1 | //===-- memprof_allocator.cpp --------------------------------------------===// |
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 | // This file is a part of MemProfiler, a memory profiler. |
10 | // |
11 | // Implementation of MemProf's memory allocator, which uses the allocator |
12 | // from sanitizer_common. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "memprof_allocator.h" |
17 | #include "memprof_mapping.h" |
18 | #include "memprof_mibmap.h" |
19 | #include "memprof_rawprofile.h" |
20 | #include "memprof_stack.h" |
21 | #include "memprof_thread.h" |
22 | #include "profile/MemProfData.inc" |
23 | #include "sanitizer_common/sanitizer_allocator_checks.h" |
24 | #include "sanitizer_common/sanitizer_allocator_interface.h" |
25 | #include "sanitizer_common/sanitizer_allocator_report.h" |
26 | #include "sanitizer_common/sanitizer_array_ref.h" |
27 | #include "sanitizer_common/sanitizer_common.h" |
28 | #include "sanitizer_common/sanitizer_errno.h" |
29 | #include "sanitizer_common/sanitizer_file.h" |
30 | #include "sanitizer_common/sanitizer_flags.h" |
31 | #include "sanitizer_common/sanitizer_internal_defs.h" |
32 | #include "sanitizer_common/sanitizer_stackdepot.h" |
33 | |
34 | #include <sched.h> |
35 | #include <time.h> |
36 | |
37 | #define MAX_HISTOGRAM_PRINT_SIZE 32U |
38 | |
39 | extern bool __memprof_histogram; |
40 | |
41 | namespace __memprof { |
42 | namespace { |
43 | using ::llvm::memprof::MemInfoBlock; |
44 | |
45 | void Print(const MemInfoBlock &M, const u64 id, bool print_terse) { |
46 | u64 p; |
47 | |
48 | if (print_terse) { |
49 | p = M.TotalSize * 100 / M.AllocCount; |
50 | Printf(format: "MIB:%llu/%u/%llu.%02llu/%u/%u/" , id, M.AllocCount, p / 100, p % 100, |
51 | M.MinSize, M.MaxSize); |
52 | p = M.TotalAccessCount * 100 / M.AllocCount; |
53 | Printf(format: "%llu.%02llu/%llu/%llu/" , p / 100, p % 100, M.MinAccessCount, |
54 | M.MaxAccessCount); |
55 | p = M.TotalLifetime * 100 / M.AllocCount; |
56 | Printf(format: "%llu.%02llu/%u/%u/" , p / 100, p % 100, M.MinLifetime, |
57 | M.MaxLifetime); |
58 | Printf(format: "%u/%u/%u/%u\n" , M.NumMigratedCpu, M.NumLifetimeOverlaps, |
59 | M.NumSameAllocCpu, M.NumSameDeallocCpu); |
60 | } else { |
61 | p = M.TotalSize * 100 / M.AllocCount; |
62 | Printf(format: "Memory allocation stack id = %llu\n" , id); |
63 | Printf(format: "\talloc_count %u, size (ave/min/max) %llu.%02llu / %u / %u\n" , |
64 | M.AllocCount, p / 100, p % 100, M.MinSize, M.MaxSize); |
65 | p = M.TotalAccessCount * 100 / M.AllocCount; |
66 | Printf(format: "\taccess_count (ave/min/max): %llu.%02llu / %llu / %llu\n" , p / 100, |
67 | p % 100, M.MinAccessCount, M.MaxAccessCount); |
68 | p = M.TotalLifetime * 100 / M.AllocCount; |
69 | Printf(format: "\tlifetime (ave/min/max): %llu.%02llu / %u / %u\n" , p / 100, |
70 | p % 100, M.MinLifetime, M.MaxLifetime); |
71 | Printf(format: "\tnum migrated: %u, num lifetime overlaps: %u, num same alloc " |
72 | "cpu: %u, num same dealloc_cpu: %u\n" , |
73 | M.NumMigratedCpu, M.NumLifetimeOverlaps, M.NumSameAllocCpu, |
74 | M.NumSameDeallocCpu); |
75 | Printf(format: "AccessCountHistogram[%u]: " , M.AccessHistogramSize); |
76 | uint32_t PrintSize = M.AccessHistogramSize > MAX_HISTOGRAM_PRINT_SIZE |
77 | ? MAX_HISTOGRAM_PRINT_SIZE |
78 | : M.AccessHistogramSize; |
79 | for (size_t i = 0; i < PrintSize; ++i) { |
80 | Printf(format: "%llu " , ((uint64_t *)M.AccessHistogram)[i]); |
81 | } |
82 | Printf(format: "\n" ); |
83 | } |
84 | } |
85 | } // namespace |
86 | |
87 | static int GetCpuId(void) { |
88 | // _memprof_preinit is called via the preinit_array, which subsequently calls |
89 | // malloc. Since this is before _dl_init calls VDSO_SETUP, sched_getcpu |
90 | // will seg fault as the address of __vdso_getcpu will be null. |
91 | if (!memprof_inited) |
92 | return -1; |
93 | return sched_getcpu(); |
94 | } |
95 | |
96 | // Compute the timestamp in ms. |
97 | static int GetTimestamp(void) { |
98 | // timespec_get will segfault if called from dl_init |
