mutex-design.rst (6160B)
1======================= 2Generic Mutex Subsystem 3======================= 4 5started by Ingo Molnar <mingo@redhat.com> 6 7updated by Davidlohr Bueso <davidlohr@hp.com> 8 9What are mutexes? 10----------------- 11 12In the Linux kernel, mutexes refer to a particular locking primitive 13that enforces serialization on shared memory systems, and not only to 14the generic term referring to 'mutual exclusion' found in academia 15or similar theoretical text books. Mutexes are sleeping locks which 16behave similarly to binary semaphores, and were introduced in 2006[1] 17as an alternative to these. This new data structure provided a number 18of advantages, including simpler interfaces, and at that time smaller 19code (see Disadvantages). 20 21[1] https://lwn.net/Articles/164802/ 22 23Implementation 24-------------- 25 26Mutexes are represented by 'struct mutex', defined in include/linux/mutex.h 27and implemented in kernel/locking/mutex.c. These locks use an atomic variable 28(->owner) to keep track of the lock state during its lifetime. Field owner 29actually contains `struct task_struct *` to the current lock owner and it is 30therefore NULL if not currently owned. Since task_struct pointers are aligned 31to at least L1_CACHE_BYTES, low bits (3) are used to store extra state (e.g., 32if waiter list is non-empty). In its most basic form it also includes a 33wait-queue and a spinlock that serializes access to it. Furthermore, 34CONFIG_MUTEX_SPIN_ON_OWNER=y systems use a spinner MCS lock (->osq), described 35below in (ii). 36 37When acquiring a mutex, there are three possible paths that can be 38taken, depending on the state of the lock: 39 40(i) fastpath: tries to atomically acquire the lock by cmpxchg()ing the owner with 41 the current task. This only works in the uncontended case (cmpxchg() checks 42 against 0UL, so all 3 state bits above have to be 0). If the lock is 43 contended it goes to the next possible path. 44 45(ii) midpath: aka optimistic spinning, tries to spin for acquisition 46 while the lock owner is running and there are no other tasks ready 47 to run that have higher priority (need_resched). The rationale is 48 that if the lock owner is running, it is likely to release the lock 49 soon. The mutex spinners are queued up using MCS lock so that only 50 one spinner can compete for the mutex. 51 52 The MCS lock (proposed by Mellor-Crummey and Scott) is a simple spinlock 53 with the desirable properties of being fair and with each cpu trying 54 to acquire the lock spinning on a local variable. It avoids expensive 55 cacheline bouncing that common test-and-set spinlock implementations 56 incur. An MCS-like lock is specially tailored for optimistic spinning 57 for sleeping lock implementation. An important feature of the customized 58 MCS lock is that it has the extra property that spinners are able to exit 59 the MCS spinlock queue when they need to reschedule. This further helps 60 avoid situations where MCS spinners that need to reschedule would continue 61 waiting to spin on mutex owner, only to go directly to slowpath upon 62 obtaining the MCS lock. 63 64 65(iii) slowpath: last resort, if the lock is still unable to be acquired, 66 the task is added to the wait-queue and sleeps until woken up by the 67 unlock path. Under normal circumstances it blocks as TASK_UNINTERRUPTIBLE. 68 69While formally kernel mutexes are sleepable locks, it is path (ii) that 70makes them more practically a hybrid type. By simply not interrupting a 71task and busy-waiting for a few cycles instead of immediately sleeping, 72the performance of this lock has been seen to significantly improve a 73number of workloads. Note that this technique is also used for rw-semaphores. 74 75Semantics 76--------- 77 78The mutex subsystem checks and enforces the following rules: 79 80 - Only one task can hold the mutex at a time. 81 - Only the owner can unlock the mutex. 82 - Multiple unlocks are not permitted. 83 - Recursive locking/unlocking is not permitted. 84 - A mutex must only be initialized via the API (see below). 85 - A task may not exit with a mutex held. 86 - Memory areas where held locks reside must not be freed. 87 - Held mutexes must not be reinitialized. 88 - Mutexes may not be used in hardware or software interrupt 89 contexts such as tasklets and timers. 90 91These semantics are fully enforced when CONFIG DEBUG_MUTEXES is enabled. 92In addition, the mutex debugging code also implements a number of other 93features that make lock debugging easier and faster: 94 95 - Uses symbolic names of mutexes, whenever they are printed 96 in debug output. 97 - Point-of-acquire tracking, symbolic lookup of function names, 98 list of all locks held in the system, printout of them. 99 - Owner tracking. 100 - Detects self-recursing locks and prints out all relevant info. 101 - Detects multi-task circular deadlocks and prints out all affected 102 locks and tasks (and only those tasks). 103 104 105Interfaces 106---------- 107Statically define the mutex:: 108 109 DEFINE_MUTEX(name); 110 111Dynamically initialize the mutex:: 112 113 mutex_init(mutex); 114 115Acquire the mutex, uninterruptible:: 116 117 void mutex_lock(struct mutex *lock); 118 void mutex_lock_nested(struct mutex *lock, unsigned int subclass); 119 int mutex_trylock(struct mutex *lock); 120 121Acquire the mutex, interruptible:: 122 123 int mutex_lock_interruptible_nested(struct mutex *lock, 124 unsigned int subclass); 125 int mutex_lock_interruptible(struct mutex *lock); 126 127Acquire the mutex, interruptible, if dec to 0:: 128 129 int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock); 130 131Unlock the mutex:: 132 133 void mutex_unlock(struct mutex *lock); 134 135Test if the mutex is taken:: 136 137 int mutex_is_locked(struct mutex *lock); 138 139Disadvantages 140------------- 141 142Unlike its original design and purpose, 'struct mutex' is among the largest 143locks in the kernel. E.g: on x86-64 it is 32 bytes, where 'struct semaphore' 144is 24 bytes and rw_semaphore is 40 bytes. Larger structure sizes mean more CPU 145cache and memory footprint. 146 147When to use mutexes 148------------------- 149 150Unless the strict semantics of mutexes are unsuitable and/or the critical 151region prevents the lock from being shared, always prefer them to any other 152locking primitive.