123 lines · plain
1======================2Lightweight PI-futexes3======================4 5We are calling them lightweight for 3 reasons:6 7 - in the user-space fastpath a PI-enabled futex involves no kernel work8 (or any other PI complexity) at all. No registration, no extra kernel9 calls - just pure fast atomic ops in userspace.10 11 - even in the slowpath, the system call and scheduling pattern is very12 similar to normal futexes.13 14 - the in-kernel PI implementation is streamlined around the mutex15 abstraction, with strict rules that keep the implementation16 relatively simple: only a single owner may own a lock (i.e. no17 read-write lock support), only the owner may unlock a lock, no18 recursive locking, etc.19 20Priority Inheritance - why?21---------------------------22 23The short reply: user-space PI helps achieving/improving determinism for24user-space applications. In the best-case, it can help achieve25determinism and well-bound latencies. Even in the worst-case, PI will26improve the statistical distribution of locking related application27delays.28 29The longer reply30----------------31 32Firstly, sharing locks between multiple tasks is a common programming33technique that often cannot be replaced with lockless algorithms. As we34can see it in the kernel [which is a quite complex program in itself],35lockless structures are rather the exception than the norm - the current36ratio of lockless vs. locky code for shared data structures is somewhere37between 1:10 and 1:100. Lockless is hard, and the complexity of lockless38algorithms often endangers to ability to do robust reviews of said code.39I.e. critical RT apps often choose lock structures to protect critical40data structures, instead of lockless algorithms. Furthermore, there are41cases (like shared hardware, or other resource limits) where lockless42access is mathematically impossible.43 44Media players (such as Jack) are an example of reasonable application45design with multiple tasks (with multiple priority levels) sharing46short-held locks: for example, a highprio audio playback thread is47combined with medium-prio construct-audio-data threads and low-prio48display-colory-stuff threads. Add video and decoding to the mix and49we've got even more priority levels.50 51So once we accept that synchronization objects (locks) are an52unavoidable fact of life, and once we accept that multi-task userspace53apps have a very fair expectation of being able to use locks, we've got54to think about how to offer the option of a deterministic locking55implementation to user-space.56 57Most of the technical counter-arguments against doing priority58inheritance only apply to kernel-space locks. But user-space locks are59different, there we cannot disable interrupts or make the task60non-preemptible in a critical section, so the 'use spinlocks' argument61does not apply (user-space spinlocks have the same priority inversion62problems as other user-space locking constructs). Fact is, pretty much63the only technique that currently enables good determinism for userspace64locks (such as futex-based pthread mutexes) is priority inheritance:65 66Currently (without PI), if a high-prio and a low-prio task shares a lock67[this is a quite common scenario for most non-trivial RT applications],68even if all critical sections are coded carefully to be deterministic69(i.e. all critical sections are short in duration and only execute a70limited number of instructions), the kernel cannot guarantee any71deterministic execution of the high-prio task: any medium-priority task72could preempt the low-prio task while it holds the shared lock and73executes the critical section, and could delay it indefinitely.74 75Implementation76--------------77 78As mentioned before, the userspace fastpath of PI-enabled pthread79mutexes involves no kernel work at all - they behave quite similarly to80normal futex-based locks: a 0 value means unlocked, and a value==TID81means locked. (This is the same method as used by list-based robust82futexes.) Userspace uses atomic ops to lock/unlock these mutexes without83entering the kernel.84 85To handle the slowpath, we have added two new futex ops:86 87 - FUTEX_LOCK_PI88 - FUTEX_UNLOCK_PI89 90If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to91TID fails], then FUTEX_LOCK_PI is called. The kernel does all the92remaining work: if there is no futex-queue attached to the futex address93yet then the code looks up the task that owns the futex [it has put its94own TID into the futex value], and attaches a 'PI state' structure to95the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware,96kernel-based synchronization object. The 'other' task is made the owner97of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the98futex value. Then this task tries to lock the rt-mutex, on which it99blocks. Once it returns, it has the mutex acquired, and it sets the100futex value to its own TID and returns. Userspace has no other work to101perform - it now owns the lock, and futex value contains102FUTEX_WAITERS|TID.103 104If the unlock side fastpath succeeds, [i.e. userspace manages to do a105TID -> 0 atomic transition of the futex value], then no kernel work is106triggered.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 the110behalf of userspace - and it also unlocks the attached111pi_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 quite115possible anyway, due to existing ABI properties of pthread mutexes.]116 117Also, under this scheme, 'robustness' and 'PI' are two orthogonal118properties 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 in122Documentation/locking/rt-mutex.rst.123