99 | if (!memprof_timestamp_inited) { |
100 | // By returning 0, this will be effectively treated as being |
101 | // timestamped at memprof init time (when memprof_init_timestamp_s |
102 | // is initialized). |
103 | return 0; |
104 | } |
105 | timespec ts; |
106 | clock_gettime(CLOCK_REALTIME, tp: &ts); |
107 | return (ts.tv_sec - memprof_init_timestamp_s) * 1000 + ts.tv_nsec / 1000000; |
108 | } |
109 | |
110 | static MemprofAllocator &get_allocator(); |
111 | |
112 | // The memory chunk allocated from the underlying allocator looks like this: |
113 | // H H U U U U U U |
114 | // H -- ChunkHeader (32 bytes) |
115 | // U -- user memory. |
116 | |
117 | // If there is left padding before the ChunkHeader (due to use of memalign), |
118 | // we store a magic value in the first uptr word of the memory block and |
119 | // store the address of ChunkHeader in the next uptr. |
120 | // M B L L L L L L L L L H H U U U U U U |
121 | // | ^ |
122 | // ---------------------| |
123 | // M -- magic value kAllocBegMagic |
124 | // B -- address of ChunkHeader pointing to the first 'H' |
125 | |
126 | constexpr uptr kMaxAllowedMallocBits = 40; |
127 | |
128 | // Should be no more than 32-bytes |
129 | struct { |
130 | // 1-st 4 bytes. |
131 | u32 ; |
132 | // 2-nd 4 bytes |
133 | u32 ; |
134 | // 3-rd 4 bytes |
135 | u32 ; |
136 | // 4-th 4 bytes |
137 | // Note only 1 bit is needed for this flag if we need space in the future for |
138 | // more fields. |
139 | u32 ; |
140 | // 5-th and 6-th 4 bytes |
141 | // The max size of an allocation is 2^40 (kMaxAllowedMallocSize), so this |
142 | // could be shrunk to kMaxAllowedMallocBits if we need space in the future for |
143 | // more fields. |
144 | atomic_uint64_t ; |
145 | // 23 bits available |
146 | // 7-th and 8-th 4 bytes |
147 | u64 ; // TODO: hash of type name |
148 | }; |
149 | |
150 | static const uptr = sizeof(ChunkHeader); |
151 | COMPILER_CHECK(kChunkHeaderSize == 32); |
152 | |
153 | struct MemprofChunk : ChunkHeader { |
154 | uptr Beg() { return reinterpret_cast<uptr>(this) + kChunkHeaderSize; } |
155 | uptr UsedSize() { |
156 | return atomic_load(a: &user_requested_size, mo: memory_order_relaxed); |
157 | } |
158 | void *AllocBeg() { |
159 | if (from_memalign) |
160 | return get_allocator().GetBlockBegin(p: reinterpret_cast<void *>(this)); |
161 | return reinterpret_cast<void *>(this); |
162 | } |
163 | }; |
164 | |
165 | class { |
166 | static constexpr uptr = |
167 | FIRST_32_SECOND_64(0xCC6E96B9, 0xCC6E96B9CC6E96B9ULL); |
168 | atomic_uintptr_t ; |
169 | MemprofChunk *; |
170 | |
171 | public: |
172 | MemprofChunk *() const { |
173 | return atomic_load(a: &magic, mo: memory_order_acquire) == kAllocBegMagic |
174 | ? chunk_header |
175 | : nullptr; |
176 | } |
177 | |
178 | void (MemprofChunk *p) { |
179 | if (p) { |
180 | chunk_header = p; |
181 | atomic_store(a: &magic, v: kAllocBegMagic, mo: memory_order_release); |
182 | return; |
183 | } |
184 | |
185 | uptr old = kAllocBegMagic; |
186 | if (!atomic_compare_exchange_strong(a: &magic, cmp: &old, xchg: 0, |
187 | mo: memory_order_release)) { |
188 | CHECK_EQ(old, kAllocBegMagic); |
189 | } |
190 | } |
191 | }; |
192 | |
193 | void FlushUnneededMemProfShadowMemory(uptr p, uptr size) { |
194 | // Since memprof's mapping is compacting, the shadow chunk may be |
195 | // not page-aligned, so we only flush the page-aligned portion. |
196 | ReleaseMemoryPagesToOS(beg: MemToShadow(p), end: MemToShadow(p: p + size)); |
197 | } |
198 | |
199 | void MemprofMapUnmapCallback::OnMap(uptr p, uptr size) const { |
200 | // Statistics. |
201 | MemprofStats &thread_stats = GetCurrentThreadStats(); |
202 | thread_stats.mmaps++; |
203 | thread_stats.mmaped += size; |
204 | } |
205 | |
206 | void MemprofMapUnmapCallback::OnUnmap(uptr p, uptr size) const { |
207 | // We are about to unmap a chunk of user memory. |
208 | // Mark the corresponding shadow memory as not needed. |
