnd < B.BlockBegin + B.BlockSize)
303         B.BlockBegin -= B.BlockSize;
304     }
305     return B;
306   }
307 
308   AtomicOptions Options;
309 
310 private:
311   static const uptr RegionSize = 1UL << Config::PrimaryRegionSizeLog;
312   static const uptr NumClasses = SizeClassMap::NumClasses;
313   static const uptr PrimarySize = RegionSize * NumClasses;
314 
315   static const uptr MapSizeIncrement = Config::PrimaryMapSizeIncrement;
316   // Fill at most this number of batches from the newly map'd memory.
317   static const u32 MaxNumBatches = SCUDO_ANDROID ? 4U : 8U;
318 
319   struct RegionStats {
320     uptr PoppedBlocks;
321     uptr PushedBlocks;
322   };
323 
324   struct ReleaseToOsInfo {
325     uptr PushedBlocksAtLastRelease;
326     uptr RangesReleased;
327     uptr LastReleasedBytes;
328     u64 LastReleaseAtNs;
329   };
330 
331   struct UnpaddedRegionInfo {
332     HybridMutex Mutex;
333     SinglyLinkedList<BatchGroup> FreeList;
334     uptr RegionBeg = 0;
335     RegionStats Stats = {};
336     u32 RandState = 0;
337     uptr MappedUser = 0;    // Bytes mapped for user memory.
338     uptr AllocatedUser = 0; // Bytes allocated for user memory.
339     MapPlatformData Data = {};
340     ReleaseToOsInfo ReleaseInfo = {};
341     bool Exhausted = false;
342   };
343   struct RegionInfo : UnpaddedRegionInfo {
344     char Padding[SCUDO_CACHE_LINE_SIZE -
345                  (sizeof(UnpaddedRegionInfo) % SCUDO_CACHE_LINE_SIZE)] = {};
346   };
347   static_assert(sizeof(RegionInfo) % SCUDO_CACHE_LINE_SIZE == 0, "");
348 
349   uptr PrimaryBase = 0;
350   MapPlatformData Data = {};
351   atomic_s32 ReleaseToOsIntervalMs = {};
352   alignas(SCUDO_CACHE_LINE_SIZE) RegionInfo RegionInfoArray[NumClasses];
353 
354   RegionInfo *getRegionInfo(uptr ClassId) {
355     DCHECK_LT(ClassId, NumClasses);
356     return &RegionInfoArray[ClassId];
357   }
358 
359   uptr getRegionBaseByClassId(uptr ClassId) const {
360     return PrimaryBase + (ClassId << Config::PrimaryRegionSizeLog);
361   }
362 
363   static CompactPtrT compactPtrInternal(uptr Base, uptr Ptr) {
364     return static_cast<CompactPtrT>((Ptr - Base) >> CompactPtrScale);
365   }
366 
367   static uptr decompactPtrInternal(uptr Base, CompactPtrT CompactPtr) {
368     return Base + (static_cast<uptr>(CompactPtr) << CompactPtrScale);
369   }
370 
371   static uptr compactPtrGroup(CompactPtrT CompactPtr) {
372     return static_cast<uptr>(CompactPtr) >> (GroupSizeLog - CompactPtrScale);
373   }
374   static uptr batchGroupBase(uptr Base, uptr GroupId) {
375     return (GroupId << GroupSizeLog) + Base;
376   }
377 
378   // Push the blocks to their batch group. The layout will be like,
379   //
380   // FreeList - > BG -> BG -> BG
381   //              |     |     |
382   //              v     v     v
383   //              TB    TB    TB
384   //              |
385   //              v
386   //              TB
387   //
388   // Each BlockGroup(BG) will associate with unique group id and the free blocks
389   // are managed by a list of TransferBatch(TB). To reduce the time of inserting
390   // blocks, BGs are sorted and the input `Array` are supposed to be sorted so
391   // that we can get better performance of maintaining sorted property.
392   // Use `SameGroup=true` to indicate that all blocks in the array are from the
393   // same group then we will skip checking the group id of each block.
394   //
395   // Note that this aims to have a better management of dirty pages, i.e., the
396   // RSS usage won't grow indefinitely. There's an exception that we may not put
397   // a block to its associated group. While populating new blocks, we may have
398   // blocks cross different groups. However, most cases will fall into same
399   // group and they are supposed to be popped soon. In that case, it's not worth
400   // sorting the array with the almost-sorted property. Therefore, we use
401   // `SameGroup=true` instead.
