Fixed a potential deadlock in folly::RWSpinLock
Summary: The current UpgradedHolder implementation may lead to a deadlock when upgrading from a reader lock, because it is blocking. A reader that has failed to set the UPGRADED bit may block the winner (upgrader) that has successfully set the UPGRADED bit, while waiting to upgrade in an infinite loop without releasing its own reader lock. The upgrader needs to wait for all reader locks to be released before it can move forward. This is the code pattern: { ReadHolder reader(lock); UpgradedHolder upgrader(std::move(reader)); WriteHolder writer(std::move(upgrader)); } To avoid this to happen, removed UpgradedHolder(ReadHolder&&) constructor and the function that impelments it unlock_shared_and_lock_upgrade. This would require a programmer explicitly release a read lock before acquiring an upgrade lock, therefore avoid the above mentioned deadlock. In addition, the current folly::RWSpinLock::try_lock_shared() implementation does not check the UPGRADED bit at all. The UPGRADED bit can be used to avoid starving writers. This diff fixed this by correctly checking the UPGRADED bit. Test Plan: Passed folly unit tests. Tested in a Facebook service that uses the folly::RWSpinLock::UpgradedHolder. Reviewed By: xliux@fb.com FB internal diff: D659875
Showing
Please register or sign in to comment