[Clang][SME2] Enable multi-vector loads & stores for SME2 (#75821)
[llvm-project.git] / compiler-rt / lib / gwp_asan / guarded_pool_allocator.cpp
blob9017ab7cf7ac052bb638c0b1f5de3d409bfa30d0
1 //===-- guarded_pool_allocator.cpp ------------------------------*- 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 //===----------------------------------------------------------------------===//
9 #include "gwp_asan/guarded_pool_allocator.h"
11 #include "gwp_asan/crash_handler.h"
12 #include "gwp_asan/options.h"
13 #include "gwp_asan/utilities.h"
15 #include <assert.h>
16 #include <stddef.h>
18 using AllocationMetadata = gwp_asan::AllocationMetadata;
19 using Error = gwp_asan::Error;
21 namespace gwp_asan {
22 namespace {
23 // Forward declare the pointer to the singleton version of this class.
24 // Instantiated during initialisation, this allows the signal handler
25 // to find this class in order to deduce the root cause of failures. Must not be
26 // referenced by users outside this translation unit, in order to avoid
27 // init-order-fiasco.
28 GuardedPoolAllocator *SingletonPtr = nullptr;
30 size_t roundUpTo(size_t Size, size_t Boundary) {
31 return (Size + Boundary - 1) & ~(Boundary - 1);
34 uintptr_t getPageAddr(uintptr_t Ptr, uintptr_t PageSize) {
35 return Ptr & ~(PageSize - 1);
38 bool isPowerOfTwo(uintptr_t X) { return (X & (X - 1)) == 0; }
39 } // anonymous namespace
41 // Gets the singleton implementation of this class. Thread-compatible until
42 // init() is called, thread-safe afterwards.
43 GuardedPoolAllocator *GuardedPoolAllocator::getSingleton() {
44 return SingletonPtr;
47 void GuardedPoolAllocator::init(const options::Options &Opts) {
48 // Note: We return from the constructor here if GWP-ASan is not available.
49 // This will stop heap-allocation of class members, as well as mmap() of the
50 // guarded slots.
51 if (!Opts.Enabled || Opts.SampleRate == 0 ||
52 Opts.MaxSimultaneousAllocations == 0)
53 return;
55 Check(Opts.SampleRate >= 0, "GWP-ASan Error: SampleRate is < 0.");
56 Check(Opts.SampleRate < (1 << 30), "GWP-ASan Error: SampleRate is >= 2^30.");
57 Check(Opts.MaxSimultaneousAllocations >= 0,
58 "GWP-ASan Error: MaxSimultaneousAllocations is < 0.");
60 SingletonPtr = this;
61 Backtrace = Opts.Backtrace;
63 State.VersionMagic = {{AllocatorVersionMagic::kAllocatorVersionMagic[0],
64 AllocatorVersionMagic::kAllocatorVersionMagic[1],
65 AllocatorVersionMagic::kAllocatorVersionMagic[2],
66 AllocatorVersionMagic::kAllocatorVersionMagic[3]},
67 AllocatorVersionMagic::kAllocatorVersion,
68 0};
70 State.MaxSimultaneousAllocations = Opts.MaxSimultaneousAllocations;
72 const size_t PageSize = getPlatformPageSize();
73 // getPageAddr() and roundUpTo() assume the page size to be a power of 2.
74 assert((PageSize & (PageSize - 1)) == 0);
75 State.PageSize = PageSize;
77 // Number of pages required =
78 // + MaxSimultaneousAllocations * maximumAllocationSize (N pages per slot)
79 // + MaxSimultaneousAllocations (one guard on the left side of each slot)
80 // + 1 (an extra guard page at the end of the pool, on the right side)
81 // + 1 (an extra page that's used for reporting internally-detected crashes,
82 // like double free and invalid free, to the signal handler; see
83 // raiseInternallyDetectedError() for more info)
84 size_t PoolBytesRequired =
85 PageSize * (2 + State.MaxSimultaneousAllocations) +
86 State.MaxSimultaneousAllocations * State.maximumAllocationSize();
87 assert(PoolBytesRequired % PageSize == 0);
88 void *GuardedPoolMemory = reserveGuardedPool(PoolBytesRequired);
90 size_t BytesRequired =
91 roundUpTo(State.MaxSimultaneousAllocations * sizeof(*Metadata), PageSize);
92 Metadata = reinterpret_cast<AllocationMetadata *>(
93 map(BytesRequired, kGwpAsanMetadataName));
95 // Allocate memory and set up the free pages queue.
