Add a bunch of new text, in particular:

Use of KM_NOSLEEP is strongly discouraged as it can
	create transient, hard to debug failures that occur
	when the system is under memory pressure.

	In situations where it is not possible to sleep, for
	example because locks are held by the caller, the
	code path should be restructured to allow the allo-
	cation to be made in another place.
This commit is contained in:
ad 2008-12-29 12:43:47 +00:00
parent a3caa093d1
commit a6dfe8170e
1 changed files with 64 additions and 7 deletions

View File

@ -1,4 +1,4 @@
.\" $NetBSD: kmem_alloc.9,v 1.6 2008/01/03 15:59:57 yamt Exp $
.\" $NetBSD: kmem_alloc.9,v 1.7 2008/12/29 12:43:47 ad Exp $
.\"
.\" Copyright (c)2006 YAMAMOTO Takashi,
.\" All rights reserved.
@ -25,7 +25,7 @@
.\" SUCH DAMAGE.
.\"
.\" ------------------------------------------------------------
.Dd January 4, 2008
.Dd December 29, 2008
.Dt KMEM_ALLOC 9
.Os
.\" ------------------------------------------------------------
@ -49,20 +49,73 @@ It takes the following arguments.
Specify the size of allocation in bytes.
.It Fa kmflags
Either of the following:
.Bl -tag -width KM_NOSLEEP -compact
.Bl -tag -width KM_NOSLEEP
.It KM_SLEEP
Can sleep until enough memory is available.
If the allocation cannot be satisfied immediatley, sleep until enough
memory is available.
.It KM_NOSLEEP
Don't sleep.
Immediately return
.Dv NULL
Immediately return NULL
if there is not enough memory available.
It should only be used when failure to allocate will not have harmful,
user-visible effects.
.Pp
.Bf -symbolic
Use of
.Dv KM_NOSLEEP
is strongly discouraged as it can create transient, hard to debug failures
that occur when the system is under memory pressure.
.Ef
.Pp
In situations where it is not possible to sleep, for example because locks
are held by the caller, the code path should be restructured to allow the
allocation to be made in another place.
.El
.El
.Pp
The contents of allocated memory are uninitialized.
.Pp
Unlike Solaris, kmem_alloc(0, flags) is illegal.
.Pp
Making
.Dv KM_SLEEP
allocations while holding mutexes or reader/writer locks is discouraged, as the
caller can sleep for an unbounded amount of time in order to satisfy the
allocation.
This can in turn block other threads that wish to acquire locks held by the
caller.
.Pp
For some locks this is permissible or even unavoidable.
For others, particularly locks that may be taken from soft interrupt context,
it is a serious problem.
As a general rule it is better not to allow this type of situation to develop.
One way to circumvent the problem is to make allocations speculative and part
of a retryable sequence.
For example:
.Bd -literal
retry:
/* speculative unlocked check */
if (need to allocate) {
new_item = kmem_alloc(sizeof(*new_item), KM_SLEEP);
} else {
new_item = NULL;
}
mutex_enter(lock);
/* check while holding lock for true status */
if (need to allocate) {
if (new_item == NULL) {
mutex_exit(lock);
goto retry;
}
consume(new_item);
new_item = NULL;
}
mutex_exit(lock);
if (new_item != NULL) {
/* did not use it after all */
kmem_free(new_item, sizeof(*new_item));
}
.Ed
.\" ------------------------------------------------------------
.Sh RETURN VALUES
On success,
@ -80,7 +133,11 @@ Otherwise, it returns
.\" ------------------------------------------------------------
.Sh CAVEATS
.Fn kmem_alloc
can not be used from interrupt context.
cannot be used from interrupt context, from a soft interrupt, or from
a callout.
Use
.Xr pool_cache 9
in these situations.
.\" ------------------------------------------------------------
.Sh SECURITY CONSIDERATION
As the allocated memory is uninitialized, it can contain security-sensitive