/* * Copyright 2001-2009 Haiku, Inc. All rights reserved. * Distributed under the terms of the MIT license. * * Authors: * Erik Jaesler, erik@cgsoftware.com */ /*! Semaphore-type class for thread safety */ #include #include #include #include #include "support_kit_config.h" // Data Member Documentation: // // The "fBenaphoreCount" member is set to 1 if the BLocker style is // semaphore. If the style is benaphore, it is initialized to 0 and // is incremented atomically when it is acquired, decremented when it // is released. By setting the benaphore count to 1 when the style is // semaphore, the benaphore effectively becomes a semaphore. I was able // to determine this is what Be's implementation does by testing the // result of the CountLockRequests() member. // // The "fSemaphoreID" member holds the sem_id returned from create_sem() // when the BLocker is constructed. It is used to acquire and release // the lock regardless of the lock style (semaphore or benaphore). // // The "fLockOwner" member holds the thread_id of the thread which // currently holds the lock. If no thread holds the lock, it is set to // B_ERROR. // // The "fRecursiveCount" member holds a count of the number of times the // thread holding the lock has acquired the lock without a matching unlock. // It is basically the number of times the thread must call Unlock() before // the lock can be acquired by a different thread. // BLocker::BLocker() { InitLocker(NULL, true); } BLocker::BLocker(const char *name) { InitLocker(name, true); } BLocker::BLocker(bool benaphoreStyle) { InitLocker(NULL, benaphoreStyle); } BLocker::BLocker(const char *name, bool benaphoreStyle) { InitLocker(name, benaphoreStyle); } /*! This constructor is not documented. The final argument is ignored for now. In Be's headers, its called "for_IPC". DO NOT USE THIS CONSTRUCTOR! */ BLocker::BLocker(const char *name, bool benaphoreStyle, bool) { InitLocker(name, benaphoreStyle); } BLocker::~BLocker() { delete_sem(fSemaphoreID); } status_t BLocker::InitCheck() const { return fSemaphoreID >= 0 ? B_OK : fSemaphoreID; } bool BLocker::Lock() { status_t result; return AcquireLock(B_INFINITE_TIMEOUT, &result); } status_t BLocker::LockWithTimeout(bigtime_t timeout) { status_t result; AcquireLock(timeout, &result); return result; } void BLocker::Unlock() { // The Be Book explicitly allows any thread, not just the lock owner, to // unlock. This is bad practice, but we must allow it for compatibility // reasons. We can at least warn the developer that something is probably // wrong. if (!IsLocked()) { fprintf(stderr, "Unlocking BLocker with sem %" B_PRId32 " from wrong thread %" B_PRId32 ", current holder %" B_PRId32 " (see issue #6400).\n", fSemaphoreID, find_thread(NULL), fLockOwner); } // Decrement the number of outstanding locks this thread holds // on this BLocker. fRecursiveCount--; // If the recursive count is now at 0, that means the BLocker has // been released by the thread. if (fRecursiveCount == 0) { // The BLocker is no longer owned by any thread. fLockOwner = B_ERROR; // Decrement the benaphore count and store the undecremented // value in oldBenaphoreCount. int32 oldBenaphoreCount = atomic_add(&fBenaphoreCount, -1); // If the oldBenaphoreCount is greater than 1, then there is // at least one thread waiting for the lock in the case of a // benaphore. if (oldBenaphoreCount > 1) { // Since there are threads waiting for the lock, it must // be released. Note, the old benaphore count will always be // greater than 1 for a semaphore so the release is always done. release_sem(fSemaphoreID); } } } thread_id BLocker::LockingThread() const { return fLockOwner; } bool BLocker::IsLocked() const { // This member returns true if the calling thread holds the lock. // The easiest way to determine this is to compare the result of // find_thread() to the fLockOwner. return find_thread(NULL) == fLockOwner; } int32 BLocker::CountLocks() const { return fRecursiveCount; } int32 BLocker::CountLockRequests() const { return fBenaphoreCount; } sem_id BLocker::Sem() const { return fSemaphoreID; } void BLocker::InitLocker(const char *name, bool benaphore) { if (name == NULL) name = "some BLocker"; if (benaphore && !BLOCKER_ALWAYS_SEMAPHORE_STYLE) { // Because this is a benaphore, initialize the benaphore count and // create the semaphore. Because this is a benaphore, the semaphore // count starts at 0 (ie acquired). fBenaphoreCount = 0; fSemaphoreID = create_sem(0, name); } else { // Because this is a semaphore, initialize the benaphore count to -1 // and create the semaphore. Because this is semaphore style, the // semaphore count starts at 1 so that one thread can acquire it and // the next thread to acquire it will block. fBenaphoreCount = 1; fSemaphoreID = create_sem(1, name); } // The lock is currently not acquired so there is no owner. fLockOwner = B_ERROR; // The lock is currently not acquired so the recursive count is zero. fRecursiveCount = 0; } bool BLocker::AcquireLock(bigtime_t timeout, status_t *error) { // By default, return no error. status_t status = B_OK; // Only try to acquire the lock if the thread doesn't already own it. if (!IsLocked()) { // Increment the benaphore count and test to see if it was already // greater than 0. If it is greater than 0, then some thread already has // the benaphore or the style is a semaphore. Either way, we need to // acquire the semaphore in this case. int32 oldBenaphoreCount = atomic_add(&fBenaphoreCount, 1); if (oldBenaphoreCount > 0) { do { status = acquire_sem_etc(fSemaphoreID, 1, B_RELATIVE_TIMEOUT, timeout); } while (status == B_INTERRUPTED); // Note, if the lock here does time out, the benaphore count // is not decremented. By doing this, the benaphore count will // never go back to zero. This means that the locking essentially // changes to semaphore style if this was a benaphore. // // Doing the decrement of the benaphore count when the acquisition // fails is a risky thing to do. If you decrement the counter at // the same time the thread which holds the benaphore does an // Unlock(), there is serious risk of a race condition. // // If the Unlock() sees a positive count and releases the semaphore // and then the timed out thread decrements the count to 0, there // is no one to take the semaphore. The next two threads will be // able to acquire the benaphore at the same time! The first will // increment the counter and acquire the lock. The second will // acquire the semaphore and therefore the lock. Not good. // // This has been discussed on the becodetalk mailing list and // Trey from Be had this to say: // // I looked at the LockWithTimeout() code, and it does not have // _this_ (ie the race condition) problem. It circumvents it by // NOT doing the atomic_add(&count, -1) if the semaphore // acquisition fails. This means that if a // BLocker::LockWithTimeout() times out, all other Lock*() attempts // turn into guaranteed semaphore grabs, _with_ the overhead of a // (now) useless atomic_add(). // // Given Trey's comments, it looks like Be took the same approach // I did. The output of CountLockRequests() of Be's implementation // confirms Trey's comments also. // // Finally some thoughts for the future with this code: // - If 2^31 timeouts occur on a 32-bit machine (ie today), // the benaphore count will wrap to a negative number. This // would have unknown consequences on the ability of the BLocker // to continue to function. // } } // If the lock has successfully been acquired. if (status == B_OK) { // Set the lock owner to this thread and increment the recursive count // by one. The recursive count is incremented because one more Unlock() // is now required to release the lock (ie, 0 => 1, 1 => 2 etc). if (fLockOwner < 0) { fLockOwner = find_thread(NULL); fRecursiveCount = 1; } else fRecursiveCount++; } if (error != NULL) *error = status; // Return true if the lock has been acquired. return (status == B_OK); }