| 1 | //===-- sanitizer_quarantine.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 | // Memory quarantine for AddressSanitizer and potentially other tools. |
| 10 | // Quarantine caches some specified amount of memory in per-thread caches, |
| 11 | // then evicts to global FIFO queue. When the queue reaches specified threshold, |
| 12 | // oldest memory is recycled. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #ifndef SANITIZER_QUARANTINE_H |
| 17 | #define SANITIZER_QUARANTINE_H |
| 18 | |
| 19 | #include "sanitizer_internal_defs.h" |
| 20 | #include "sanitizer_mutex.h" |
| 21 | #include "sanitizer_list.h" |
| 22 | |
| 23 | namespace __sanitizer { |
| 24 | |
| 25 | template<typename Node> class QuarantineCache; |
| 26 | |
| 27 | struct QuarantineBatch { |
| 28 | static const uptr kSize = 1021; |
| 29 | QuarantineBatch *next; |
| 30 | uptr size; |
| 31 | uptr count; |
| 32 | void *batch[kSize]; |
| 33 | |
| 34 | void init(void *ptr, uptr size) { |
| 35 | count = 1; |
| 36 | batch[0] = ptr; |
| 37 | this->size = size + sizeof(QuarantineBatch); // Account for the batch size. |
| 38 | } |
| 39 | |
| 40 | // The total size of quarantined nodes recorded in this batch. |
| 41 | uptr quarantined_size() const { |
| 42 | return size - sizeof(QuarantineBatch); |
| 43 | } |
| 44 | |
| 45 | void push_back(void *ptr, uptr size) { |
| 46 | CHECK_LT(count, kSize); |
| 47 | batch[count++] = ptr; |
| 48 | this->size += size; |
| 49 | } |
| 50 | |
| 51 | bool can_merge(const QuarantineBatch* const from) const { |
| 52 | return count + from->count <= kSize; |
| 53 | } |
| 54 | |
| 55 | void merge(QuarantineBatch* const from) { |
| 56 | CHECK_LE(count + from->count, kSize); |
| 57 | CHECK_GE(size, sizeof(QuarantineBatch)); |
| 58 | |
| 59 | for (uptr i = 0; i < from->count; ++i) |
| 60 | batch[count + i] = from->batch[i]; |
| 61 | count += from->count; |
| 62 | size += from->quarantined_size(); |
| 63 | |
| 64 | from->count = 0; |
| 65 | from->size = sizeof(QuarantineBatch); |
| 66 | } |
| 67 | }; |
| 68 | |
| 69 | COMPILER_CHECK(sizeof(QuarantineBatch) <= (1 << 13)); // 8Kb. |
| 70 | |
| 71 | template<typename Callback, typename Node> |
| 72 | class Quarantine { |
| 73 | public: |
| 74 | typedef QuarantineCache<Callback> Cache; |
| 75 | |
| 76 | explicit Quarantine(LinkerInitialized) |
| 77 | : cache_(LINKER_INITIALIZED) { |
| 78 | } |
| 79 | |
| 80 | void Init(uptr size, uptr cache_size) { |
| 81 | // Thread local quarantine size can be zero only when global quarantine size |
| 82 | // is zero (it allows us to perform just one atomic read per Put() call). |
| 83 | CHECK((size == 0 && cache_size == 0) || cache_size != 0); |
| 84 | |
| 85 | atomic_store_relaxed(a: &max_size_, v: size); |
| 86 | atomic_store_relaxed(a: &min_size_, v: size / 10 * 9); // 90% of max size. |
| 87 | atomic_store_relaxed(a: &max_cache_size_, v: cache_size); |
| 88 | |
| 89 | cache_mutex_.Init(); |
| 90 | recycle_mutex_.Init(); |
| 91 | } |
| 92 | |
| 93 | uptr GetMaxSize() const { return atomic_load_relaxed(a: &max_size_); } |
| 94 | uptr GetMaxCacheSize() const { return atomic_load_relaxed(a: &max_cache_size_); } |
| 95 | |
| 96 | void Put(Cache *c, Callback cb, Node *ptr, uptr size) { |
| 97 | uptr max_cache_size = GetMaxCacheSize(); |
| 98 | if (max_cache_size && size <= GetMaxSize()) { |
| 99 | cb.PreQuarantine(ptr); |
| 100 | c->Enqueue(cb, ptr, size); |
| 101 | } else { |
| 102 | // GetMaxCacheSize() == 0 only when GetMaxSize() == 0 (see Init). |
| 103 | cb.RecyclePassThrough(ptr); |
| 104 | } |
| 105 | // Check cache size anyway to accommodate for runtime cache_size change. |
| 106 | if (c->Size() > max_cache_size) |
| 107 | Drain(c, cb); |
| 108 | } |
| 109 | |
| 110 | void NOINLINE Drain(Cache *c, Callback cb) { |
| 111 | { |
| 112 | SpinMutexLock l(&cache_mutex_); |