209 | FlushUnneededMemProfShadowMemory(p, size); |
210 | // Statistics. |
211 | MemprofStats &thread_stats = GetCurrentThreadStats(); |
212 | thread_stats.munmaps++; |
213 | thread_stats.munmaped += size; |
214 | } |
215 | |
216 | AllocatorCache *GetAllocatorCache(MemprofThreadLocalMallocStorage *ms) { |
217 | CHECK(ms); |
218 | return &ms->allocator_cache; |
219 | } |
220 | |
221 | // Accumulates the access count from the shadow for the given pointer and size. |
222 | u64 GetShadowCount(uptr p, u32 size) { |
223 | u64 *shadow = (u64 *)MEM_TO_SHADOW(p); |
224 | u64 *shadow_end = (u64 *)MEM_TO_SHADOW(p + size); |
225 | u64 count = 0; |
226 | for (; shadow <= shadow_end; shadow++) |
227 | count += *shadow; |
228 | return count; |
229 | } |
230 | |
231 | // Accumulates the access count from the shadow for the given pointer and size. |
232 | // See memprof_mapping.h for an overview on histogram counters. |
233 | u64 GetShadowCountHistogram(uptr p, u32 size) { |
234 | u8 *shadow = (u8 *)HISTOGRAM_MEM_TO_SHADOW(p); |
235 | u8 *shadow_end = (u8 *)HISTOGRAM_MEM_TO_SHADOW(p + size); |
236 | u64 count = 0; |
237 | for (; shadow <= shadow_end; shadow++) |
238 | count += *shadow; |
239 | return count; |
240 | } |
241 | |
242 | // Clears the shadow counters (when memory is allocated). |
243 | void ClearShadow(uptr addr, uptr size) { |
244 | CHECK(AddrIsAlignedByGranularity(addr)); |
245 | CHECK(AddrIsInMem(addr)); |
246 | CHECK(AddrIsAlignedByGranularity(addr + size)); |
247 | CHECK(AddrIsInMem(addr + size - SHADOW_GRANULARITY)); |
248 | CHECK(REAL(memset)); |
249 | uptr shadow_beg; |
250 | uptr shadow_end; |
251 | if (__memprof_histogram) { |
252 | shadow_beg = HISTOGRAM_MEM_TO_SHADOW(addr); |
253 | shadow_end = HISTOGRAM_MEM_TO_SHADOW(addr + size); |
254 | } else { |
255 | shadow_beg = MEM_TO_SHADOW(addr); |
256 | shadow_end = MEM_TO_SHADOW(addr + size - SHADOW_GRANULARITY) + 1; |
257 | } |
258 | |
259 | if (shadow_end - shadow_beg < common_flags()->clear_shadow_mmap_threshold) { |
260 | REAL(memset)((void *)shadow_beg, 0, shadow_end - shadow_beg); |
261 | } else { |
262 | uptr page_size = GetPageSizeCached(); |
263 | uptr page_beg = RoundUpTo(size: shadow_beg, boundary: page_size); |
264 | uptr page_end = RoundDownTo(x: shadow_end, boundary: page_size); |
265 | |
266 | if (page_beg >= page_end) { |
267 | REAL(memset)((void *)shadow_beg, 0, shadow_end - shadow_beg); |
268 | } else { |
269 | if (page_beg != shadow_beg) { |
270 | REAL(memset)((void *)shadow_beg, 0, page_beg - shadow_beg); |
271 | } |
272 | if (page_end != shadow_end) { |
273 | REAL(memset)((void *)page_end, 0, shadow_end - page_end); |
274 | } |
275 | ReserveShadowMemoryRange(beg: page_beg, end: page_end - 1, name: nullptr); |
276 | } |
277 | } |
278 | } |
279 | |
280 | struct Allocator { |
281 | static const uptr kMaxAllowedMallocSize = 1ULL << kMaxAllowedMallocBits; |
282 | |
283 | MemprofAllocator allocator; |
284 | StaticSpinMutex fallback_mutex; |
285 | AllocatorCache fallback_allocator_cache; |
286 | |
287 | uptr max_user_defined_malloc_size; |
288 | |
289 | // Holds the mapping of stack ids to MemInfoBlocks. |
290 | MIBMapTy MIBMap; |
291 | |
292 | atomic_uint8_t destructing; |
293 | atomic_uint8_t constructed; |
294 | bool print_text; |
295 | |
296 | // ------------------- Initialization ------------------------ |
297 | explicit Allocator(LinkerInitialized) : print_text(flags()->print_text) { |
298 | atomic_store_relaxed(a: &destructing, v: 0); |
299 | atomic_store_relaxed(a: &constructed, v: 1); |
300 | } |
301 | |
302 | ~Allocator() { |
303 | atomic_store_relaxed(a: &destructing, v: 1); |
304 | FinishAndWrite(); |
305 | } |
306 | |
307 | static void PrintCallback(const uptr Key, LockedMemInfoBlock *const &Value, |
308 | void *Arg) { |
309 | SpinMutexLock l(&Value->mutex); |
310 | Print(M: Value->mib, id: Key, print_terse: bool(Arg)); |
311 | } |
312 | |
313 | // See memprof_mapping.h for an overview on histogram counters. |