402   //
403   // The region mutex needs to be held while calling this method.
404   void pushBlocksImpl(CacheT *C, uptr ClassId, CompactPtrT *Array, u32 Size,
405                       bool SameGroup = false) {
406     DCHECK_GT(Size, 0U);
407     RegionInfo *Region = getRegionInfo(ClassId);
408 
409     auto CreateGroup = [&](uptr GroupId) {
410       BatchGroup *BG = nullptr;
411       TransferBatch *TB = nullptr;
412       if (ClassId == SizeClassMap::BatchClassId) {
413         DCHECK_GE(Size, 2U);
414         BG = reinterpret_cast<BatchGroup *>(
415             decompactPtr(ClassId, Array[Size - 1]));
416         BG->Batches.clear();
417 
418         TB = reinterpret_cast<TransferBatch *>(
419             decompactPtr(ClassId, Array[Size - 2]));
420         TB->clear();
421       } else {
422         BG = C->createGroup();
423         BG->Batches.clear();
424 
425         TB = C->createBatch(ClassId, nullptr);
426         TB->clear();
427       }
428 
429       BG->GroupId = GroupId;
430       BG->Batches.push_front(TB);
431       BG->PushedBlocks = 0;
432       BG->PushedBlocksAtLastCheckpoint = 0;
433       BG->MaxCachedPerBatch =
434           TransferBatch::getMaxCached(getSizeByClassId(ClassId));
435 
436       return BG;
437     };
438 
439     auto InsertBlocks = [&](BatchGroup *BG, CompactPtrT *Array, u32 Size) {
440       SinglyLinkedList<TransferBatch> &Batches = BG->Batches;
441       TransferBatch *CurBatch = Batches.front();
442       DCHECK_NE(CurBatch, nullptr);
443 
444       for (u32 I = 0; I < Size;) {
445         DCHECK_GE(BG->MaxCachedPerBatch, CurBatch->getCount());
446         u16 UnusedSlots =
447             static_cast<u16>(BG->MaxCachedPerBatch - CurBatch->getCount());
448         if (UnusedSlots == 0) {
449           CurBatch = C->createBatch(
450               ClassId,
451               reinterpret_cast<void *>(decompactPtr(ClassId, Array[I])));
452           CurBatch->clear();
453           Batches.push_front(CurBatch);
454           UnusedSlots = BG->MaxCachedPerBatch;
455         }
456         // `UnusedSlots` is u16 so the result will be also fit in u16.
457         u16 AppendSize = static_cast<u16>(Min<u32>(UnusedSlots, Size - I));
458         CurBatch->appendFromArray(&Array[I], AppendSize);
459         I += AppendSize;
460       }
461 
462       BG->PushedBlocks += Size;
463     };
464 
465     BatchGroup *Cur = Region->FreeList.front();
466 
467     if (ClassId == SizeClassMap::BatchClassId) {
468       if (Cur == nullptr) {
469         // Don't need to classify BatchClassId.
470         Cur = CreateGroup(/*GroupId=*/0);
471         Region->FreeList.push_front(Cur);
472       }
473       InsertBlocks(Cur, Array, Size);
474       return;
475     }
476 
477     // In the following, `Cur` always points to the BatchGroup for blocks that
478     // will be pushed next. `Prev` is the element right before `Cur`.
479     BatchGroup *Prev = nullptr;
480 
481     while (Cur != nullptr && compactPtrGroup(Array[0]) > Cur->GroupId) {
482       Prev = Cur;
483       Cur = Cur->Next;
484     }
485 
486     if (Cur == nullptr || compactPtrGroup(Array[0]) != Cur->GroupId) {
487       Cur = CreateGroup(compactPtrGroup(Array[0]));
488       if (Prev == nullptr)
489         Region->FreeList.push_front(Cur);
490       else
491         Region->FreeList.insert(Prev, Cur);
492     }
493 
494     // All the blocks are from the same group, just push without checking group
495     // id.
496     if (SameGroup) {
497       InsertBlocks(Cur, Array, Size);
498       return;
499     }
500 
501     // The blocks are sorted by group id. Determine the segment of group and
502     // push them to their group together.
