/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- * vim: sw=4 ts=4 et : * This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ #ifndef mozilla_DeadlockDetector_h #define mozilla_DeadlockDetector_h #include "mozilla/Attributes.h" #include #include "plhash.h" #include "prlock.h" #include "nsTArray.h" #ifdef NS_TRACE_MALLOC # include "nsTraceMalloc.h" #endif // ifdef NS_TRACE_MALLOC namespace mozilla { // FIXME bug 456272: split this off into a convenience API on top of // nsStackWalk? class NS_COM_GLUE CallStack { private: #ifdef NS_TRACE_MALLOC typedef nsTMStackTraceID callstack_id; // needs to be a macro to avoid disturbing the backtrace # define NS_GET_BACKTRACE() NS_TraceMallocGetStackTrace() # define NS_DEADLOCK_DETECTOR_CONSTEXPR #else typedef void* callstack_id; # define NS_GET_BACKTRACE() 0 # define NS_DEADLOCK_DETECTOR_CONSTEXPR MOZ_CONSTEXPR #endif // ifdef NS_TRACE_MALLOC callstack_id mCallStack; public: /** * CallStack * *ALWAYS* *ALWAYS* *ALWAYS* call this with no arguments. This * constructor takes an argument *ONLY* so that |GET_BACKTRACE()| * can be evaluated in the stack frame of the caller, rather than * that of the constructor. * * *BEWARE*: this means that calling this constructor with no * arguments is not the same as a "default, do-nothing" * constructor: it *will* construct a backtrace. This can cause * unexpected performance issues. */ NS_DEADLOCK_DETECTOR_CONSTEXPR CallStack(const callstack_id aCallStack = NS_GET_BACKTRACE()) : mCallStack(aCallStack) { } NS_DEADLOCK_DETECTOR_CONSTEXPR CallStack(const CallStack& aFrom) : mCallStack(aFrom.mCallStack) { } CallStack& operator=(const CallStack& aFrom) { mCallStack = aFrom.mCallStack; return *this; } bool operator==(const CallStack& aOther) const { return mCallStack == aOther.mCallStack; } bool operator!=(const CallStack& aOther) const { return mCallStack != aOther.mCallStack; } // FIXME bug 456272: if this is split off, // NS_TraceMallocPrintStackTrace should be modified to print into // an nsACString void Print(FILE* f) const { #ifdef NS_TRACE_MALLOC if (this != &kNone && mCallStack) { NS_TraceMallocPrintStackTrace(f, mCallStack); return; } #endif fputs(" [stack trace unavailable]\n", f); } /** The "null" callstack. */ static const CallStack kNone; }; /** * DeadlockDetector * * The following is an approximate description of how the deadlock detector * works. * * The deadlock detector ensures that all blocking resources are * acquired according to a partial order P. One type of blocking * resource is a lock. If a lock l1 is acquired (locked) before l2, * then we say that |l1 <_P l2|. The detector flags an error if two * locks l1 and l2 have an inconsistent ordering in P; that is, if * both |l1 <_P l2| and |l2 <_P l1|. This is a potential error * because a thread acquiring l1,l2 according to the first order might * race with a thread acquiring them according to the second order. * If this happens under the right conditions, then the acquisitions * will deadlock. * * This deadlock detector doesn't know at compile-time what P is. So, * it tries to discover the order at run time. More precisely, it * finds some order P, then tries to find chains of resource * acquisitions that violate P. An example acquisition sequence, and * the orders they impose, is * l1.lock() // current chain: [ l1 ] * // order: { } * * l2.lock() // current chain: [ l1, l2 ] * // order: { l1 <_P l2 } * * l3.lock() // current chain: [ l1, l2, l3 ] * // order: { l1 <_P l2, l2 <_P l3, l1 <_P l3 } * // (note: <_P is transitive, so also |l1 <_P l3|) * * l2.unlock() // current chain: [ l1, l3 ] * // order: { l1 <_P l2, l2 <_P l3, l1 <_P l3 } * // (note: it's OK, but weird, that l2 was unlocked out * // of order. we still have l1 <_P