314 | static MemInfoBlock CreateNewMIB(uptr p, MemprofChunk *m, u64 user_size) { |
315 | if (__memprof_histogram) { |
316 | return CreateNewMIBWithHistogram(p, m, user_size); |
317 | } else { |
318 | return CreateNewMIBWithoutHistogram(p, m, user_size); |
319 | } |
320 | } |
321 | |
322 | static MemInfoBlock CreateNewMIBWithHistogram(uptr p, MemprofChunk *m, |
323 | u64 user_size) { |
324 | |
325 | u64 c = GetShadowCountHistogram(p, size: user_size); |
326 | long curtime = GetTimestamp(); |
327 | uint32_t HistogramSize = |
328 | RoundUpTo(size: user_size, HISTOGRAM_GRANULARITY) / HISTOGRAM_GRANULARITY; |
329 | uintptr_t Histogram = |
330 | (uintptr_t)InternalAlloc(size: HistogramSize * sizeof(uint64_t)); |
331 | memset(s: (void *)Histogram, c: 0, n: HistogramSize * sizeof(uint64_t)); |
332 | for (size_t i = 0; i < HistogramSize; ++i) { |
333 | u8 Counter = |
334 | *((u8 *)HISTOGRAM_MEM_TO_SHADOW(p + HISTOGRAM_GRANULARITY * i)); |
335 | ((uint64_t *)Histogram)[i] = (uint64_t)Counter; |
336 | } |
337 | MemInfoBlock newMIB(user_size, c, m->timestamp_ms, curtime, m->cpu_id, |
338 | GetCpuId(), Histogram, HistogramSize); |
339 | return newMIB; |
340 | } |
341 | |
342 | static MemInfoBlock CreateNewMIBWithoutHistogram(uptr p, MemprofChunk *m, |
343 | u64 user_size) { |
344 | u64 c = GetShadowCount(p, size: user_size); |
345 | long curtime = GetTimestamp(); |
346 | MemInfoBlock newMIB(user_size, c, m->timestamp_ms, curtime, m->cpu_id, |
347 | GetCpuId(), 0, 0); |
348 | return newMIB; |
349 | } |
350 | |
351 | void FinishAndWrite() { |
352 | if (print_text && common_flags()->print_module_map) |
353 | DumpProcessMap(); |
354 | |
355 | allocator.ForceLock(); |
356 | |
357 | InsertLiveBlocks(); |
358 | if (print_text) { |
359 | if (!flags()->print_terse) |
360 | Printf(format: "Recorded MIBs (incl. live on exit):\n" ); |
361 | MIBMap.ForEach(cb: PrintCallback, |
362 | arg: reinterpret_cast<void *>(flags()->print_terse)); |
363 | StackDepotPrintAll(); |
364 | } else { |
365 | // Serialize the contents to a raw profile. Format documented in |
366 | // memprof_rawprofile.h. |
367 | char *Buffer = nullptr; |
368 | |
369 | __sanitizer::ListOfModules List; |
370 | List.init(); |
371 | ArrayRef<LoadedModule> Modules(List.begin(), List.end()); |
372 | u64 BytesSerialized = SerializeToRawProfile(BlockCache&: MIBMap, Modules, Buffer); |
373 | CHECK(Buffer && BytesSerialized && "could not serialize to buffer" ); |
374 | report_file.Write(buffer: Buffer, length: BytesSerialized); |
375 | } |
376 | |
377 | allocator.ForceUnlock(); |
378 | } |
379 | |
380 | // Inserts any blocks which have been allocated but not yet deallocated. |
381 | void InsertLiveBlocks() { |
382 | allocator.ForEachChunk( |
383 | callback: [](uptr chunk, void *alloc) { |
384 | u64 user_requested_size; |
385 | Allocator *A = (Allocator *)alloc; |
386 | MemprofChunk *m = |
387 | A->GetMemprofChunk(alloc_beg: (void *)chunk, user_requested_size); |
388 | if (!m) |
389 | return; |
390 | uptr user_beg = ((uptr)m) + kChunkHeaderSize; |
391 | MemInfoBlock newMIB = CreateNewMIB(p: user_beg, m, user_size: user_requested_size); |
392 | InsertOrMerge(Id: m->alloc_context_id, Block: newMIB, Map&: A->MIBMap); |
393 | }, |
394 | arg: this); |
395 | } |
396 | |
397 | void InitLinkerInitialized() { |
398 | SetAllocatorMayReturnNull(common_flags()->allocator_may_return_null); |
399 | allocator.InitLinkerInitialized( |
400 | release_to_os_interval_ms: common_flags()->allocator_release_to_os_interval_ms); |
401 | max_user_defined_malloc_size = common_flags()->max_allocation_size_mb |
402 | ? common_flags()->max_allocation_size_mb |
403 | << 20 |
404 | : kMaxAllowedMallocSize; |
405 | } |
406 | |
407 | // -------------------- Allocation/Deallocation routines --------------- |
408 | void *Allocate(uptr size, uptr alignment, BufferedStackTrace *stack, |
409 | AllocType alloc_type) { |
410 | if (UNLIKELY(!memprof_inited)) |
411 | MemprofInitFromRtl(); |
412 | if (UNLIKELY(IsRssLimitExceeded())) { |
413 | if (AllocatorMayReturnNull()) |
