xref: /freebsd/share/man/man9/locking.9 (revision 56bc43f5d02bd0745d597db44c285bf78d083762)
1.\" Copyright (c) 2007 Julian Elischer  (julian -  freebsd org )
2.\" All rights reserved.
3.\"
4.\" Redistribution and use in source and binary forms, with or without
5.\" modification, are permitted provided that the following conditions
6.\" are met:
7.\" 1. Redistributions of source code must retain the above copyright
8.\"    notice, this list of conditions and the following disclaimer.
9.\" 2. Redistributions in binary form must reproduce the above copyright
10.\"    notice, this list of conditions and the following disclaimer in the
11.\"    documentation and/or other materials provided with the distribution.
12.\"
13.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16.\" ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17.\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18.\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19.\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20.\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21.\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22.\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23.\" SUCH DAMAGE.
24.\"
25.Dd December 28, 2025
26.Dt LOCKING 9
27.Os
28.Sh NAME
29.Nm locking
30.Nd kernel synchronization primitives
31.Sh DESCRIPTION
32The
33.Em FreeBSD
34kernel is written to run across multiple CPUs and as such provides
35several different synchronization primitives to allow developers
36to safely access and manipulate many data types.
37.Ss Mutexes
38Mutexes (also called "blocking mutexes") are the most commonly used
39synchronization primitive in the kernel.
40A thread acquires (locks) a mutex before accessing data shared with other
41threads (including interrupt threads), and releases (unlocks) it afterwards.
42If the mutex cannot be acquired, the thread requesting it will wait.
43Mutexes are adaptive by default, meaning that
44if the owner of a contended mutex is currently running on another CPU,
45then a thread attempting to acquire the mutex will spin rather than yielding
46the processor.
47Mutexes fully support priority propagation.
48.Pp
49See
50.Xr mutex 9
51for details.
52.Ss Spin Mutexes
53Spin mutexes are a variation of basic mutexes; the main difference between
54the two is that spin mutexes never block.
55Instead, they spin while waiting for the lock to be released.
56To avoid deadlock, a thread that holds a spin mutex must never yield its CPU.
57Unlike ordinary mutexes, spin mutexes disable interrupts when acquired.
58Since disabling interrupts can be expensive, they are generally slower to
59acquire and release.
60Spin mutexes should be used only when absolutely necessary,
61e.g. to protect data shared
62with interrupt filter code (see
63.Xr bus_setup_intr 9
64for details),
65or for scheduler internals.
66.Ss Mutex Pools
67With most synchronization primitives, such as mutexes, the programmer must
68provide memory to hold the primitive.
69For example, a mutex may be embedded inside the structure it protects.
70Mutex pools provide a preallocated set of mutexes to avoid this
71requirement.
72Note that mutexes from a pool may only be used as leaf locks.
73.Pp
74See
75.Xr mtx_pool 9
76for details.
77.Ss Reader/Writer Locks
78Reader/writer locks allow shared access to protected data by multiple threads
79or exclusive access by a single thread.
80The threads with shared access are known as
81.Em readers
82since they should only read the protected data.
83A thread with exclusive access is known as a
84.Em writer
85since it may modify protected data.
86.Pp
87Reader/writer locks can be treated as mutexes (see above and
88.Xr mutex 9 )
89with shared/exclusive semantics.
90Reader/writer locks support priority propagation like mutexes,
91but priority is propagated only to an exclusive holder.
92This limitation comes from the fact that shared owners
93are anonymous.
94.Pp
95See
96.Xr rwlock 9
97for details.
98.Ss Read-Mostly Locks
99Read-mostly locks are similar to
100.Em reader/writer
101locks but optimized for very infrequent write locking.
102.Em Read-mostly
103locks implement full priority propagation by tracking shared owners
104using a caller-supplied
105.Em tracker
106data structure.
107.Pp
108See
109.Xr rmlock 9
110for details.
111.Ss Sleepable Read-Mostly Locks
112Sleepable read-mostly locks are a variation on read-mostly locks.