| 113 | cache_.Transfer(c); |
| 114 | } |
| 115 | if (cache_.Size() > GetMaxSize() && recycle_mutex_.TryLock()) |
| 116 | Recycle(min_size: atomic_load_relaxed(a: &min_size_), cb); |
| 117 | } |
| 118 | |
| 119 | void NOINLINE DrainAndRecycle(Cache *c, Callback cb) { |
| 120 | { |
| 121 | SpinMutexLock l(&cache_mutex_); |
| 122 | cache_.Transfer(c); |
| 123 | } |
| 124 | recycle_mutex_.Lock(); |
| 125 | Recycle(min_size: 0, cb); |
| 126 | } |
| 127 | |
| 128 | void PrintStats() const { |
| 129 | // It assumes that the world is stopped, just as the allocator's PrintStats. |
| 130 | Printf("Quarantine limits: global: %zdMb; thread local: %zdKb\n" , |
| 131 | GetMaxSize() >> 20, GetMaxCacheSize() >> 10); |
| 132 | cache_.PrintStats(); |
| 133 | } |
| 134 | |
| 135 | private: |
| 136 | // Read-only data. |
| 137 | char pad0_[kCacheLineSize]; |
| 138 | atomic_uintptr_t max_size_; |
| 139 | atomic_uintptr_t min_size_; |
| 140 | atomic_uintptr_t max_cache_size_; |
| 141 | char pad1_[kCacheLineSize]; |
| 142 | StaticSpinMutex cache_mutex_; |
| 143 | StaticSpinMutex recycle_mutex_; |
| 144 | Cache cache_; |
| 145 | char pad2_[kCacheLineSize]; |
| 146 | |
| 147 | void NOINLINE Recycle(uptr min_size, Callback cb) |
| 148 | SANITIZER_REQUIRES(recycle_mutex_) SANITIZER_RELEASE(recycle_mutex_) { |
| 149 | Cache tmp; |
| 150 | { |
| 151 | SpinMutexLock l(&cache_mutex_); |
| 152 | // Go over the batches and merge partially filled ones to |
| 153 | // save some memory, otherwise batches themselves (since the memory used |
| 154 | // by them is counted against quarantine limit) can overcome the actual |
| 155 | // user's quarantined chunks, which diminishes the purpose of the |
| 156 | // quarantine. |
| 157 | uptr cache_size = cache_.Size(); |
| 158 | uptr overhead_size = cache_.OverheadSize(); |
| 159 | CHECK_GE(cache_size, overhead_size); |
| 160 | // Do the merge only when overhead exceeds this predefined limit (might |
| 161 | // require some tuning). It saves us merge attempt when the batch list |
| 162 | // quarantine is unlikely to contain batches suitable for merge. |
| 163 | const uptr kOverheadThresholdPercents = 100; |
| 164 | if (cache_size > overhead_size && |
| 165 | overhead_size * (100 + kOverheadThresholdPercents) > |
| 166 | cache_size * kOverheadThresholdPercents) { |
| 167 | cache_.MergeBatches(&tmp); |
| 168 | } |
| 169 | // Extract enough chunks from the quarantine to get below the max |
| 170 | // quarantine size and leave some leeway for the newly quarantined chunks. |
| 171 | while (cache_.Size() > min_size) { |
| 172 | tmp.EnqueueBatch(cache_.DequeueBatch()); |
| 173 | } |
| 174 | } |
| 175 | recycle_mutex_.Unlock(); |
| 176 | DoRecycle(c: &tmp, cb); |
| 177 | } |
| 178 | |
| 179 | void NOINLINE DoRecycle(Cache *c, Callback cb) { |
| 180 | while (QuarantineBatch *b = c->DequeueBatch()) { |
| 181 | const uptr kPrefetch = 16; |
| 182 | CHECK(kPrefetch <= ARRAY_SIZE(b->batch)); |
| 183 | for (uptr i = 0; i < kPrefetch; i++) |
| 184 | PREFETCH(b->batch[i]); |
| 185 | for (uptr i = 0, count = b->count; i < count; i++) { |
| 186 | if (i + kPrefetch < count) |
| 187 | PREFETCH(b->batch[i + kPrefetch]); |
| 188 | cb.Recycle((Node*)b->batch[i]); |
| 189 | } |
| 190 | cb.Deallocate(b); |
| 191 | } |
| 192 | } |
| 193 | }; |
| 194 | |
| 195 | // Per-thread cache of memory blocks. |
| 196 | template<typename Callback> |
| 197 | class QuarantineCache { |
| 198 | public: |
| 199 | explicit QuarantineCache(LinkerInitialized) { |
| 200 | } |
| 201 | |
| 202 | QuarantineCache() |
| 203 | : size_() { |
| 204 | list_.clear(); |
| 205 | } |
| 206 | |
| 207 | // Total memory used, including internal accounting. |
| 208 | uptr Size() const { |
| 209 | return atomic_load_relaxed(a: &size_); |
| 210 | } |
| 211 | |
| 212 | // Memory used for internal accounting. |
