summaryrefslogtreecommitdiffstats
path: root/share/man/man9
diff options
context:
space:
mode:
authordillon <dillon@FreeBSD.org>1998-12-23 00:24:59 +0000
committerdillon <dillon@FreeBSD.org>1998-12-23 00:24:59 +0000
commit9f074776ec98ccdd3b640c8c715fe6faa96531f3 (patch)
tree8dfe41bef0f21b4f2929f028567b2aee2b930ce8 /share/man/man9
parentfb5b48c1cb415d1abe438657ad64e5ed6327ec9d (diff)
downloadFreeBSD-src-9f074776ec98ccdd3b640c8c715fe6faa96531f3.zip
FreeBSD-src-9f074776ec98ccdd3b640c8c715fe6faa96531f3.tar.gz
update asleep/await sections
Diffstat (limited to 'share/man/man9')
-rw-r--r--share/man/man9/sleep.977
1 files changed, 46 insertions, 31 deletions
diff --git a/share/man/man9/sleep.9 b/share/man/man9/sleep.9
index 0e352e4..2d9410b 100644
--- a/share/man/man9/sleep.9
+++ b/share/man/man9/sleep.9
@@ -23,7 +23,7 @@
.\" (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
.\" THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.\"
-.\" $Id: sleep.9,v 1.11 1998/12/21 10:29:28 dillon Exp $
+.\" $Id: sleep.9,v 1.12 1998/12/21 10:34:53 dillon Exp $
.\" "
.Dd December 17, 1998
.Os
@@ -133,15 +133,18 @@ value of NULL
to remove any previously queued condition.
.Pp
.Nm Await
-implements the new asynchronous wait function. If you
+implements the new asynchronous wait function. When
.Fn asleep
-on an identifier,
+is called on an identifier it associates the process with that
+identifier but does not block.
.Fn await
-will actually block the process until someone calls
+will actually block the process until
.Fn wakeup
-on that identifier. If someone calls
+is called on that identifier any time after the
+.Fn asleep .
+If
.Fn wakeup
-after you
+is called after you
.Fn asleep
but before you
.Fn await
@@ -151,10 +154,10 @@ call is effectively a NOP.
If
.Fn await
is called multiple times without an intervening
-.Fn asleep
+.Fn asleep ,
the
.Fn await
-is effective a NOP, but will call
+is effectively a NOP but will also call
.Fn mswitch
for safety. The
.Fn await
@@ -162,13 +165,13 @@ function allows you to override the priority and timeout values to be used.
If the value -1 is specified for an argument, the value is taken from the
previous
.Fn asleep
-call. If you pass -1 for the priority you must be prepared to catch signal
+call. If -1 is passed for the priority you must be prepared to catch signal
conditions if the prior call to
.Fn asleep
-specified it in its priority. If you pass -1 for the timeout you must be
+specified it in its priority. If -1 is passed for the timeout you must be
prepared to catch a timeout condition if the prior call to
.Fn asleep
-specified a timeout. When you use -1, you should generally not make
+specified a timeout. When you use -1, it is usually a good idea to not make
assumptions as to the arguments used by the prior
.Fn asleep
call.
@@ -177,29 +180,39 @@ The
.Fn asleep
and
.Fn await
-functions are used by the kernel code for various purposes but the main one is
-to allow complex interlocking code to 'backout' of a temporary resource failure
-(such as lack of memory or trying to access a block that is not in the buffer
-cache) in order to release major locks prior to blocking, and to then retry
-the call that failed on wakeup. This involves subroutines deep in the kernel
-calling
+functions are mainly used by the kernel to shift the burden of blocking
+away from extremely low level routines and to push it onto their callers.
+This in turn allows more complex interlocking code to
+.Em backout
+of a temporary resource failure
+(such as lack of memory) in order to release major locks prior to actually
+blocking, and to then retry the operation on wakeup. This key feature is
+expected to be heavily used in SMP situations in order to allow code to make
+better use of spinlocks. A spinlock, by its very nature, cannot be used
+around code that might block. It is hoped that these capabilities will
+make it easier to migrate the SMP master locks deeper into the kernel.
+.Pp
+These routines may also be used to avoid nasty spl*() calls to get around
+race conditions with simple conditional test/wait interlocks. You simple
+call
+.Fn asleep
+prior to your test, then conditonally
+.Fn await
+only if the test fails. It is usually a good idea to cancel an
.Fn asleep
-and returning a temporary failure, then popping back up through a number
-of call levels before calling
+if you wind up never calling the related
.Fn await ,
-then retrying. The kernel might also use these functions to avoid using
-spinlocks in a check-condition interlock. That is, in case the case where
-the kernel wishes to check the condition of something and then block on it.
-To avoid the race between the check and the blocking, the kernel can first
-check the condition, then call
-.Fn asleep ,
-then check the condition a second time before calling
-.Fn await .
-The overlap makes the race condition impossible.
+but it is not required. If you do not want to waste cpu calling
+.Fn asleep
+unnecessarily, you can surround the whole thing with a second test. The
+race condition is still handled by the inside
+.Fn asleep
+call.
.Sh RETURN VALUES
See above.
.Sh SEE ALSO
-.Xr ps 1
+.Xr ps 1 ,
+.Xr malloc 9
.Sh HISTORY
The sleep/wakeup process synchronization mechanism is very old. It
appeared in a very early version of Unix.
@@ -209,8 +222,10 @@ appeared in
.Bx 4.4 .
.Pp
.Nm Asleep/await
-first appeared in FreeBSD-3.0.1
+first appeared in FreeBSD-3.0.1 and is designed to shift the burden of blocking
+away from extremely low level routines and push it up to their callers.
.Sh AUTHORS
This man page has been written by
.ie t J\(:org Wunsch.
-.el Joerg Wunsch. asleep/await portions were written by Matt Dillon
+.el Joerg Wunsch.
+asleep/await designed and written by Matthew Dillon.
OpenPOWER on IntegriCloud