113Threads holding an exclusive lock may sleep,
114but threads holding a shared lock may not.
115Priority is propagated to shared owners but not to exclusive owners.
116.Ss Shared/exclusive locks
117Shared/exclusive locks are similar to reader/writer locks; the main difference
118between them is that shared/exclusive locks may be held during unbounded sleep.
119Acquiring a contested shared/exclusive lock can perform an unbounded sleep.
120These locks do not support priority propagation.
121.Pp
122See
123.Xr sx 9
124for details.
125.Ss Lockmanager locks
126Lockmanager locks are sleepable shared/exclusive locks used mostly in
127.Xr VFS 9
128.Po
129as a
130.Xr vnode 9
131lock
132.Pc
133and in the buffer cache
134.Po
135.Xr BUF_LOCK 9
136.Pc .
137They have features other lock types do not have such as sleep
138timeouts, blocking upgrades,
139writer starvation avoidance, draining, and an interlock mutex,
140but this makes them complicated both to use and to implement;
141for this reason, they should be avoided.
142.Pp
143See
144.Xr lock 9
145for details.
146.Ss Non-blocking synchronization
147The kernel has two facilities,
148.Xr epoch 9
149and
150.Xr smr 9 ,
151which can be used to provide read-only access to a data structure while one or
152more writers are concurrently modifying the data structure.
153Specifically, readers using
154.Xr epoch 9
155and
156.Xr smr 9
157to synchronize accesses do not block writers, in contrast with reader/writer
158locks, and they help ensure that memory freed by writers is not reused until
159all readers which may be accessing it have finished.
160Thus, they are a useful building block in the construction of lock-free
161data structures.
162.Pp
163These facilities are difficult to use correctly and should be avoided
164in preference to traditional mutual exclusion-based synchronization,
165except when performance or non-blocking guarantees are a major concern.
166.Pp
167See
168.Xr epoch 9
169and
170.Xr smr 9
171for details.
172.Ss Counting semaphores
173Counting semaphores provide a mechanism for synchronizing access
174to a pool of resources.
175Unlike mutexes, semaphores do not have the concept of an owner,
176so they can be useful in situations where one thread needs
177to acquire a resource, and another thread needs to release it.
178They are largely deprecated.
179.Pp
180See
181.Xr sema 9
182for details.
183.Ss Condition variables
184Condition variables are used in conjunction with locks to wait for
185a condition to become true.
186A thread must hold the associated lock before calling one of the
187.Fn cv_wait ,
188functions.
189When a thread waits on a condition, the lock
190is atomically released before the thread yields the processor
191and reacquired before the function call returns.
192Condition variables may be used with blocking mutexes,
193reader/writer locks, read-mostly locks, and shared/exclusive locks.
194.Pp
195See
196.Xr condvar 9
197for details.
198.Ss Sleep/Wakeup
199The functions
200.Fn tsleep ,
201.Fn msleep ,
202.Fn msleep_spin ,
203.Fn pause ,
204.Fn wakeup ,
205and
206.Fn wakeup_one
207also handle event-based thread blocking.
208If a thread must wait for an external event, it is put to sleep by
209.Fn tsleep ,
210.Fn msleep ,
211.Fn msleep_spin ,
212or
213.Fn pause .
214Threads may also wait using one of the locking primitive sleep routines
215.Xr mtx_sleep 9 ,
216.Xr rw_sleep 9 ,
217or
218.Xr sx_sleep 9 .
219.Pp
220Unlike condition variables,
221arbitrary addresses may be used as wait channels and a dedicated
222structure does not need to be allocated.
223However, care must be taken to ensure that wait channel addresses are
224unique to an event.
225For example, the memory address of a sleepable lock such as a
226.Xr sx 9
227must not be used as a sleep channel, because the lock implementation
228will internally use the same address as a wait channel.
229.Pp
230The parameter
231.Fa chan
232is an arbitrary address that uniquely identifies the event on which
233the thread is being put to sleep.