96 BytesRequired = roundUpTo(
97 State.MaxSimultaneousAllocations * sizeof(*FreeSlots), PageSize);
98 FreeSlots =
99 reinterpret_cast<size_t *>(map(BytesRequired, kGwpAsanFreeSlotsName));
101 // Multiply the sample rate by 2 to give a good, fast approximation for (1 /
102 // SampleRate) chance of sampling.
103 if (Opts.SampleRate != 1)
104 AdjustedSampleRatePlusOne = static_cast<uint32_t>(Opts.SampleRate) * 2 + 1;
105 else
106 AdjustedSampleRatePlusOne = 2;
108 initPRNG();
109 getThreadLocals()->NextSampleCounter =
110 ((getRandomUnsigned32() % (AdjustedSampleRatePlusOne - 1)) + 1) &
111 ThreadLocalPackedVariables::NextSampleCounterMask;
113 State.GuardedPagePool = reinterpret_cast<uintptr_t>(GuardedPoolMemory);
114 State.GuardedPagePoolEnd =
115 reinterpret_cast<uintptr_t>(GuardedPoolMemory) + PoolBytesRequired;
117 if (Opts.InstallForkHandlers)
118 installAtFork();
121 void GuardedPoolAllocator::disable() {
122 PoolMutex.lock();
123 BacktraceMutex.lock();
126 void GuardedPoolAllocator::enable() {
127 PoolMutex.unlock();
128 BacktraceMutex.unlock();
131 void GuardedPoolAllocator::iterate(void *Base, size_t Size, iterate_callback Cb,
132 void *Arg) {
133 uintptr_t Start = reinterpret_cast<uintptr_t>(Base);
134 for (size_t i = 0; i < State.MaxSimultaneousAllocations; ++i) {
135 const AllocationMetadata &Meta = Metadata[i];
136 if (Meta.Addr && !Meta.IsDeallocated && Meta.Addr >= Start &&
137 Meta.Addr < Start + Size)
138 Cb(Meta.Addr, Meta.RequestedSize, Arg);
142 void GuardedPoolAllocator::uninitTestOnly() {
143 if (State.GuardedPagePool) {
144 unreserveGuardedPool();
145 State.GuardedPagePool = 0;
146 State.GuardedPagePoolEnd = 0;
148 if (Metadata) {
149 unmap(Metadata,
150 roundUpTo(State.MaxSimultaneousAllocations * sizeof(*Metadata),
151 State.PageSize));
152 Metadata = nullptr;
154 if (FreeSlots) {
155 unmap(FreeSlots,
156 roundUpTo(State.MaxSimultaneousAllocations * sizeof(*FreeSlots),
157 State.PageSize));
158 FreeSlots = nullptr;
160 *getThreadLocals() = ThreadLocalPackedVariables();
163 // Note, minimum backing allocation size in GWP-ASan is always one page, and
164 // each slot could potentially be multiple pages (but always in
165 // page-increments). Thus, for anything that requires less than page size
166 // alignment, we don't need to allocate extra padding to ensure the alignment
167 // can be met.
