On Mon, Feb 10, 2025 at 10:38:41AM +0100, Peter Zijlstra wrote: > On Thu, Feb 06, 2025 at 02:54:08AM -0800, Kumar Kartikeya Dwivedi wrote: > > > > Deadlock Detection > > ~~~~~~~~~~~~~~~~~~ > > We handle two cases of deadlocks: AA deadlocks (attempts to acquire the > > same lock again), and ABBA deadlocks (attempts to acquire two locks in > > the opposite order from two distinct threads). Variants of ABBA > > deadlocks may be encountered with more than two locks being held in the > > incorrect order. These are not diagnosed explicitly, as they reduce to > > ABBA deadlocks. > > > > Deadlock detection is triggered immediately when beginning the waiting > > loop of a lock slow path. > > > > While timeouts ensure that any waiting loops in the locking slow path > > terminate and return to the caller, it can be excessively long in some > > situations. While the default timeout is short (0.5s), a stall for this > > duration inside the kernel can set off alerts for latency-critical > > services with strict SLOs. Ideally, the kernel should recover from an > > undesired state of the lock as soon as possible. > > > > A multi-step strategy is used to recover the kernel from waiting loops > > in the locking algorithm which may fail to terminate in a bounded amount > > of time. > > > > * Each CPU maintains a table of held locks. Entries are inserted and > > removed upon entry into lock, and exit from unlock, respectively. > > * Deadlock detection for AA locks is thus simple: we have an AA > > deadlock if we find a held lock entry for the lock we’re attempting > > to acquire on the same CPU. > > * During deadlock detection for ABBA, we search through the tables of > > all other CPUs to find situations where we are holding a lock the > > remote CPU is attempting to acquire, and they are holding a lock we > > are attempting to acquire. Upon encountering such a condition, we > > report an ABBA deadlock. > > * We divide the duration between entry time point into the waiting loop > > and the timeout time point into intervals of 1 ms, and perform > > deadlock detection until timeout happens. Upon entry into the slow > > path, and then completion of each 1 ms interval, we perform detection > > of both AA and ABBA deadlocks. In the event that deadlock detection > > yields a positive result, the recovery happens sooner than the > > timeout. Otherwise, it happens as a last resort upon completion of > > the timeout. > > > > Timeouts > > ~~~~~~~~ > > Timeouts act as final line of defense against stalls for waiting loops. > > The ‘ktime_get_mono_fast_ns’ function is used to poll for the current > > time, and it is compared to the timestamp indicating the end time in the > > waiter loop. Each waiting loop is instrumented to check an extra > > condition using a macro. Internally, the macro implementation amortizes > > the checking of the timeout to avoid sampling the clock in every > > iteration. Precisely, the timeout checks are invoked every 64k > > iterations. > > > > Recovery > > ~~~~~~~~ > > I'm probably bad at reading, but I failed to find anything that > explained how you recover from a deadlock. > > Do you force unload the BPF program? Even the simple AB-BA case, CPU0 CPU1 lock-A lock-B lock-B lock-A <- just having a random lock op return -ETIMO doesn't actually solve anything. Suppose CPU1's lock-A will time out; it will have to unwind and release lock-B before CPU0 can make progress. Worse, if CPU1 isn't quick enough to unwind and release B, then CPU0's lock-B will also time out. At which point they'll both try again and you're stuck in the same place, no? Given you *have* to unwind to make progress; why not move the entire thing to a wound-wait style lock? Then you also get rid of the whole timeout mess.
