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