]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - share/man/man9/epoch.9
Merge libc++ trunk r338150, and resolve conflicts.
[FreeBSD/FreeBSD.git] / share / man / man9 / epoch.9
1 .\"
2 .\" Copyright (C) 2018 Matthew Macy <mmacy@FreeBSD.org>.
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(s), this list of conditions and the following disclaimer as
9 .\"    the first lines of this file unmodified other than the possible
10 .\"    addition of one or more copyright notices.
11 .\" 2. Redistributions in binary form must reproduce the above copyright
12 .\"    notice(s), this list of conditions and the following disclaimer in the
13 .\"    documentation and/or other materials provided with the distribution.
14 .\"
15 .\" THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
16 .\" EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
17 .\" WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
18 .\" DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
19 .\" DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
20 .\" (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
21 .\" SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
22 .\" CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 .\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 .\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
25 .\" DAMAGE.
26 .\"
27 .\" $FreeBSD$
28 .\"
29 .Dd June 25, 2018
30 .Dt EPOCH 9
31 .Os
32 .Sh NAME
33 .Nm epoch ,
34 .Nm epoch_context ,
35 .Nm epoch_alloc ,
36 .Nm epoch_free ,
37 .Nm epoch_enter ,
38 .Nm epoch_exit ,
39 .Nm epoch_wait ,
40 .Nm epoch_call ,
41 .Nm in_epoch ,
42 .Nd kernel epoch based reclamation
43 .Sh SYNOPSIS
44 .In sys/param.h
45 .In sys/proc.h
46 .In sys/epoch.h
47 .Ft epoch_t
48 .Fn epoch_alloc "int flags"
49 .Ft void
50 .Fn epoch_enter "epoch_t epoch"
51 .Ft void
52 .Fn epoch_enter_preempt "epoch_t epoch" "epoch_tracker_t et"
53 .Ft void
54 .Fn epoch_exit "epoch_t epoch"
55 .Ft void
56 .Fn epoch_exit_preempt "epoch_t epoch" "epoch_tracker_t et"
57 .Ft void
58 .Fn epoch_wait "epoch_t epoch"
59 .Ft void
60 .Fn epoch_wait_preempt "epoch_t epoch"
61 .Ft void
62 .Fn epoch_call "epoch_t epoch" "epoch_context_t ctx" "void (*callback) (epoch_context_t)"
63 .Ft int
64 .Fn in_epoch "epoch_t epoch"
65 .Sh DESCRIPTION
66 Epochs are used to guarantee liveness and immutability of data by
67 deferring reclamation and mutation until a grace period has elapsed.
68 Epochs do not have any lock ordering issues.
69 Entering and leaving an epoch section will never block.
70 .Pp
71 Epochs are allocated with
72 .Fn epoch_alloc
73 and freed with
74 .Fn epoch_free .
75 The flags passed to epoch_alloc determine whether preemption is
76 allowed during a section or not (the default), as specified by
77 EPOCH_PREEMPT.
78 Threads indicate the start of an epoch critical section by calling
79 .Fn epoch_enter .
80 The end of a critical section is indicated by calling
81 .Fn epoch_exit .
82 The _preempt variants can be used around code which requires preemption.
83 A thread can wait until a grace period has elapsed
84 since any threads have entered
85 the epoch by calling
86 .Fn epoch_wait
87 or
88 .Fn epoch_wait_preempt ,
89 depending on the epoch_type.
90 The use of a default epoch type allows one to use
91 .Fn epoch_wait
92 which is guaranteed to have much shorter completion times since
93 we know that none of the threads in an epoch section will be preempted
94 before completing its section.
95 If the thread can't sleep or is otherwise in a performance sensitive
96 path it can ensure that a grace period has elapsed by calling
97 .Fn epoch_call
98 with a callback with any work that needs to wait for an epoch to elapse.
99 Only non-sleepable locks can be acquired during a section protected by
100 .Fn epoch_enter_preempt
101 and
102 .Fn epoch_exit_preempt .
103 INVARIANTS can assert that a thread is in an epoch by using
104 .Fn in_epoch .
105 .Pp
106 The epoch API currently does not support sleeping in epoch_preempt sections.
107 A caller should never call
108 .Fn epoch_wait
109 in the middle of an epoch section for the same epoch as this will lead to a deadlock.
110 .Pp
111 By default mutexes cannot be held across
112 .Fn epoch_wait_preempt .
113 To permit this the epoch must be allocated with
114 EPOCH_LOCKED.
115 When doing this one must be cautious of creating a situation where a deadlock is
116 possible. Note that epochs are not a straight replacement for read locks.
117 Callers must use safe list and tailq traversal routines in an epoch (see ck_queue).
118 When modifying a list referenced from an epoch section safe removal
119 routines must be used and the caller can no longer modify a list entry
120 in place.
121 An item to be modified must be handled with copy on write
122 and frees must be deferred until after a grace period has elapsed.
123 .Sh RETURN VALUES
124 .Fn in_epoch curepoch
125 will return 1 if curthread is in curepoch, 0 otherwise.
126 .Sh CAVEATS
127 One must be cautious when using
128 .Fn epoch_wait_preempt
129 threads are pinned during epoch sections so if a thread in a section is then
130 preempted by a higher priority compute bound thread on that CPU it can be
131 prevented from leaving the section.
132 Thus the wait time for the waiter is
133 potentially unbounded.
134 .Sh EXAMPLES
135 Async free example:
136 Thread 1:
137 .Bd -literal
138 int
139 in_pcbladdr(struct inpcb *inp, struct in_addr *faddr, struct in_laddr *laddr,
140     struct ucred *cred)
141 {
142    /* ... */
143    epoch_enter(net_epoch);
144     CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
145         sa = ifa->ifa_addr;
146         if (sa->sa_family != AF_INET)
147             continue;
148         sin = (struct sockaddr_in *)sa;
149         if (prison_check_ip4(cred, &sin->sin_addr) == 0) {
150              ia = (struct in_ifaddr *)ifa;
151              break;
152         }
153     }
154     epoch_exit(net_epoch);
155    /* ... */
156 }
157 .Ed
158 Thread 2:
159 .Bd -literal
160 void
161 ifa_free(struct ifaddr *ifa)
162 {
163
164     if (refcount_release(&ifa->ifa_refcnt))
165         epoch_call(net_epoch, &ifa->ifa_epoch_ctx, ifa_destroy);
166 }
167
168 void
169 if_purgeaddrs(struct ifnet *ifp)
170 {
171
172     /* .... *
173     IF_ADDR_WLOCK(ifp);
174     CK_STAILQ_REMOVE(&ifp->if_addrhead, ifa, ifaddr, ifa_link);
175     IF_ADDR_WUNLOCK(ifp);
176     ifa_free(ifa);
177 }
178 .Ed
179 .Pp
180 Thread 1 traverses the ifaddr list in an epoch.
181 Thread 2 unlinks with the corresponding epoch safe macro, marks as logically free,
182 and then defers deletion.
183 More general mutation or a synchronous
184 free would have to follow a call to
185 .Fn epoch_wait .
186 .Sh ERRORS
187 None.
188 .El
189 .Sh SEE ALSO
190 .Xr locking 9 ,
191 .Xr mtx_pool 9 ,
192 .Xr mutex 9 ,
193 .Xr rwlock 9 ,
194 .Xr sema 9 ,
195 .Xr sleep 9 ,
196 .Xr sx 9 ,
197 .Xr timeout 9