168 size_t GuardedPoolAllocator::getRequiredBackingSize(size_t Size,
169 size_t Alignment,
170 size_t PageSize) {
171 assert(isPowerOfTwo(Alignment) && "Alignment must be a power of two!");
172 assert(Alignment != 0 && "Alignment should be non-zero");
173 assert(Size != 0 && "Size should be non-zero");
175 if (Alignment <= PageSize)
176 return Size;
178 return Size + Alignment - PageSize;
181 uintptr_t GuardedPoolAllocator::alignUp(uintptr_t Ptr, size_t Alignment) {
182 assert(isPowerOfTwo(Alignment) && "Alignment must be a power of two!");
183 assert(Alignment != 0 && "Alignment should be non-zero");
184 if ((Ptr & (Alignment - 1)) == 0)
185 return Ptr;
187 Ptr += Alignment - (Ptr & (Alignment - 1));
188 return Ptr;
191 uintptr_t GuardedPoolAllocator::alignDown(uintptr_t Ptr, size_t Alignment) {
192 assert(isPowerOfTwo(Alignment) && "Alignment must be a power of two!");
193 assert(Alignment != 0 && "Alignment should be non-zero");
194 if ((Ptr & (Alignment - 1)) == 0)
195 return Ptr;
197 Ptr -= Ptr & (Alignment - 1);
198 return Ptr;
201 void *GuardedPoolAllocator::allocate(size_t Size, size_t Alignment) {
202 // GuardedPagePoolEnd == 0 when GWP-ASan is disabled. If we are disabled, fall
203 // back to the supporting allocator.
204 if (State.GuardedPagePoolEnd == 0) {
205 getThreadLocals()->NextSampleCounter =
206 (AdjustedSampleRatePlusOne - 1) &
207 ThreadLocalPackedVariables::NextSampleCounterMask;
208 return nullptr;
211 if (Size == 0)
212 Size = 1;
213 if (Alignment == 0)
214 Alignment = alignof(max_align_t);
216 if (!isPowerOfTwo(Alignment) || Alignment > State.maximumAllocationSize() ||
217 Size > State.maximumAllocationSize())
218 return nullptr;
220 size_t BackingSize = getRequiredBackingSize(Size, Alignment, State.PageSize);
221 if (BackingSize > State.maximumAllocationSize())
222 return nullptr;
224 // Protect against recursivity.
225 if (getThreadLocals()->RecursiveGuard)
226 return nullptr;
227 ScopedRecursiveGuard SRG;
229 size_t Index;
231 ScopedLock L(PoolMutex);
232 Index = reserveSlot();
235 if (Index == kInvalidSlotID)
236 return nullptr;
238 uintptr_t SlotStart = State.slotToAddr(Index);
239 AllocationMetadata *Meta = addrToMetadata(SlotStart);
240 uintptr_t SlotEnd = State.slotToAddr(Index) + State.maximumAllocationSize();
241 uintptr_t UserPtr;
242 // Randomly choose whether to left-align or right-align the allocation, and
243 // then apply the necessary adjustments to get an aligned pointer.
244 if (getRandomUnsigned32() % 2 == 0)
245 UserPtr = alignUp(SlotStart, Alignment);
246 else
247 UserPtr = alignDown(SlotEnd - Size, Alignment);
249 assert(UserPtr >= SlotStart);
250 assert(UserPtr + Size <= SlotEnd);
252 // If a slot is multiple pages in size, and the allocation takes up a single
253 // page, we can improve overflow detection by leaving the unused pages as
254 // unmapped.
255 const size_t PageSize = State.PageSize;
256 allocateInGuardedPool(
257 reinterpret_cast<void *>(getPageAddr(UserPtr, PageSize)),
258 roundUpTo(Size, PageSize));
260 Meta->RecordAllocation(UserPtr, Size);
262 ScopedLock UL(BacktraceMutex);
263 Meta->AllocationTrace.RecordBacktrace(Backtrace);
266 return reinterpret_cast<void *>(UserPtr);
269 void GuardedPoolAllocator::raiseInternallyDetectedError(uintptr_t Address,
270 Error E) {
271 // Disable the allocator before setting the internal failure state. In
272 // non-recoverable mode, the allocator will be permanently disabled, and so
273 // things will be accessed without locks.