414 | return nullptr; |
415 | ReportRssLimitExceeded(stack); |
416 | } |
417 | CHECK(stack); |
418 | const uptr min_alignment = MEMPROF_ALIGNMENT; |
419 | if (alignment < min_alignment) |
420 | alignment = min_alignment; |
421 | if (size == 0) { |
422 | // We'd be happy to avoid allocating memory for zero-size requests, but |
423 | // some programs/tests depend on this behavior and assume that malloc |
424 | // would not return NULL even for zero-size allocations. Moreover, it |
425 | // looks like operator new should never return NULL, and results of |
426 | // consecutive "new" calls must be different even if the allocated size |
427 | // is zero. |
428 | size = 1; |
429 | } |
430 | CHECK(IsPowerOfTwo(alignment)); |
431 | uptr rounded_size = RoundUpTo(size, boundary: alignment); |
432 | uptr needed_size = rounded_size + kChunkHeaderSize; |
433 | if (alignment > min_alignment) |
434 | needed_size += alignment; |
435 | CHECK(IsAligned(needed_size, min_alignment)); |
436 | if (size > kMaxAllowedMallocSize || needed_size > kMaxAllowedMallocSize || |
437 | size > max_user_defined_malloc_size) { |
438 | if (AllocatorMayReturnNull()) { |
439 | Report(format: "WARNING: MemProfiler failed to allocate 0x%zx bytes\n" , size); |
440 | return nullptr; |
441 | } |
442 | uptr malloc_limit = |
443 | Min(a: kMaxAllowedMallocSize, b: max_user_defined_malloc_size); |
444 | ReportAllocationSizeTooBig(user_size: size, max_size: malloc_limit, stack); |
445 | } |
446 | |
447 | MemprofThread *t = GetCurrentThread(); |
448 | void *allocated; |
449 | if (t) { |
450 | AllocatorCache *cache = GetAllocatorCache(ms: &t->malloc_storage()); |
451 | allocated = allocator.Allocate(cache, size: needed_size, alignment: 8); |
452 | } else { |
453 | SpinMutexLock l(&fallback_mutex); |
454 | AllocatorCache *cache = &fallback_allocator_cache; |
455 | allocated = allocator.Allocate(cache, size: needed_size, alignment: 8); |
456 | } |
457 | if (UNLIKELY(!allocated)) { |
458 | SetAllocatorOutOfMemory(); |
459 | if (AllocatorMayReturnNull()) |
460 | return nullptr; |
461 | ReportOutOfMemory(requested_size: size, stack); |
462 | } |
463 | |
464 | uptr alloc_beg = reinterpret_cast<uptr>(allocated); |
465 | uptr alloc_end = alloc_beg + needed_size; |
466 | uptr = alloc_beg + kChunkHeaderSize; |
467 | uptr user_beg = beg_plus_header; |
468 | if (!IsAligned(a: user_beg, alignment)) |
469 | user_beg = RoundUpTo(size: user_beg, boundary: alignment); |
470 | uptr user_end = user_beg + size; |
471 | CHECK_LE(user_end, alloc_end); |
472 | uptr chunk_beg = user_beg - kChunkHeaderSize; |
473 | MemprofChunk *m = reinterpret_cast<MemprofChunk *>(chunk_beg); |
474 | m->from_memalign = alloc_beg != chunk_beg; |
475 | CHECK(size); |
476 | |
477 | m->cpu_id = GetCpuId(); |
478 | m->timestamp_ms = GetTimestamp(); |
479 | m->alloc_context_id = StackDepotPut(stack: *stack); |
480 | |
481 | uptr size_rounded_down_to_granularity = |
482 | RoundDownTo(x: size, SHADOW_GRANULARITY); |
483 | if (size_rounded_down_to_granularity) |
484 | ClearShadow(addr: user_beg, size: size_rounded_down_to_granularity); |
485 | |
486 | MemprofStats &thread_stats = GetCurrentThreadStats(); |
487 | thread_stats.mallocs++; |
488 | thread_stats.malloced += size; |
489 | thread_stats.malloced_overhead += needed_size - size; |
490 | if (needed_size > SizeClassMap::kMaxSize) |
491 | thread_stats.malloc_large++; |
492 | else |
493 | thread_stats.malloced_by_size[SizeClassMap::ClassID(size: needed_size)]++; |
494 | |
495 | void *res = reinterpret_cast<void *>(user_beg); |
496 | atomic_store(a: &m->user_requested_size, v: size, mo: memory_order_release); |
497 | if (alloc_beg != chunk_beg) { |
498 | CHECK_LE(alloc_beg + sizeof(LargeChunkHeader), chunk_beg); |
499 | reinterpret_cast<LargeChunkHeader *>(alloc_beg)->Set(m); |
500 | } |
501 | RunMallocHooks(ptr: res, size); |
502 | return res; |
503 | } |
504 | |
505 | void Deallocate(void *ptr, uptr delete_size, uptr delete_alignment, |