503     u32 Count = 1;
504     for (u32 I = 1; I < Size; ++I) {
505       if (compactPtrGroup(Array[I - 1]) != compactPtrGroup(Array[I])) {
506         DCHECK_EQ(compactPtrGroup(Array[I - 1]), Cur->GroupId);
507         InsertBlocks(Cur, Array + I - Count, Count);
508 
509         while (Cur != nullptr && compactPtrGroup(Array[I]) > Cur->GroupId) {
510           Prev = Cur;
511           Cur = Cur->Next;
512         }
513 
514         if (Cur == nullptr || compactPtrGroup(Array[I]) != Cur->GroupId) {
515           Cur = CreateGroup(compactPtrGroup(Array[I]));
516           DCHECK_NE(Prev, nullptr);
517           Region->FreeList.insert(Prev, Cur);
518         }
519 
520         Count = 1;
521       } else {
522         ++Count;
523       }
524     }
525 
526     InsertBlocks(Cur, Array + Size - Count, Count);
527   }
528 
529   // Pop one TransferBatch from a BatchGroup. The BatchGroup with the smallest
530   // group id will be considered first.
531   //
532   // The region mutex needs to be held while calling this method.
533   TransferBatch *popBatchImpl(CacheT *C, uptr ClassId) {
534     RegionInfo *Region = getRegionInfo(ClassId);
535     if (Region->FreeList.empty())
536       return nullptr;
537 
538     SinglyLinkedList<TransferBatch> &Batches =
539         Region->FreeList.front()->Batches;
540     DCHECK(!Batches.empty());
541 
542     TransferBatch *B = Batches.front();
543     Batches.pop_front();
544     DCHECK_NE(B, nullptr);
545     DCHECK_GT(B->getCount(), 0U);
546 
547     if (Batches.empty()) {
548       BatchGroup *BG = Region->FreeList.front();
549       Region->FreeList.pop_front();
550 
551       // We don't keep BatchGroup with zero blocks to avoid empty-checking while
552       // allocating. Note that block used by constructing BatchGroup is recorded
553       // as free blocks in the last element of BatchGroup::Batches. Which means,
554       // once we pop the last TransferBatch, the block is implicitly
555       // deallocated.
556       if (ClassId != SizeClassMap::BatchClassId)
557         C->deallocate(SizeClassMap::BatchClassId, BG);
558     }
559 
560     return B;
561   }
562 
563   NOINLINE bool populateFreeList(CacheT *C, uptr ClassId, RegionInfo *Region) {
564     const uptr Size = getSizeByClassId(ClassId);
565     const u16 MaxCount = TransferBatch::getMaxCached(Size);
566 
567     const uptr RegionBeg = Region->RegionBeg;
568     const uptr MappedUser = Region->MappedUser;
569     const uptr TotalUserBytes = Region->AllocatedUser + MaxCount * Size;
570     // Map more space for blocks, if necessary.
571     if (TotalUserBytes > MappedUser) {
572       // Do the mmap for the user memory.
573       const uptr MapSize =
574           roundUpTo(TotalUserBytes - MappedUser, MapSizeIncrement);
575       const uptr RegionBase = RegionBeg - getRegionBaseByClassId(ClassId);
576       if (UNLIKELY(RegionBase + MappedUser + MapSize > RegionSize)) {
577         if (!Region->Exhausted) {
578           Region->Exhausted = true;
579           ScopedString Str;
580           getStats(&Str);
581           Str.append(
582               "Scudo OOM: The process has exhausted %zuM for size class %zu.\n",
583               RegionSize >> 20, Size);
584           Str.output();
585         }
586         return false;
587       }
588       if (MappedUser == 0)
589         Region->Data = Data;
590       if (UNLIKELY(!map(
591               reinterpret_cast<void *>(RegionBeg + MappedUser), MapSize,
592               "scudo:primary",
593               MAP_ALLOWNOMEM | MAP_RESIZABLE |
594                   (useMemoryTagging<Config>(Options.load()) ? MAP_MEMTAG : 0),
595               &Region->Data))) {
596         return false;
597       }
598       Region->MappedUser += MapSize;
599       C->getStats().add(StatMapped, MapSize);
600     }
601 
602     const u32 NumberOfBlocks = Min(
603         MaxNumBatches * MaxCount,
604         static_cast<u32>((Region->MappedUser - Region->AllocatedUser) / Size));
605     DCHECK_GT(NumberOfBlocks, 0);
606 
607     constexpr u32 ShuffleArraySize =
608         MaxNumBatches * TransferBatch::MaxNumCached;
609     CompactPtrT ShuffleArray[ShuffleArraySize];
610     DCHECK_LE(NumberOfBlocks, ShuffleArraySize);
611 
612     const uptr CompactPtrBase = getCompactPtrBaseByClassId(ClassId);
613     uptr P = RegionBeg + Region->AllocatedUser;
614     for (u32 I = 0; I < NumberOfBlocks; I++, P += Size)
615       ShuffleArray[I] = compactPtrInternal(CompactPtrBase, P);
616     // No need to shuffle the batches size class.