274 disable();
276 // Races between internally- and externally-raised faults can happen. Right
277 // now, in this thread we've locked the allocator in order to raise an
278 // internally-detected fault, and another thread could SIGSEGV to raise an
279 // externally-detected fault. What will happen is that the other thread will
280 // wait in the signal handler, as we hold the allocator's locks from the
281 // disable() above. We'll trigger the signal handler by touching the
282 // internal-signal-raising address below, and the signal handler from our
283 // thread will get to run first as we will continue to hold the allocator
284 // locks until the enable() at the end of this function. Be careful though, if
285 // this thread receives another SIGSEGV after the disable() above, but before
286 // touching the internal-signal-raising address below, then this thread will
287 // get an "externally-raised" SIGSEGV while *also* holding the allocator
288 // locks, which means this thread's signal handler will deadlock. This could
289 // be resolved with a re-entrant lock, but asking platforms to implement this
290 // seems unnecessary given the only way to get a SIGSEGV in this critical
291 // section is either a memory safety bug in the couple lines of code below (be
292 // careful!), or someone outside uses `kill(this_thread, SIGSEGV)`, which
293 // really shouldn't happen.
295 State.FailureType = E;
296 State.FailureAddress = Address;
298 // Raise a SEGV by touching a specific address that identifies to the crash
299 // handler that this is an internally-raised fault. Changing this address?
300 // Don't forget to update __gwp_asan_get_internal_crash_address.
301 volatile char *p =
302 reinterpret_cast<char *>(State.internallyDetectedErrorFaultAddress());
303 *p = 0;
305 // This should never be reached in non-recoverable mode. Ensure that the
306 // signal handler called handleRecoverablePostCrashReport(), which was
307 // responsible for re-setting these fields.
308 assert(State.FailureType == Error::UNKNOWN);
309 assert(State.FailureAddress == 0u);
311 // In recoverable mode, the signal handler (after dumping the crash) marked
312 // the page containing the InternalFaultSegvAddress as read/writeable, to
313 // allow the second touch to succeed after returning from the signal handler.
314 // Now, we need to mark the page as non-read/write-able again, so future
315 // internal faults can be raised.
316 deallocateInGuardedPool(
317 reinterpret_cast<void *>(getPageAddr(
318 State.internallyDetectedErrorFaultAddress(), State.PageSize)),
319 State.PageSize);
321 // And now we're done with patching ourselves back up, enable the allocator.
322 enable();
325 void GuardedPoolAllocator::deallocate(void *Ptr) {
326 assert(pointerIsMine(Ptr) && "Pointer is not mine!");
327 uintptr_t UPtr = reinterpret_cast<uintptr_t>(Ptr);
328 size_t Slot = State.getNearestSlot(UPtr);
329 uintptr_t SlotStart = State.slotToAddr(Slot);
330 AllocationMetadata *Meta = addrToMetadata(UPtr);
332 // If this allocation is responsible for crash, never recycle it. Turn the
333 // deallocate() call into a no-op.
334 if (Meta->HasCrashed)
335 return;
337 if (Meta->Addr != UPtr) {
338 raiseInternallyDetectedError(UPtr, Error::INVALID_FREE);
339 return;
341 if (Meta->IsDeallocated) {
342 raiseInternallyDetectedError(UPtr, Error::DOUBLE_FREE);
343 return;
346 // Intentionally scope the mutex here, so that other threads can access the
347 // pool during the expensive markInaccessible() call.
349 ScopedLock L(PoolMutex);
351 // Ensure that the deallocation is recorded before marking the page as
352 // inaccessible. Otherwise, a racy use-after-free will have inconsistent
353 // metadata.
354 Meta->RecordDeallocation();
356 // Ensure that the unwinder is not called if the recursive flag is set,
357 // otherwise non-reentrant unwinders may deadlock.
358 if (!getThreadLocals()->RecursiveGuard) {
359 ScopedRecursiveGuard SRG;
360 ScopedLock UL(BacktraceMutex);
361 Meta->DeallocationTrace.RecordBacktrace(Backtrace);
365 deallocateInGuardedPool(reinterpret_cast<void *>(SlotStart),
366 State.maximumAllocationSize());
368 // And finally, lock again to release the slot back into the pool.
369 ScopedLock L(PoolMutex);
370 freeSlot(Slot);
373 // Thread-compatible, protected by PoolMutex.