506 | BufferedStackTrace *stack, AllocType alloc_type) { |
507 | uptr p = reinterpret_cast<uptr>(ptr); |
508 | if (p == 0) |
509 | return; |
510 | |
511 | RunFreeHooks(ptr); |
512 | |
513 | uptr chunk_beg = p - kChunkHeaderSize; |
514 | MemprofChunk *m = reinterpret_cast<MemprofChunk *>(chunk_beg); |
515 | |
516 | u64 user_requested_size = |
517 | atomic_exchange(a: &m->user_requested_size, v: 0, mo: memory_order_acquire); |
518 | if (memprof_inited && atomic_load_relaxed(a: &constructed) && |
519 | !atomic_load_relaxed(a: &destructing)) { |
520 | MemInfoBlock newMIB = this->CreateNewMIB(p, m, user_size: user_requested_size); |
521 | InsertOrMerge(Id: m->alloc_context_id, Block: newMIB, Map&: MIBMap); |
522 | } |
523 | |
524 | MemprofStats &thread_stats = GetCurrentThreadStats(); |
525 | thread_stats.frees++; |
526 | thread_stats.freed += user_requested_size; |
527 | |
528 | void *alloc_beg = m->AllocBeg(); |
529 | if (alloc_beg != m) { |
530 | // Clear the magic value, as allocator internals may overwrite the |
531 | // contents of deallocated chunk, confusing GetMemprofChunk lookup. |
532 | reinterpret_cast<LargeChunkHeader *>(alloc_beg)->Set(nullptr); |
533 | } |
534 | |
535 | MemprofThread *t = GetCurrentThread(); |
536 | if (t) { |
537 | AllocatorCache *cache = GetAllocatorCache(ms: &t->malloc_storage()); |
538 | allocator.Deallocate(cache, p: alloc_beg); |
539 | } else { |
540 | SpinMutexLock l(&fallback_mutex); |
541 | AllocatorCache *cache = &fallback_allocator_cache; |
542 | allocator.Deallocate(cache, p: alloc_beg); |
543 | } |
544 | } |
545 | |
546 | void *Reallocate(void *old_ptr, uptr new_size, BufferedStackTrace *stack) { |
547 | CHECK(old_ptr && new_size); |
548 | uptr p = reinterpret_cast<uptr>(old_ptr); |
549 | uptr chunk_beg = p - kChunkHeaderSize; |
550 | MemprofChunk *m = reinterpret_cast<MemprofChunk *>(chunk_beg); |
551 | |
552 | MemprofStats &thread_stats = GetCurrentThreadStats(); |
553 | thread_stats.reallocs++; |
554 | thread_stats.realloced += new_size; |
555 | |
556 | void *new_ptr = Allocate(size: new_size, alignment: 8, stack, alloc_type: FROM_MALLOC); |
557 | if (new_ptr) { |
558 | CHECK_NE(REAL(memcpy), nullptr); |
559 | uptr memcpy_size = Min(a: new_size, b: m->UsedSize()); |
560 | REAL(memcpy)(new_ptr, old_ptr, memcpy_size); |
561 | Deallocate(ptr: old_ptr, delete_size: 0, delete_alignment: 0, stack, alloc_type: FROM_MALLOC); |
562 | } |
563 | return new_ptr; |
564 | } |
565 | |
566 | void *Calloc(uptr nmemb, uptr size, BufferedStackTrace *stack) { |
567 | if (UNLIKELY(CheckForCallocOverflow(size, nmemb))) { |
568 | if (AllocatorMayReturnNull()) |
569 | return nullptr; |
570 | ReportCallocOverflow(count: nmemb, size, stack); |
571 | } |
572 | void *ptr = Allocate(size: nmemb * size, alignment: 8, stack, alloc_type: FROM_MALLOC); |
573 | // If the memory comes from the secondary allocator no need to clear it |
574 | // as it comes directly from mmap. |
575 | if (ptr && allocator.FromPrimary(p: ptr)) |
576 | REAL(memset)(ptr, 0, nmemb * size); |
577 | return ptr; |
578 | } |
579 | |
580 | void CommitBack(MemprofThreadLocalMallocStorage *ms) { |
581 | AllocatorCache *ac = GetAllocatorCache(ms); |
582 | allocator.SwallowCache(cache: ac); |
583 | } |
584 | |
585 | // -------------------------- Chunk lookup ---------------------- |
586 | |
587 | // Assumes alloc_beg == allocator.GetBlockBegin(alloc_beg). |
588 | MemprofChunk *GetMemprofChunk(void *alloc_beg, u64 &user_requested_size) { |
589 | if (!alloc_beg) |
590 | return nullptr; |
591 | MemprofChunk *p = reinterpret_cast<LargeChunkHeader *>(alloc_beg)->Get(); |
592 | if (!p) { |
593 | if (!allocator.FromPrimary(p: alloc_beg)) |
594 | return nullptr; |
595 | p = reinterpret_cast<MemprofChunk *>(alloc_beg); |
596 | } |
597 | // The size is reset to 0 on deallocation (and a min of 1 on |
598 | // allocation). |
599 | user_requested_size = |
600 | atomic_load(a: &p->user_requested_size, mo: memory_order_acquire); |
601 | if (user_requested_size) |
