epoch.9: Add missing functions, clean up documentation

Various rototilling.
This commit is contained in:
Conrad Meyer 2019-12-28 01:35:32 +00:00
parent 5541eb27d6
commit 2c1962aba6
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=356140

View File

@ -26,7 +26,7 @@
.\"
.\" $FreeBSD$
.\"
.Dd June 28, 2019
.Dd December 27, 2019
.Dt EPOCH 9
.Os
.Sh NAME
@ -37,34 +37,58 @@
.Nm epoch_enter ,
.Nm epoch_exit ,
.Nm epoch_wait ,
.Nm epoch_enter_preempt ,
.Nm epoch_exit_preempt ,
.Nm epoch_wait_preempt ,
.Nm epoch_call ,
.Nm epoch_drain_callbacks ,
.Nm in_epoch ,
.Nm in_epoch_verbose ,
.Nd kernel epoch based reclamation
.Sh SYNOPSIS
.In sys/param.h
.In sys/proc.h
.In sys/epoch.h
.\" Types
.Bd -literal
struct epoch; /* Opaque */
.Ed
.Vt typedef "struct epoch *epoch_t" ;
.Bd -literal
struct epoch_context {
void *data[2];
};
.Ed
.Vt typedef "struct epoch_context *epoch_context_t" ;
.Bd -literal
struct epoch_tracker; /* Opaque */
.Ed
.Vt typedef "struct epoch_tracker *epoch_tracker_t" ;
.\" Declarations
.Ft epoch_t
.Fn epoch_alloc "int flags"
.Fn epoch_alloc "const char *name" "int flags"
.Ft void
.Fn epoch_free "epoch_t epoch"
.Ft void
.Fn epoch_enter "epoch_t epoch"
.Ft void
.Fn epoch_enter_preempt "epoch_t epoch" "epoch_tracker_t et"
.Ft void
.Fn epoch_exit "epoch_t epoch"
.Ft void
.Fn epoch_exit_preempt "epoch_t epoch" "epoch_tracker_t et"
.Ft void
.Fn epoch_wait "epoch_t epoch"
.Ft void
.Fn epoch_enter_preempt "epoch_t epoch" "epoch_tracker_t et"
.Ft void
.Fn epoch_exit_preempt "epoch_t epoch" "epoch_tracker_t et"
.Ft void
.Fn epoch_wait_preempt "epoch_t epoch"
.Ft void
.Fn epoch_call "epoch_t epoch" "epoch_context_t ctx" "void (*callback) (epoch_context_t)"
.Fn epoch_call "epoch_t epoch" "epoch_context_t ctx" "void (*callback)(epoch_context_t)"
.Ft void
.Fn epoch_drain_callbacks "epoch_t epoch"
.Ft int
.Fn in_epoch "epoch_t epoch"
.Ft int
.Fn in_epoch_verbose "epoch_t epoch" "int dump_onfail"
.Sh DESCRIPTION
Epochs are used to guarantee liveness and immutability of data by
deferring reclamation and mutation until a grace period has elapsed.
@ -72,58 +96,105 @@ Epochs do not have any lock ordering issues.
Entering and leaving an epoch section will never block.
.Pp
Epochs are allocated with
.Fn epoch_alloc
and freed with
.Fn epoch_alloc .
The
.Fa name
argument is used for debugging convenience when the
.Cd EPOCH_TRACE
kernel option is configured.
By default, epochs do not allow preemption during sections.
By default mutexes cannot be held across
.Fn epoch_wait_preempt .
The
.Fa flags
specified are formed by
.Em OR Ns 'ing
the following values:
.Bl -tag -offset indent -width Ds
.It Dv EPOCH_LOCKED
Permit holding mutexes across
.Fn epoch_wait_preempt
(requires
.Dv EPOCH_PREEMPT ) .
When doing this one must be cautious of creating a situation where a deadlock
is possible.
.It Dv EPOCH_PREEMPT
The
.Vt epoch
will allow preemption during sections.
Only non-sleepable locks may be acquired during a preemptible epoch.
The functions
.Fn epoch_enter_preempt ,
.Fn epoch_exit_preempt ,
and
.Fn epoch_wait_preempt
must be used in place of
.Fn epoch_enter ,
.Fn epoch_exit ,
and
.Fn epoch_wait ,
respectively.
.El
.Pp
.Vt epoch Ns s
are freed with
.Fn epoch_free .
The flags passed to epoch_alloc determine whether preemption is
allowed during a section or not (the default), as specified by
EPOCH_PREEMPT.
.Pp
Threads indicate the start of an epoch critical section by calling
.Fn epoch_enter .
The end of a critical section is indicated by calling
.Fn epoch_exit .
The _preempt variants can be used around code which requires preemption.
A thread can wait until a grace period has elapsed
since any threads have entered
the epoch by calling
.Fn epoch_wait
or
.Fn epoch_wait_preempt ,
depending on the epoch_type.
The use of a default epoch type allows one to use
.Fn epoch_wait
which is guaranteed to have much shorter completion times since
we know that none of the threads in an epoch section will be preempted
before completing its section.
If the thread can't sleep or is otherwise in a performance sensitive
path it can ensure that a grace period has elapsed by calling
.Fn epoch_call