374 static bool PreviousRecursiveGuard;
376 void GuardedPoolAllocator::preCrashReport(void *Ptr) {
377 assert(pointerIsMine(Ptr) && "Pointer is not mine!");
378 uintptr_t InternalCrashAddr = __gwp_asan_get_internal_crash_address(
379 &State, reinterpret_cast<uintptr_t>(Ptr));
380 if (!InternalCrashAddr)
381 disable();
383 // If something in the signal handler calls malloc() while dumping the
384 // GWP-ASan report (e.g. backtrace_symbols()), make sure that GWP-ASan doesn't
385 // service that allocation. `PreviousRecursiveGuard` is protected by the
386 // allocator locks taken in disable(), either explicitly above for
387 // externally-raised errors, or implicitly in raiseInternallyDetectedError()
388 // for internally-detected errors.
389 PreviousRecursiveGuard = getThreadLocals()->RecursiveGuard;
390 getThreadLocals()->RecursiveGuard = true;
393 void GuardedPoolAllocator::postCrashReportRecoverableOnly(void *SignalPtr) {
394 uintptr_t SignalUPtr = reinterpret_cast<uintptr_t>(SignalPtr);
395 uintptr_t InternalCrashAddr =
396 __gwp_asan_get_internal_crash_address(&State, SignalUPtr);
397 uintptr_t ErrorUptr = InternalCrashAddr ?: SignalUPtr;
399 AllocationMetadata *Metadata = addrToMetadata(ErrorUptr);
400 Metadata->HasCrashed = true;
402 allocateInGuardedPool(
403 reinterpret_cast<void *>(getPageAddr(SignalUPtr, State.PageSize)),
404 State.PageSize);
406 // Clear the internal state in order to not confuse the crash handler if a
407 // use-after-free or buffer-overflow comes from a different allocation in the
408 // future.
409 if (InternalCrashAddr) {
410 State.FailureType = Error::UNKNOWN;
411 State.FailureAddress = 0;
414 size_t Slot = State.getNearestSlot(ErrorUptr);
415 // If the slot is available, remove it permanently.
416 for (size_t i = 0; i < FreeSlotsLength; ++i) {
417 if (FreeSlots[i] == Slot) {
418 FreeSlots[i] = FreeSlots[FreeSlotsLength - 1];
419 FreeSlotsLength -= 1;
420 break;
424 getThreadLocals()->RecursiveGuard = PreviousRecursiveGuard;
425 if (!InternalCrashAddr)
426 enable();
429 size_t GuardedPoolAllocator::getSize(const void *Ptr) {
430 assert(pointerIsMine(Ptr));
431 ScopedLock L(PoolMutex);
432 AllocationMetadata *Meta = addrToMetadata(reinterpret_cast<uintptr_t>(Ptr));
433 assert(Meta->Addr == reinterpret_cast<uintptr_t>(Ptr));
434 return Meta->RequestedSize;
437 AllocationMetadata *GuardedPoolAllocator::addrToMetadata(uintptr_t Ptr) const {
438 return &Metadata[State.getNearestSlot(Ptr)];
441 size_t GuardedPoolAllocator::reserveSlot() {
442 // Avoid potential reuse of a slot before we have made at least a single
443 // allocation in each slot. Helps with our use-after-free detection.
444 if (NumSampledAllocations < State.MaxSimultaneousAllocations)
445 return NumSampledAllocations++;
447 if (FreeSlotsLength == 0)
448 return kInvalidSlotID;
450 size_t ReservedIndex = getRandomUnsigned32() % FreeSlotsLength;
451 size_t SlotIndex = FreeSlots[ReservedIndex];
452 FreeSlots[ReservedIndex] = FreeSlots[--FreeSlotsLength];
453 return SlotIndex;
456 void GuardedPoolAllocator::freeSlot(size_t SlotIndex) {
457 assert(FreeSlotsLength < State.MaxSimultaneousAllocations);
458 FreeSlots[FreeSlotsLength++] = SlotIndex;
461 uint32_t GuardedPoolAllocator::getRandomUnsigned32() {
462 uint32_t RandomState = getThreadLocals()->RandomState;
463 RandomState ^= RandomState << 13;
464 RandomState ^= RandomState >> 17;
465 RandomState ^= RandomState << 5;
466 getThreadLocals()->RandomState = RandomState;
467 return RandomState;
469 } // namespace gwp_asan