602 | return p; |
603 | return nullptr; |
604 | } |
605 | |
606 | MemprofChunk *GetMemprofChunkByAddr(uptr p, u64 &user_requested_size) { |
607 | void *alloc_beg = allocator.GetBlockBegin(p: reinterpret_cast<void *>(p)); |
608 | return GetMemprofChunk(alloc_beg, user_requested_size); |
609 | } |
610 | |
611 | uptr AllocationSize(uptr p) { |
612 | u64 user_requested_size; |
613 | MemprofChunk *m = GetMemprofChunkByAddr(p, user_requested_size); |
614 | if (!m) |
615 | return 0; |
616 | if (m->Beg() != p) |
617 | return 0; |
618 | return user_requested_size; |
619 | } |
620 | |
621 | uptr AllocationSizeFast(uptr p) { |
622 | return reinterpret_cast<MemprofChunk *>(p - kChunkHeaderSize)->UsedSize(); |
623 | } |
624 | |
625 | void Purge() { allocator.ForceReleaseToOS(); } |
626 | |
627 | void PrintStats() { allocator.PrintStats(); } |
628 | |
629 | void ForceLock() SANITIZER_NO_THREAD_SAFETY_ANALYSIS { |
630 | allocator.ForceLock(); |
631 | fallback_mutex.Lock(); |
632 | } |
633 | |
634 | void ForceUnlock() SANITIZER_NO_THREAD_SAFETY_ANALYSIS { |
635 | fallback_mutex.Unlock(); |
636 | allocator.ForceUnlock(); |
637 | } |
638 | }; |
639 | |
640 | static Allocator instance(LINKER_INITIALIZED); |
641 | |
642 | static MemprofAllocator &get_allocator() { return instance.allocator; } |
643 | |
644 | void InitializeAllocator() { instance.InitLinkerInitialized(); } |
645 | |
646 | void MemprofThreadLocalMallocStorage::CommitBack() { |
647 | instance.CommitBack(ms: this); |
648 | } |
649 | |
650 | void PrintInternalAllocatorStats() { instance.PrintStats(); } |
651 | |
652 | void memprof_free(void *ptr, BufferedStackTrace *stack, AllocType alloc_type) { |
653 | instance.Deallocate(ptr, delete_size: 0, delete_alignment: 0, stack, alloc_type); |
654 | } |
655 | |
656 | void memprof_delete(void *ptr, uptr size, uptr alignment, |
657 | BufferedStackTrace *stack, AllocType alloc_type) { |
658 | instance.Deallocate(ptr, delete_size: size, delete_alignment: alignment, stack, alloc_type); |
659 | } |
660 | |
661 | void *memprof_malloc(uptr size, BufferedStackTrace *stack) { |
662 | return SetErrnoOnNull(instance.Allocate(size, alignment: 8, stack, alloc_type: FROM_MALLOC)); |
663 | } |
664 | |
665 | void *memprof_calloc(uptr nmemb, uptr size, BufferedStackTrace *stack) { |
666 | return SetErrnoOnNull(instance.Calloc(nmemb, size, stack)); |
667 | } |
668 | |
669 | void *memprof_reallocarray(void *p, uptr nmemb, uptr size, |
670 | BufferedStackTrace *stack) { |
671 | if (UNLIKELY(CheckForCallocOverflow(size, nmemb))) { |
672 | errno = errno_ENOMEM; |
673 | if (AllocatorMayReturnNull()) |
674 | return nullptr; |
675 | ReportReallocArrayOverflow(count: nmemb, size, stack); |
676 | } |
677 | return memprof_realloc(p, size: nmemb * size, stack); |
678 | } |
679 | |
680 | void *memprof_realloc(void *p, uptr size, BufferedStackTrace *stack) { |
681 | if (!p) |
682 | return SetErrnoOnNull(instance.Allocate(size, alignment: 8, stack, alloc_type: FROM_MALLOC)); |
683 | if (size == 0) { |
684 | if (flags()->allocator_frees_and_returns_null_on_realloc_zero) { |
685 | instance.Deallocate(ptr: p, delete_size: 0, delete_alignment: 0, stack, alloc_type: FROM_MALLOC); |
686 | return nullptr; |
687 | } |
688 | // Allocate a size of 1 if we shouldn't free() on Realloc to 0 |
689 | size = 1; |
690 | } |
691 | return SetErrnoOnNull(instance.Reallocate(old_ptr: p, new_size: size, stack)); |
692 | } |
693 | |
694 | void *memprof_valloc(uptr size, BufferedStackTrace *stack) { |
695 | return SetErrnoOnNull( |
696 | instance.Allocate(size, alignment: GetPageSizeCached(), stack, alloc_type: FROM_MALLOC)); |
697 | } |
698 | |
699 | void *memprof_pvalloc(uptr size, BufferedStackTrace *stack) { |
700 | uptr PageSize = GetPageSizeCached(); |
701 | if (UNLIKELY(CheckForPvallocOverflow(size, PageSize))) { |
702 | errno = errno_ENOMEM; |
703 | if (AllocatorMayReturnNull()) |
704 | return nullptr; |
705 | ReportPvallocOverflow(size, stack); |
706 | } |