234All threads sleeping on a single
235.Fa chan
236are woken up later by
237.Fn wakeup
238.Pq often called from inside an interrupt routine
239to indicate that the
240event the thread was blocking on has occurred.
241.Pp
242Several of the sleep functions including
243.Fn msleep ,
244.Fn msleep_spin ,
245and the locking primitive sleep routines specify an additional lock
246parameter.
247The lock will be released before sleeping and reacquired
248before the sleep routine returns.
249If
250.Fa priority
251includes the
252.Dv PDROP
253flag, then the lock will not be reacquired before returning.
254The lock is used to ensure that a condition can be checked atomically,
255and that the current thread can be suspended without missing a
256change to the condition or an associated wakeup.
257In addition, all of the sleep routines will fully drop the
258.Va Giant
259mutex
260.Pq even if recursed
261while the thread is suspended and will reacquire the
262.Va Giant
263mutex
264.Pq restoring any recursion
265before the function returns.
266.Pp
267The
268.Fn pause
269function is a special sleep function that waits for a specified
270amount of time to pass before the thread resumes execution.
271This sleep cannot be terminated early by either an explicit
272.Fn wakeup
273or a signal.
274.Pp
275See
276.Xr sleep 9
277for details.
278.Ss Giant
279Giant is a special mutex used to protect data structures that do not
280yet have their own locks.
281Since it provides semantics akin to the old
282.Xr spl 9
283interface,
284Giant has special characteristics:
285.Bl -enum
286.It
287It is recursive.
288.It
289Drivers can request that Giant be locked around them
290by not marking themselves MPSAFE.
291Note that infrastructure to do this is slowly going away as non-MPSAFE
292drivers either became properly locked or disappear.
293.It
294Giant must be locked before other non-sleepable locks.
295.It
296Giant is dropped during unbounded sleeps and reacquired after wakeup.
297.It
298There are places in the kernel that drop Giant and pick it back up
299again.
300Sleep locks will do this before sleeping.
301Parts of the network or VM code may do this as well.
302This means that you cannot count on Giant keeping other code from
303running if your code sleeps, even if you want it to.
304.El
305.Sh INTERACTIONS
306The primitives can interact and have a number of rules regarding how
307they can and can not be combined.
308Many of these rules are checked by
309.Xr witness 4 .
310.Ss Bounded vs. Unbounded Sleep
311In a bounded sleep
312.Po also referred to as
313.Dq blocking
314.Pc
315the only resource needed to resume execution of a thread
316is CPU time for the owner of a lock that the thread is waiting to acquire.
317In an unbounded sleep
318.Po
319often referred to as simply
320.Dq sleeping
321.Pc
322a thread waits for an external event or for a condition
323to become true.
324In particular,
325a dependency chain of threads in bounded sleeps should always make forward
326progress,
327since there is always CPU time available.
328This requires that no thread in a bounded sleep is waiting for a lock held
329by a thread in an unbounded sleep.
330To avoid priority inversions,
331a thread in a bounded sleep lends its priority to the owner of the lock
332that it is waiting for.
333.Pp
334The following primitives perform bounded sleeps:
335mutexes, reader/writer locks and read-mostly locks.
336.Pp
337The following primitives perform unbounded sleeps:
338sleepable read-mostly locks, shared/exclusive locks, lockmanager locks,
339counting semaphores, condition variables, and sleep/wakeup.
340.Ss General Principles
341.Bl -bullet
342.It
343It is an error to do any operation that could result in yielding the processor
344while holding a spin mutex.
345.It
346It is an error to do any operation that could result in unbounded sleep
347while holding any primitive from the 'bounded sleep' group.
348For example, it is an error to try to acquire a shared/exclusive lock while
349holding a mutex, or to try to allocate memory with M_WAITOK while holding a
350reader/writer lock.
351.Pp
352Note that the lock passed to one of the
353.Fn sleep
354or
355.Fn cv_wait
356functions is dropped before the thread enters the unbounded sleep and does
357not violate this rule.