| 213 | uptr OverheadSize() const { |
| 214 | return list_.size() * sizeof(QuarantineBatch); |
| 215 | } |
| 216 | |
| 217 | void Enqueue(Callback cb, void *ptr, uptr size) { |
| 218 | if (list_.empty() || list_.back()->count == QuarantineBatch::kSize) { |
| 219 | QuarantineBatch *b = (QuarantineBatch *)cb.Allocate(sizeof(*b)); |
| 220 | CHECK(b); |
| 221 | b->init(ptr, size); |
| 222 | EnqueueBatch(b); |
| 223 | } else { |
| 224 | list_.back()->push_back(ptr, size); |
| 225 | SizeAdd(add: size); |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | void Transfer(QuarantineCache *from_cache) { |
| 230 | list_.append_back(l: &from_cache->list_); |
| 231 | SizeAdd(add: from_cache->Size()); |
| 232 | |
| 233 | atomic_store_relaxed(&from_cache->size_, 0); |
| 234 | } |
| 235 | |
| 236 | void EnqueueBatch(QuarantineBatch *b) { |
| 237 | list_.push_back(x: b); |
| 238 | SizeAdd(add: b->size); |
| 239 | } |
| 240 | |
| 241 | QuarantineBatch *DequeueBatch() { |
| 242 | if (list_.empty()) |
| 243 | return nullptr; |
| 244 | QuarantineBatch *b = list_.front(); |
| 245 | list_.pop_front(); |
| 246 | SizeSub(sub: b->size); |
| 247 | return b; |
| 248 | } |
| 249 | |
| 250 | void MergeBatches(QuarantineCache *to_deallocate) { |
| 251 | uptr = 0; |
| 252 | QuarantineBatch *current = list_.front(); |
| 253 | while (current && current->next) { |
| 254 | if (current->can_merge(from: current->next)) { |
| 255 | QuarantineBatch * = current->next; |
| 256 | // Move all the chunks into the current batch. |
| 257 | current->merge(from: extracted); |
| 258 | CHECK_EQ(extracted->count, 0); |
| 259 | CHECK_EQ(extracted->size, sizeof(QuarantineBatch)); |
| 260 | // Remove the next batch from the list and account for its size. |
| 261 | list_.extract(prev: current, x: extracted); |
| 262 | extracted_size += extracted->size; |
| 263 | // Add it to deallocation list. |
| 264 | to_deallocate->EnqueueBatch(extracted); |
| 265 | } else { |
| 266 | current = current->next; |
| 267 | } |
| 268 | } |
| 269 | SizeSub(sub: extracted_size); |
| 270 | } |
| 271 | |
| 272 | void PrintStats() const { |
| 273 | uptr batch_count = 0; |
| 274 | uptr total_overhead_bytes = 0; |
| 275 | uptr total_bytes = 0; |
| 276 | uptr total_quarantine_chunks = 0; |
| 277 | for (List::ConstIterator it = list_.begin(); it != list_.end(); ++it) { |
| 278 | batch_count++; |
| 279 | total_bytes += (*it).size; |
| 280 | total_overhead_bytes += (*it).size - (*it).quarantined_size(); |
| 281 | total_quarantine_chunks += (*it).count; |
| 282 | } |
| 283 | uptr quarantine_chunks_capacity = batch_count * QuarantineBatch::kSize; |
| 284 | int chunks_usage_percent = quarantine_chunks_capacity == 0 ? |
| 285 | 0 : total_quarantine_chunks * 100 / quarantine_chunks_capacity; |
| 286 | uptr total_quarantined_bytes = total_bytes - total_overhead_bytes; |
| 287 | int memory_overhead_percent = total_quarantined_bytes == 0 ? |
| 288 | 0 : total_overhead_bytes * 100 / total_quarantined_bytes; |
| 289 | Printf(format: "Global quarantine stats: batches: %zd; bytes: %zd (user: %zd); " |
| 290 | "chunks: %zd (capacity: %zd); %d%% chunks used; %d%% memory overhead" |
| 291 | "\n" , |
| 292 | batch_count, total_bytes, total_quarantined_bytes, |
| 293 | total_quarantine_chunks, quarantine_chunks_capacity, |
| 294 | chunks_usage_percent, memory_overhead_percent); |
| 295 | } |
| 296 | |
| 297 | private: |
| 298 | typedef IntrusiveList<QuarantineBatch> List; |
| 299 | |
| 300 | List list_; |
| 301 | atomic_uintptr_t size_; |
| 302 | |
| 303 | void SizeAdd(uptr add) { |
| 304 | atomic_store_relaxed(&size_, Size() + add); |
| 305 | } |
| 306 | void SizeSub(uptr sub) { |
| 307 | atomic_store_relaxed(&size_, Size() - sub); |
| 308 | } |
| 309 | }; |
| 310 | |
| 311 | } // namespace __sanitizer |
| 312 | |
| 313 | #endif // SANITIZER_QUARANTINE_H |
| 314 | |