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