1====================== 2Lightweight PI-futexes 3====================== 4 5We are calling them lightweight for 3 reasons: 6 7 - in the user-space fastpath a PI-enabled futex involves no kernel work 8 (or any other PI complexity) at all. No registration, no extra kernel 9 calls - just pure fast atomic ops in userspace. 10 11 - even in the slowpath, the system call and scheduling pattern is very 12 similar to normal futexes. 13 14 - the in-kernel PI implementation is streamlined around the mutex 15 abstraction, with strict rules that keep the implementation 16 relatively simple: only a single owner may own a lock (i.e. no 17 read-write lock support), only the owner may unlock a lock, no 18 recursive locking, etc. 19 20Priority Inheritance - why? 21--------------------------- 22 23The short reply: user-space PI helps achieving/improving determinism for 24user-space applications. In the best-case, it can help achieve 25determinism and well-bound latencies. Even in the worst-case, PI will 26improve the statistical distribution of locking related application 27delays. 28 29The longer reply 30---------------- 31 32Firstly, sharing locks between multiple tasks is a common programming 33technique that often cannot be replaced with lockless algorithms. As we 34can see it in the kernel [which is a quite complex program in itself], 35lockless structures are rather the exception than the norm - the current 36ratio of lockless vs. locky code for shared data structures is somewhere 37between 1:10 and 1:100. Lockless is hard, and the complexity of lockless 38algorithms often endangers to ability to do robust reviews of said code. 39I.e. critical RT apps often choose lock structures to protect critical 40data structures, instead of lockless algorithms. Furthermore, there are 41cases (like shared hardware, or other resource limits) where lockless 42access is mathematically impossible. 43 44Media players (such as Jack) are an example of reasonable application 45design with multiple tasks (with multiple priority levels) sharing 46short-held locks: for example, a highprio audio playback thread is 47combined with medium-prio construct-audio-data threads and low-prio 48display-colory-stuff threads. Add video and decoding to the mix and 49we've got even more priority levels. 50 51So once we accept that synchronization objects (locks) are an 52unavoidable fact of life, and once we accept that multi-task userspace 53apps have a very fair expectation of being able to use locks, we've got 54to think about how to offer the option of a deterministic locking 55implementation to user-space. 56 57Most of the technical counter-arguments against doing priority 58inheritance only apply to kernel-space locks. But user-space locks are 59different, there we cannot disable interrupts or make the task 60non-preemptible in a critical section, so the 'use spinlocks' argument 61does not apply (user-space spinlocks have the same priority inversion 62problems as other user-space locking constructs). Fact is, pretty much 63the only technique that currently enables good determinism for userspace 64locks (such as futex-based pthread mutexes) is priority inheritance: 65 66Currently (without PI), if a high-prio and a low-prio task shares a lock 67[this is a quite common scenario for most non-trivial RT applications], 68even if all critical sections are coded carefully to be deterministic 69(i.e. all critical sections are short in duration and only execute a 70limited number of instructions), the kernel cannot guarantee any 71deterministic execution of the high-prio task: any medium-priority task 72could preempt the low-prio task while it holds the shared lock and 73executes the critical section, and could delay it indefinitely. 74 75Implementation 76-------------- 77 78As mentioned before, the userspace fastpath of PI-enabled pthread 79mutexes involves no kernel work at all - they behave quite similarly to 80normal futex-based locks: a 0 value means unlocked, and a value==TID 81means locked. (This is the same method as used by list-based robust 82futexes.) Userspace uses atomic ops to lock/unlock these mutexes without 83entering the kernel. 84 85To handle the slowpath, we have added two new futex ops: 86 87 - FUTEX_LOCK_PI 88 - FUTEX_UNLOCK_PI 89 90If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to 91TID fails], then FUTEX_LOCK_PI is called. The kernel does all the 92remaining work: if there is no futex-queue attached to the futex address 93yet then the code looks up the task that owns the futex [it has put its 94own TID into the futex value], and attaches a 'PI state' structure to 95the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware, 96kernel-based synchronization object. The 'other' task is made the owner 97of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the 98futex value. Then this task tries to lock the rt-mutex, on which it 99blocks. Once it returns, it has the mutex acquired, and it sets the 100futex value to its own TID and returns. Userspace has no other work to 101perform - it now owns the lock, and futex value contains 102FUTEX_WAITERS|TID. 103 104If the unlock side fastpath succeeds, [i.e. userspace manages to do a 105TID -> 0 atomic transition of the futex value], then no kernel work is 106triggered. 107 108If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), 109then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the 110behalf of userspace - and it also unlocks the attached 111pi_state->rt_mutex and thus wakes up any potential waiters. 112 113Note that under this approach, contrary to previous PI-futex approaches, 114there is no prior 'registration' of a PI-futex. [which is not quite 115possible anyway, due to existing ABI properties of pthread mutexes.] 116 117Also, under this scheme, 'robustness' and 'PI' are two orthogonal 118properties of futexes, and all four combinations are possible: futex, 119robust-futex, PI-futex, robust+PI-futex. 120 121More details about priority inheritance can be found in 122Documentation/locking/rt-mutex.txt. 123