707 | // pvalloc(0) should allocate one page. |
708 | size = size ? RoundUpTo(size, boundary: PageSize) : PageSize; |
709 | return SetErrnoOnNull(instance.Allocate(size, alignment: PageSize, stack, alloc_type: FROM_MALLOC)); |
710 | } |
711 | |
712 | void *memprof_memalign(uptr alignment, uptr size, BufferedStackTrace *stack, |
713 | AllocType alloc_type) { |
714 | if (UNLIKELY(!IsPowerOfTwo(alignment))) { |
715 | errno = errno_EINVAL; |
716 | if (AllocatorMayReturnNull()) |
717 | return nullptr; |
718 | ReportInvalidAllocationAlignment(alignment, stack); |
719 | } |
720 | return SetErrnoOnNull(instance.Allocate(size, alignment, stack, alloc_type)); |
721 | } |
722 | |
723 | void *memprof_aligned_alloc(uptr alignment, uptr size, |
724 | BufferedStackTrace *stack) { |
725 | if (UNLIKELY(!CheckAlignedAllocAlignmentAndSize(alignment, size))) { |
726 | errno = errno_EINVAL; |
727 | if (AllocatorMayReturnNull()) |
728 | return nullptr; |
729 | ReportInvalidAlignedAllocAlignment(size, alignment, stack); |
730 | } |
731 | return SetErrnoOnNull(instance.Allocate(size, alignment, stack, alloc_type: FROM_MALLOC)); |
732 | } |
733 | |
734 | int memprof_posix_memalign(void **memptr, uptr alignment, uptr size, |
735 | BufferedStackTrace *stack) { |
736 | if (UNLIKELY(!CheckPosixMemalignAlignment(alignment))) { |
737 | if (AllocatorMayReturnNull()) |
738 | return errno_EINVAL; |
739 | ReportInvalidPosixMemalignAlignment(alignment, stack); |
740 | } |
741 | void *ptr = instance.Allocate(size, alignment, stack, alloc_type: FROM_MALLOC); |
742 | if (UNLIKELY(!ptr)) |
743 | // OOM error is already taken care of by Allocate. |
744 | return errno_ENOMEM; |
745 | CHECK(IsAligned((uptr)ptr, alignment)); |
746 | *memptr = ptr; |
747 | return 0; |
748 | } |
749 | |
750 | static const void *memprof_malloc_begin(const void *p) { |
751 | u64 user_requested_size; |
752 | MemprofChunk *m = |
753 | instance.GetMemprofChunkByAddr(p: (uptr)p, user_requested_size); |
754 | if (!m) |
755 | return nullptr; |
756 | if (user_requested_size == 0) |
757 | return nullptr; |
758 | |
759 | return (const void *)m->Beg(); |
760 | } |
761 | |
762 | uptr memprof_malloc_usable_size(const void *ptr) { |
763 | if (!ptr) |
764 | return 0; |
765 | uptr usable_size = instance.AllocationSize(p: reinterpret_cast<uptr>(ptr)); |
766 | return usable_size; |
767 | } |
768 | |
769 | } // namespace __memprof |
770 | |
771 | // ---------------------- Interface ---------------- {{{1 |
772 | using namespace __memprof; |
773 | |
774 | uptr __sanitizer_get_estimated_allocated_size(uptr size) { return size; } |
775 | |
776 | int __sanitizer_get_ownership(const void *p) { |
777 | return memprof_malloc_usable_size(ptr: p) != 0; |
778 | } |
779 | |
780 | const void *__sanitizer_get_allocated_begin(const void *p) { |
781 | return memprof_malloc_begin(p); |
782 | } |
783 | |
784 | uptr __sanitizer_get_allocated_size(const void *p) { |
785 | return memprof_malloc_usable_size(ptr: p); |
786 | } |
787 | |
788 | uptr __sanitizer_get_allocated_size_fast(const void *p) { |
789 | DCHECK_EQ(p, __sanitizer_get_allocated_begin(p)); |
790 | uptr ret = instance.AllocationSizeFast(p: reinterpret_cast<uptr>(p)); |
791 | DCHECK_EQ(ret, __sanitizer_get_allocated_size(p)); |
792 | return ret; |
793 | } |
794 | |
795 | void __sanitizer_purge_allocator() { instance.Purge(); } |
796 | |
797 | int __memprof_profile_dump() { |
798 | instance.FinishAndWrite(); |
799 | // In the future we may want to return non-zero if there are any errors |
800 | // detected during the dumping process. |
801 | return 0; |
802 | } |
803 | |
804 | void __memprof_profile_reset() { |
805 | if (report_file.fd != kInvalidFd && report_file.fd != kStdoutFd && |
806 | report_file.fd != kStderrFd) { |
807 | CloseFile(report_file.fd); |
808 | // Setting the file descriptor to kInvalidFd ensures that we will reopen the |
809 | // file when invoking Write again. |
810 | report_file.fd = kInvalidFd; |
811 | } |
812 | } |
813 | |