with a callback with any work that needs to wait for an epoch to elapse.
Only non-sleepable locks can be acquired during a section protected by
.Fn epoch_enter
(or
.Fn epoch_enter_preempt
for preemptible epochs).
Threads call
.Fn epoch_exit
(or
.Fn epoch_exit_preempt
for preemptible epochs)
to indicate the end of a critical section.
.Vt struct epoch_tracker Ns s
are stack objects whose pointers are passed to
.Fn epoch_enter_preempt
and
.Fn epoch_exit_preempt .
.Fn epoch_exit_preempt
(much like
.Vt struct rm_priotracker ) .
.Pp
Threads can defer work until a grace period has expired since any thread has
entered the epoch either synchronously or asynchronously.
.Fn epoch_call
defers work asynchronously by invoking the provided
.Fa callback
at a later time.
.Fn epoch_wait
(or
.Fn epoch_wait_preempt )
blocks the current thread until the grace period has expired and the work can be
done safely.
.Pp
Default, non-preemptible epoch wait
.Fn ( epoch_wait )
is guaranteed to have much shorter completion times relative to
preemptible epoch wait
.Fn ( epoch_wait_preempt ) .
(In the default type, none of the threads in an epoch section will be preempted
before completing its section.)
.Pp
INVARIANTS can assert that a thread is in an epoch by using
.Fn in_epoch .
.Fn in_epoch "epoch"
is equivalent to invoking
.Fn in_epoch_verbose "epoch" "0" .
If
.Cd EPOCH_TRACE
is enabled,
.Fn in_epoch_verbose "epoch" "1"
provides additional verbose debugging information.
.Pp
The epoch API currently does not support sleeping in epoch_preempt sections.
A caller should never call
.Fn epoch_wait
in the middle of an epoch section for the same epoch as this will lead to a deadlock.
.Pp
By default mutexes cannot be held across
.Fn epoch_wait_preempt .
To permit this the epoch must be allocated with
EPOCH_LOCKED.
When doing this one must be cautious of creating a situation where a deadlock is
possible. Note that epochs are not a straight replacement for read locks.
Callers must use safe list and tailq traversal routines in an epoch (see ck_queue).
When modifying a list referenced from an epoch section safe removal
routines must be used and the caller can no longer modify a list entry
in place.
An item to be modified must be handled with copy on write
and frees must be deferred until after a grace period has elapsed.
.Pp
The
.Fn epoch_drain_callbacks
function is used to drain all pending callbacks which have been invoked by prior
@ -140,12 +211,18 @@ This function can sleep and is not optimized for performance.
will return 1 if curthread is in curepoch, 0 otherwise.
.Sh CAVEATS
One must be cautious when using
.Fn epoch_wait_preempt
threads are pinned during epoch sections so if a thread in a section is then
preempted by a higher priority compute bound thread on that CPU it can be
prevented from leaving the section.
Thus the wait time for the waiter is
potentially unbounded.
.Fn epoch_wait_preempt .
Threads are pinned during epoch sections, so if a thread in a section is then
preempted by a higher priority compute bound thread on that CPU, it can be
prevented from leaving the section indefinitely.
.Pp
Epochs are not a straight replacement for read locks.
Callers must use safe list and tailq traversal routines in an epoch (see ck_queue).
When modifying a list referenced from an epoch section safe removal
routines must be used and the caller can no longer modify a list entry
in place.
An item to be modified must be handled with copy on write
and frees must be deferred until after a grace period has elapsed.
.Sh EXAMPLES
Async free example:
Thread 1:
@ -154,8 +231,8 @@ int
in_pcbladdr(struct inpcb *inp, struct in_addr *faddr, struct in_laddr *laddr,
struct ucred *cred)
{
/* ... */
epoch_enter(net_epoch);
/* ... */
epoch_enter(net_epoch);
CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
sa = ifa->ifa_addr;
if (sa->sa_family != AF_INET)
@ -167,7 +244,7 @@ in_pcbladdr(struct inpcb *inp, struct in_addr *faddr, struct in_laddr *laddr,
}
}
epoch_exit(net_epoch);
/* ... */
/* ... */
}
.Ed
Thread 2:
@ -198,8 +275,6 @@ and then defers deletion.
More general mutation or a synchronous
free would have to follow a call to
.Fn epoch_wait .
.Sh ERRORS
None.
.Sh NOTES
The
.Nm