l3). * * l2.lock() // current chain: [ l1, l3, l2 ] * // order: { l1 <_P l2, l2 <_P l3, l1 <_P l3, * l3 <_P l2 (!!!) } * BEEP BEEP! Here the detector will flag a potential error, since * l2 and l3 were used inconsistently (and potentially in ways that * would deadlock). */ template class DeadlockDetector { public: /** * ResourceAcquisition * Consists simply of a resource and the calling context from * which it was acquired. We pack this information together so * that it can be returned back to the caller when a potential * deadlock has been found. */ struct ResourceAcquisition { const T* mResource; CallStack mCallContext; ResourceAcquisition( const T* aResource, const CallStack aCallContext=CallStack::kNone) : mResource(aResource), mCallContext(aCallContext) { } ResourceAcquisition(const ResourceAcquisition& aFrom) : mResource(aFrom.mResource), mCallContext(aFrom.mCallContext) { } ResourceAcquisition& operator=(const ResourceAcquisition& aFrom) { mResource = aFrom.mResource; mCallContext = aFrom.mCallContext; return *this; } }; typedef nsTArray ResourceAcquisitionArray; private: typedef nsTArray HashEntryArray; typedef typename HashEntryArray::index_type index_type; typedef typename HashEntryArray::size_type size_type; enum { NoIndex = HashEntryArray::NoIndex }; /** * Value type for the ordering table. Contains the other * resources on which an ordering constraint |key < other| * exists. The catch is that we also store the calling context at * which the other resource was acquired; this improves the * quality of error messages when potential deadlock is detected. */ struct OrderingEntry { OrderingEntry() : mFirstSeen(CallStack::kNone), mOrderedLT() // FIXME bug 456272: set to empirical { // dep size? } ~OrderingEntry() { } CallStack mFirstSeen; // first site from which the resource appeared HashEntryArray mOrderedLT; // this <_o Other }; static void* TableAlloc(void* /*pool*/, size_t size) { return operator new(size); } static void TableFree(void* /*pool*/, void* item) { operator delete(item); } static PLHashEntry* EntryAlloc(void* /*pool*/, const void* key) { return new PLHashEntry; } static void EntryFree(void* /*pool*/, PLHashEntry* entry, unsigned flag) { delete static_cast(const_cast(entry->key)); delete static_cast(entry->value); entry->value = 0; if (HT_FREE_ENTRY == flag) delete entry; } static PLHashNumber HashKey(const void* aKey) { return NS_PTR_TO_INT32(aKey) >> 2; } static const PLHashAllocOps kAllocOps; // Hash table "interface" the rest of the code should use PLHashEntry** GetEntry(const T* aKey) { return PL_HashTableRawLookup(mOrdering, HashKey(aKey), aKey); } void PutEntry(T* aKey) { PL_HashTableAdd(mOrdering, aKey, new OrderingEntry()); } // XXX need these helper methods because OrderingEntry doesn't have // XXX access to underlying PLHashEntry /** * Add the order |aFirst <_o aSecond|. * * WARNING: this does not check whether it's sane to add this * order. In the "best" bad case, when this order already exists, * adding it anyway may unnecessarily result in O(n^2) space. In * the "worst" bad case, adding it anyway will cause * |InTransitiveClosure()| to diverge. */ void AddOrder(PLHashEntry* aLT, PLHashEntry* aGT) { static_cast(aLT->value)->mOrderedLT .InsertElementSorted(aGT); } /** * Return true iff the order |aFirst < aSecond| has been * *explicitly* added. * * Does not consider transitivity. */ bool IsOrdered(const PLHashEntry* aFirst, const PLHashEntry* aSecond) const { return NoIndex != static_cast(aFirst->value)->mOrderedLT .BinaryIndexOf(aSecond); } /** * Return a pointer to the array of all elements "that" for * which the order |this < that| has been explicitly added. * * NOTE: this does *not* consider transitive orderings. */ PLHashEntry* const* GetOrders(const