358.It
359It is an error to do any operation that could result in yielding of
360the processor when running inside an interrupt filter.
361.It
362It is an error to do any operation that could result in unbounded sleep when
363running inside an interrupt thread.
364.El
365.Ss Interaction table
366The following table shows what you can and can not do while holding
367one of the locking primitives discussed.
368Note that
369.Dq sleep
370includes
371.Fn sema_wait ,
372.Fn sema_timedwait ,
373any of the
374.Fn cv_wait
375functions,
376and any of the
377.Fn sleep
378functions.
379.Bl -column "       You want:" "spin mtx " "mutex/rw " "rmlock " "sleep rm " "sx/lk "  -offset 3n
380.It Em "       You want:" Ta spin mtx Ta mutex/rw Ta rmlock Ta sleep rm Ta sx/lk Ta sleep
381.It Em "You have:"        Ta -------- Ta -------- Ta ------ Ta -------- Ta ----- Ta ------
382.It spin mtx  Ta \&ok Ta \&no Ta \&no Ta \&no Ta \&no Ta \&no-1
383.It mutex/rw  Ta \&ok Ta \&ok Ta \&ok Ta \&no Ta \&no Ta \&no-1
384.It rmlock    Ta \&ok Ta \&ok Ta \&ok Ta \&no Ta \&no Ta \&no-1
385.It sleep rm  Ta \&ok Ta \&ok Ta \&ok Ta \&ok-2 Ta \&ok-2 Ta \&ok-2/3
386.It sx        Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok-3
387.It lockmgr   Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok
388.El
389.Pp
390.Em *1
391There are calls that atomically release this primitive when going to sleep
392and reacquire it on wakeup
393.Po
394.Fn mtx_sleep ,
395.Fn rw_sleep ,
396.Fn msleep_spin ,
397etc.
398.Pc .
399.Pp
400.Em *2
401These cases are only allowed while holding a write lock on a sleepable
402read-mostly lock.
403.Pp
404.Em *3
405Though one can sleep while holding this lock,
406one can also use a
407.Fn sleep
408function to atomically release this primitive when going to sleep and
409reacquire it on wakeup.
410.Pp
411Note that non-blocking try operations on locks are always permitted.
412.Ss Context mode table
413The next table shows what can be used in different contexts.
414At this time this is a rather easy to remember table.
415.Bl -column "interrupt filter: " "spin mtx " "mutex/rw " "rmlock " "sleep rm " "sx/lk " -offset 3n
416.It Em "Context:"  Ta spin mtx Ta mutex/rw Ta rmlock Ta sleep rm Ta sx/lk Ta sleep
417.It interrupt filter:  Ta \&ok Ta \&no Ta \&no Ta \&no Ta \&no Ta \&no
418.It interrupt thread:  Ta \&ok Ta \&ok Ta \&ok Ta \&no Ta \&no Ta \&no
419.It callout:           Ta \&ok Ta \&ok Ta \&ok Ta \&no Ta \&no Ta \&no
420.It direct callout:    Ta \&ok Ta \&no Ta \&no Ta \&no Ta \&no Ta \&no
421.It system call:       Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok Ta \&ok
422.El
423.Sh SEE ALSO
424.Xr lockstat 1 ,
425.Xr witness 4 ,
426.Xr atomic 9 ,
427.Xr BUS_SETUP_INTR 9 ,
428.Xr callout 9 ,
429.Xr condvar 9 ,
430.Xr epoch 9 ,
431.Xr lock 9 ,
432.Xr LOCK_PROFILING 9 ,
433.Xr mtx_pool 9 ,
434.Xr mutex 9 ,
435.Xr rmlock 9 ,
436.Xr rwlock 9 ,
437.Xr sema 9 ,
438.Xr sleep 9 ,
439.Xr smr 9 ,
440.Xr sx 9
441.Sh BUGS
442There are too many locking primitives to choose from.
443