0001 Lightweight PI-futexes
0004 We are calling them lightweight for 3 reasons:
0006 - in the user-space fastpath a PI-enabled futex involves no kernel work
0007 (or any other PI complexity) at all. No registration, no extra kernel
0008 calls - just pure fast atomic ops in userspace.
0010 - even in the slowpath, the system call and scheduling pattern is very
0011 similar to normal futexes.
0013 - the in-kernel PI implementation is streamlined around the mutex
0014 abstraction, with strict rules that keep the implementation
0015 relatively simple: only a single owner may own a lock (i.e. no
0016 read-write lock support), only the owner may unlock a lock, no
0017 recursive locking, etc.
0019 Priority Inheritance - why?
0022 The short reply: user-space PI helps achieving/improving determinism for
0023 user-space applications. In the best-case, it can help achieve
0024 determinism and well-bound latencies. Even in the worst-case, PI will
0025 improve the statistical distribution of locking related application
0028 The longer reply:
0031 Firstly, sharing locks between multiple tasks is a common programming
0032 technique that often cannot be replaced with lockless algorithms. As we
0033 can see it in the kernel [which is a quite complex program in itself],
0034 lockless structures are rather the exception than the norm - the current
0035 ratio of lockless vs. locky code for shared data structures is somewhere
0036 between 1:10 and 1:100. Lockless is hard, and the complexity of lockless
0037 algorithms often endangers to ability to do robust reviews of said code.
0038 I.e. critical RT apps often choose lock structures to protect critical
0039 data structures, instead of lockless algorithms. Furthermore, there are
0040 cases (like shared hardware, or other resource limits) where lockless
0041 access is mathematically impossible.
0043 Media players (such as Jack) are an example of reasonable application
0044 design with multiple tasks (with multiple priority levels) sharing
0045 short-held locks: for example, a highprio audio playback thread is
0046 combined with medium-prio construct-audio-data threads and low-prio
0047 display-colory-stuff threads. Add video and decoding to the mix and
0048 we've got even more priority levels.
0050 So once we accept that synchronization objects (locks) are an
0051 unavoidable fact of life, and once we accept that multi-task userspace
0052 apps have a very fair expectation of being able to use locks, we've got
0053 to think about how to offer the option of a deterministic locking
0054 implementation to user-space.
0056 Most of the technical counter-arguments against doing priority
0057 inheritance only apply to kernel-space locks. But user-space locks are
0058 different, there we cannot disable interrupts or make the task
0059 non-preemptible in a critical section, so the 'use spinlocks' argument
0060 does not apply (user-space spinlocks have the same priority inversion
0061 problems as other user-space locking constructs). Fact is, pretty much
0062 the only technique that currently enables good determinism for userspace
0063 locks (such as futex-based pthread mutexes) is priority inheritance:
0065 Currently (without PI), if a high-prio and a low-prio task shares a lock
0066 [this is a quite common scenario for most non-trivial RT applications],
0067 even if all critical sections are coded carefully to be deterministic
0068 (i.e. all critical sections are short in duration and only execute a
0069 limited number of instructions), the kernel cannot guarantee any
0070 deterministic execution of the high-prio task: any medium-priority task
0071 could preempt the low-prio task while it holds the shared lock and
0072 executes the critical section, and could delay it indefinitely.
0077 As mentioned before, the userspace fastpath of PI-enabled pthread
0078 mutexes involves no kernel work at all - they behave quite similarly to
0079 normal futex-based locks: a 0 value means unlocked, and a value==TID
0080 means locked. (This is the same method as used by list-based robust
0081 futexes.) Userspace uses atomic ops to lock/unlock these mutexes without
0082 entering the kernel.
0084 To handle the slowpath, we have added two new futex ops:
0089 If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to
0090 TID fails], then FUTEX_LOCK_PI is called. The kernel does all the
0091 remaining work: if there is no futex-queue attached to the futex address
0092 yet then the code looks up the task that owns the futex [it has put its
0093 own TID into the futex value], and attaches a 'PI state' structure to
0094 the futex-queue. The pi_state includes an rt-mutex, which is a PI-aware,
0095 kernel-based synchronization object. The 'other' task is made the owner
0096 of the rt-mutex, and the FUTEX_WAITERS bit is atomically set in the
0097 futex value. Then this task tries to lock the rt-mutex, on which it
0098 blocks. Once it returns, it has the mutex acquired, and it sets the
0099 futex value to its own TID and returns. Userspace has no other work to
0100 perform - it now owns the lock, and futex value contains
0103 If the unlock side fastpath succeeds, [i.e. userspace manages to do a
0104 TID -> 0 atomic transition of the futex value], then no kernel work is
0107 If the unlock fastpath fails (because the FUTEX_WAITERS bit is set),
0108 then FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the
0109 behalf of userspace - and it also unlocks the attached
0110 pi_state->rt_mutex and thus wakes up any potential waiters.
0112 Note that under this approach, contrary to previous PI-futex approaches,
0113 there is no prior 'registration' of a PI-futex. [which is not quite
0114 possible anyway, due to existing ABI properties of pthread mutexes.]
0116 Also, under this scheme, 'robustness' and 'PI' are two orthogonal
0117 properties of futexes, and all four combinations are possible: futex,
0118 robust-futex, PI-futex, robust+PI-futex.
0120 More details about priority inheritance can be found in