617     if (ClassId != SizeClassMap::BatchClassId)
618       shuffle(ShuffleArray, NumberOfBlocks, &Region->RandState);
619     for (u32 I = 0; I < NumberOfBlocks;) {
620       // `MaxCount` is u16 so the result will also fit in u16.
621       const u16 N = static_cast<u16>(Min<u32>(MaxCount, NumberOfBlocks - I));
622       // Note that the N blocks here may have different group ids. Given that
623       // it only happens when it crosses the group size boundary. Instead of
624       // sorting them, treat them as same group here to avoid sorting the
625       // almost-sorted blocks.
626       pushBlocksImpl(C, ClassId, &ShuffleArray[I], N, /*SameGroup=*/true);
627       I += N;
628     }
629 
630     const uptr AllocatedUser = Size * NumberOfBlocks;
631     C->getStats().add(StatFree, AllocatedUser);
632     Region->AllocatedUser += AllocatedUser;
633 
634     return true;
635   }
636 
637   void getStats(ScopedString *Str, uptr ClassId, uptr Rss) {
638     RegionInfo *Region = getRegionInfo(ClassId);
639     if (Region->MappedUser == 0)
640       return;
641     const uptr InUse = Region->Stats.PoppedBlocks - Region->Stats.PushedBlocks;
642     const uptr TotalChunks = Region->AllocatedUser / getSizeByClassId(ClassId);
643     Str->append("%s %02zu (%6zu): mapped: %6zuK popped: %7zu pushed: %7zu "
644                 "inuse: %6zu total: %6zu rss: %6zuK releases: %6zu last "
645                 "released: %6zuK region: 0x%zx (0x%zx)\n",
646                 Region->Exhausted ? "F" : " ", ClassId,
647                 getSizeByClassId(ClassId), Region->MappedUser >> 10,
648                 Region->Stats.PoppedBlocks, Region->Stats.PushedBlocks, InUse,
649                 TotalChunks, Rss >> 10, Region->ReleaseInfo.RangesReleased,
650                 Region->ReleaseInfo.LastReleasedBytes >> 10, Region->RegionBeg,
651                 getRegionBaseByClassId(ClassId));
652   }
653 
654   NOINLINE uptr releaseToOSMaybe(RegionInfo *Region, uptr ClassId,
655                                  bool Force = false) {
656     const uptr BlockSize = getSizeByClassId(ClassId);
657     const uptr PageSize = getPageSizeCached();
658 
659     DCHECK_GE(Region->Stats.PoppedBlocks, Region->Stats.PushedBlocks);
660     const uptr BytesInFreeList =
661         Region->AllocatedUser -
662         (Region->Stats.PoppedBlocks - Region->Stats.PushedBlocks) * BlockSize;
663     if (BytesInFreeList < PageSize)
664       return 0; // No chance to release anything.
665     const uptr BytesPushed = (Region->Stats.PushedBlocks -
666                               Region->ReleaseInfo.PushedBlocksAtLastRelease) *
667                              BlockSize;
668     if (BytesPushed < PageSize)
669       return 0; // Nothing new to release.
670 
671     bool CheckDensity = BlockSize < PageSize / 16U;
672     // Releasing smaller blocks is expensive, so we want to make sure that a
673     // significant amount of bytes are free, and that there has been a good
674     // amount of batches pushed to the freelist before attempting to release.
675     if (CheckDensity) {
676       if (!Force && BytesPushed < Region->AllocatedUser / 16U)
677         return 0;
678     }
679 
680     if (!Force) {
681       const s32 IntervalMs = atomic_load_relaxed(&ReleaseToOsIntervalMs);
682       if (IntervalMs < 0)
683         return 0;
684       if (Region->ReleaseInfo.LastReleaseAtNs +
685               static_cast<u64>(IntervalMs) * 1000000 >
686           getMonotonicTime()) {
687         return 0; // Memory was returned recently.