PLHashEntry* aEntry) const { return static_cast(aEntry->value)->mOrderedLT .Elements(); } /** * Return the number of elements "that" for which the order * |this < that| has been explicitly added. * * NOTE: this does *not* consider transitive orderings. */ size_type NumOrders(const PLHashEntry* aEntry) const { return static_cast(aEntry->value)->mOrderedLT .Length(); } /** Make a ResourceAcquisition out of |aEntry|. */ ResourceAcquisition MakeResourceAcquisition(const PLHashEntry* aEntry) const { return ResourceAcquisition( static_cast(aEntry->key), static_cast(aEntry->value)->mFirstSeen); } // Throwaway RAII lock to make the following code safer. struct PRAutoLock { PRAutoLock(PRLock* aLock) : mLock(aLock) { PR_Lock(mLock); } ~PRAutoLock() { PR_Unlock(mLock); } PRLock* mLock; }; public: static const uint32_t kDefaultNumBuckets; /** * DeadlockDetector * Create a new deadlock detector. * * @param aNumResourcesGuess Guess at approximate number of resources * that will be checked. */ DeadlockDetector(uint32_t aNumResourcesGuess = kDefaultNumBuckets) { mOrdering = PL_NewHashTable(aNumResourcesGuess, HashKey, PL_CompareValues, PL_CompareValues, &kAllocOps, 0); if (!mOrdering) NS_RUNTIMEABORT("couldn't initialize resource ordering table"); mLock = PR_NewLock(); if (!mLock) NS_RUNTIMEABORT("couldn't allocate deadlock detector lock"); } /** * ~DeadlockDetector * * *NOT* thread safe. */ ~DeadlockDetector() { PL_HashTableDestroy(mOrdering); PR_DestroyLock(mLock); } /** * Add * Make the deadlock detector aware of |aResource|. * * WARNING: The deadlock detector owns |aResource|. * * Thread safe. * * @param aResource Resource to make deadlock detector aware of. */ void Add(T* aResource) { PRAutoLock _(mLock); PutEntry(aResource); } // Nb: implementing a Remove() method makes the detector "more // unsound." By removing a resource from the orderings, deadlocks // may be missed that would otherwise have been found. However, // removing resources possibly reduces the # of false positives, // and additionally saves space. So it's a trade off; we have // chosen to err on the side of caution and not implement Remove(). /** * CheckAcquisition This method is called after acquiring |aLast|, * but before trying to acquire |aProposed| from |aCallContext|. * It determines whether actually trying to acquire |aProposed| * will create problems. It is OK if |aLast| is NULL; this is * interpreted as |aProposed| being the thread's first acquisition * of its current chain. * * Iff acquiring |aProposed| may lead to deadlock for some thread * interleaving (including the current one!), the cyclical * dependency from which this was deduced is returned. Otherwise, * 0 is returned. * * If a potential deadlock is detected and a resource cycle is * returned, it is the *caller's* responsibility to free it. * * Thread safe. * * @param aLast Last resource acquired by calling thread (or 0). * @param aProposed Resource calling thread proposes to acquire. * @param aCallContext Calling context whence acquisiton request came. */ ResourceAcquisitionArray* CheckAcquisition(const T* aLast, const T* aProposed, const CallStack& aCallContext) { NS_ASSERTION(aProposed, "null resource"); PRAutoLock _(mLock); PLHashEntry* second = *GetEntry(aProposed); OrderingEntry* e = static_cast(second->value); if (CallStack::kNone == e->mFirstSeen) e->mFirstSeen = aCallContext; if (!aLast) // don't check if |0 < proposed|; just vamoose return 0; PLHashEntry* first = *GetEntry(aLast); // this is the crux of the deadlock detector algorithm if (first == second) { // reflexive deadlock. fastpath b/c InTransitiveClosure is // not applicable here. ResourceAcquisitionArray* cycle = new ResourceAcquisitionArray(); if (!cycle) NS_RUNTIMEABORT("can't allocate dep. cycle