688       }
689     }
690 
691     const uptr GroupSize = (1U << GroupSizeLog);
692     const uptr AllocatedUserEnd = Region->AllocatedUser + Region->RegionBeg;
693     ReleaseRecorder Recorder(Region->RegionBeg, &Region->Data);
694     PageReleaseContext Context(BlockSize, Region->AllocatedUser,
695                                /*NumberOfRegions=*/1U);
696 
697     const uptr CompactPtrBase = getCompactPtrBaseByClassId(ClassId);
698     auto DecompactPtr = [CompactPtrBase](CompactPtrT CompactPtr) {
699       return decompactPtrInternal(CompactPtrBase, CompactPtr);
700     };
701     for (BatchGroup &BG : Region->FreeList) {
702       const uptr PushedBytesDelta =
703           BG.PushedBlocks - BG.PushedBlocksAtLastCheckpoint;
704       if (PushedBytesDelta * BlockSize < PageSize)
705         continue;
706 
707       // Group boundary does not necessarily have the same alignment as Region.
708       // It may sit across a Region boundary. Which means that we may have the
709       // following two cases,
710       //
711       // 1. Group boundary sits before RegionBeg.
712       //
713       //                (BatchGroupBeg)
714       // batchGroupBase  RegionBeg       BatchGroupEnd
715       //        |            |                |
716       //        v            v                v
717       //        +------------+----------------+
718       //         \                           /
719       //          ------   GroupSize   ------
720       //
721       // 2. Group boundary sits after RegionBeg.
722       //
723       //               (BatchGroupBeg)
724       //    RegionBeg  batchGroupBase               BatchGroupEnd
725       //        |           |                             |
726       //        v           v                             v
727       //        +-----------+-----------------------------+
728       //                     \                           /
729       //                      ------   GroupSize   ------
730       //
731       // Note that in the first case, the group range before RegionBeg is never
732       // used. Therefore, while calculating the used group size, we should
733       // exclude that part to get the correct size.
734       const uptr BatchGroupBeg =
735           Max(batchGroupBase(CompactPtrBase, BG.GroupId), Region->RegionBeg);
736       DCHECK_GE(AllocatedUserEnd, BatchGroupBeg);
737       const uptr BatchGroupEnd =
738           batchGroupBase(CompactPtrBase, BG.GroupId) + GroupSize;
739       const uptr AllocatedGroupSize = AllocatedUserEnd >= BatchGroupEnd
740                                           ? BatchGroupEnd - BatchGroupBeg
741                                           : AllocatedUserEnd - BatchGroupBeg;
742       if (AllocatedGroupSize == 0)
743         continue;
744 
745       // TransferBatches are pushed in front of BG.Batches. The first one may
746       // not have all caches used.
747       const uptr NumBlocks = (BG.Batches.size() - 1) * BG.MaxCachedPerBatch +
748                              BG.Batches.front()->getCount();
749       const uptr BytesInBG = NumBlocks * BlockSize;
750       // Given the randomness property, we try to release the pages only if the
751       // bytes used by free blocks exceed certain proportion of group size. Note
752       // that this heuristic only applies when all the spaces in a BatchGroup
753       // are allocated.
754       if (CheckDensity && (BytesInBG * 100U) / AllocatedGroupSize <
755                               (100U - 1U - BlockSize / 16U)) {
756         continue;
757       }
758 
759       BG.PushedBlocksAtLastCheckpoint = BG.PushedBlocks;
760       // Note that we don't always visit blocks in each BatchGroup so that we
761       // may miss the chance of releasing certain pages that cross BatchGroups.
762       Context.markFreeBlocks(BG.Batches, DecompactPtr, Region->RegionBeg);
763     }
764 
765     if (!Context.hasBlockMarked())
766       return 0;
767 
768     auto SkipRegion = [](UNUSED uptr RegionIndex) { return false; };
769     releaseFreeMemoryToOS(Context, Recorder, SkipRegion);
770 
771     if (Recorder.getReleasedRangesCount() > 0) {
772       Region->ReleaseInfo.PushedBlocksAtLastRelease =
773           Region->Stats.PushedBlocks;
774       Region->ReleaseInfo.RangesReleased += Recorder.getReleasedRangesCount();
775       Region->ReleaseInfo.LastReleasedBytes = Recorder.getReleasedBytes();
776     }
777     Region->ReleaseInfo.LastReleaseAtNs = getMonotonicTime();
778     return Recorder.getReleasedBytes();
779   }
780 };
781 
782 } // namespace scudo
783 
784 #endif // SCUDO_PRIMARY64_H_
785