array"); cycle->AppendElement(MakeResourceAcquisition(first)); cycle->AppendElement(ResourceAcquisition(aProposed, aCallContext)); return cycle; } if (InTransitiveClosure(first, second)) { // we've already established |last < proposed|. all is well. return 0; } if (InTransitiveClosure(second, first)) { // the order |proposed < last| has been deduced, perhaps // transitively. we're attempting to violate that // constraint by acquiring resources in the order // |last < proposed|, and thus we may deadlock under the // right conditions. ResourceAcquisitionArray* cycle = GetDeductionChain(second, first); // show how acquiring |proposed| would complete the cycle cycle->AppendElement(ResourceAcquisition(aProposed, aCallContext)); return cycle; } // |last|, |proposed| are unordered according to our // poset. this is fine, but we now need to add this // ordering constraint. AddOrder(first, second); return 0; } /** * Return true iff |aTarget| is in the transitive closure of |aStart| * over the ordering relation `<_this'. * * @precondition |aStart != aTarget| */ bool InTransitiveClosure(const PLHashEntry* aStart, const PLHashEntry* aTarget) const { if (IsOrdered(aStart, aTarget)) return true; index_type i = 0; size_type len = NumOrders(aStart); for (const PLHashEntry* const* it = GetOrders(aStart); i < len; ++i, ++it) if (InTransitiveClosure(*it, aTarget)) return true; return false; } /** * Return an array of all resource acquisitions * aStart <_this r1 <_this r2 <_ ... <_ aTarget * from which |aStart <_this aTarget| was deduced, including * |aStart| and |aTarget|. * * Nb: there may be multiple deductions of |aStart <_this * aTarget|. This function returns the first ordering found by * depth-first search. * * Nb: |InTransitiveClosure| could be replaced by this function. * However, this one is more expensive because we record the DFS * search stack on the heap whereas the other doesn't. * * @precondition |aStart != aTarget| */ ResourceAcquisitionArray* GetDeductionChain( const PLHashEntry* aStart, const PLHashEntry* aTarget) { ResourceAcquisitionArray* chain = new ResourceAcquisitionArray(); if (!chain) NS_RUNTIMEABORT("can't allocate dep. cycle array"); chain->AppendElement(MakeResourceAcquisition(aStart)); NS_ASSERTION(GetDeductionChain_Helper(aStart, aTarget, chain), "GetDeductionChain called when there's no deadlock"); return chain; } // precondition: |aStart != aTarget| // invariant: |aStart| is the last element in |aChain| bool GetDeductionChain_Helper(const PLHashEntry* aStart, const PLHashEntry* aTarget, ResourceAcquisitionArray* aChain) { if (IsOrdered(aStart, aTarget)) { aChain->AppendElement(MakeResourceAcquisition(aTarget)); return true; } index_type i = 0; size_type len = NumOrders(aStart); for (const PLHashEntry* const* it = GetOrders(aStart); i < len; ++i, ++it) { aChain->AppendElement(MakeResourceAcquisition(*it)); if (GetDeductionChain_Helper(*it, aTarget, aChain)) return true; aChain->RemoveElementAt(aChain->Length() - 1); } return false; } /** * The partial order on resource acquisitions used by the deadlock * detector. */ PLHashTable* mOrdering; // T* -> PLHashEntry /** * Protects contentious methods. * Nb: can't use mozilla::Mutex since we are used as its deadlock * detector. */ PRLock* mLock; private: DeadlockDetector(const DeadlockDetector& aDD) MOZ_DELETE; DeadlockDetector& operator=(const DeadlockDetector& aDD) MOZ_DELETE; }; template const PLHashAllocOps DeadlockDetector::kAllocOps = { DeadlockDetector::TableAlloc, DeadlockDetector::TableFree, DeadlockDetector::EntryAlloc, DeadlockDetector::EntryFree }; template // FIXME bug 456272: tune based on average workload const uint32_t DeadlockDetector::kDefaultNumBuckets = 64; } // namespace mozilla #endif // ifndef mozilla_DeadlockDetector_h