summaryrefslogtreecommitdiffstats
path: root/ipc
diff options
context:
space:
mode:
Diffstat (limited to 'ipc')
-rw-r--r--ipc/sem.c103
1 files changed, 53 insertions, 50 deletions
diff --git a/ipc/sem.c b/ipc/sem.c
index a744eb5..5b33228 100644
--- a/ipc/sem.c
+++ b/ipc/sem.c
@@ -3,56 +3,6 @@
* Copyright (C) 1992 Krishna Balasubramanian
* Copyright (C) 1995 Eric Schenk, Bruno Haible
*
- * IMPLEMENTATION NOTES ON CODE REWRITE (Eric Schenk, January 1995):
- * This code underwent a massive rewrite in order to solve some problems
- * with the original code. In particular the original code failed to
- * wake up processes that were waiting for semval to go to 0 if the
- * value went to 0 and was then incremented rapidly enough. In solving
- * this problem I have also modified the implementation so that it
- * processes pending operations in a FIFO manner, thus give a guarantee
- * that processes waiting for a lock on the semaphore won't starve
- * unless another locking process fails to unlock.
- * In addition the following two changes in behavior have been introduced:
- * - The original implementation of semop returned the value
- * last semaphore element examined on success. This does not
- * match the manual page specifications, and effectively
- * allows the user to read the semaphore even if they do not
- * have read permissions. The implementation now returns 0
- * on success as stated in the manual page.
- * - There is some confusion over whether the set of undo adjustments
- * to be performed at exit should be done in an atomic manner.
- * That is, if we are attempting to decrement the semval should we queue
- * up and wait until we can do so legally?
- * The original implementation attempted to do this.
- * The current implementation does not do so. This is because I don't
- * think it is the right thing (TM) to do, and because I couldn't
- * see a clean way to get the old behavior with the new design.
- * The POSIX standard and SVID should be consulted to determine
- * what behavior is mandated.
- *
- * Further notes on refinement (Christoph Rohland, December 1998):
- * - The POSIX standard says, that the undo adjustments simply should
- * redo. So the current implementation is o.K.
- * - The previous code had two flaws:
- * 1) It actively gave the semaphore to the next waiting process
- * sleeping on the semaphore. Since this process did not have the
- * cpu this led to many unnecessary context switches and bad
- * performance. Now we only check which process should be able to
- * get the semaphore and if this process wants to reduce some
- * semaphore value we simply wake it up without doing the
- * operation. So it has to try to get it later. Thus e.g. the
- * running process may reacquire the semaphore during the current
- * time slice. If it only waits for zero or increases the semaphore,
- * we do the operation in advance and wake it up.
- * 2) It did not wake up all zero waiting processes. We try to do
- * better but only get the semops right which only wait for zero or
- * increase. If there are decrement operations in the operations
- * array we do the same as before.
- *
- * With the incarnation of O(1) scheduler, it becomes unnecessary to perform
- * check/retry algorithm for waking up blocked processes as the new scheduler
- * is better at handling thread switch than the old one.
- *
* /proc/sysvipc/sem support (c) 1999 Dragos Acostachioaie <dragos@iname.com>
*
* SMP-threaded, sysctl's added
@@ -61,6 +11,8 @@
* (c) 2001 Red Hat Inc
* Lockless wakeup
* (c) 2003 Manfred Spraul <manfred@colorfullife.com>
+ * Further wakeup optimizations, documentation
+ * (c) 2010 Manfred Spraul <manfred@colorfullife.com>
*
* support for audit of ipc object properties and permission changes
* Dustin Kirkland <dustin.kirkland@us.ibm.com>
@@ -68,6 +20,57 @@
* namespaces support
* OpenVZ, SWsoft Inc.
* Pavel Emelianov <xemul@openvz.org>
+ *
+ * Implementation notes: (May 2010)
+ * This file implements System V semaphores.
+ *
+ * User space visible behavior:
+ * - FIFO ordering for semop() operations (just FIFO, not starvation
+ * protection)
+ * - multiple semaphore operations that alter the same semaphore in
+ * one semop() are handled.
+ * - sem_ctime (time of last semctl()) is updated in the IPC_SET, SETVAL and
+ * SETALL calls.
+ * - two Linux specific semctl() commands: SEM_STAT, SEM_INFO.
+ * - undo adjustments at process exit are limited to 0..SEMVMX.
+ * - namespace are supported.
+ * - SEMMSL, SEMMNS, SEMOPM and SEMMNI can be configured at runtine by writing
+ * to /proc/sys/kernel/sem.
+ * - statistics about the usage are reported in /proc/sysvipc/sem.
+ *
+ * Internals:
+ * - scalability:
+ * - all global variables are read-mostly.
+ * - semop() calls and semctl(RMID) are synchronized by RCU.
+ * - most operations do write operations (actually: spin_lock calls) to
+ * the per-semaphore array structure.
+ * Thus: Perfect SMP scaling between independent semaphore arrays.
+ * If multiple semaphores in one array are used, then cache line
+ * trashing on the semaphore array spinlock will limit the scaling.
+ * - semncnt and semzcnt are calculated on demand in count_semncnt() and
+ * count_semzcnt()
+ * - the task that performs a successful semop() scans the list of all
+ * sleeping tasks and completes any pending operations that can be fulfilled.
+ * Semaphores are actively given to waiting tasks (necessary for FIFO).
+ * (see update_queue())
+ * - To improve the scalability, the actual wake-up calls are performed after
+ * dropping all locks. (see wake_up_sem_queue_prepare(),
+ * wake_up_sem_queue_do())
+ * - All work is done by the waker, the woken up task does not have to do
+ * anything - not even acquiring a lock or dropping a refcount.
+ * - A woken up task may not even touch the semaphore array anymore, it may
+ * have been destroyed already by a semctl(RMID).
+ * - The synchronizations between wake-ups due to a timeout/signal and a
+ * wake-up due to a completed semaphore operation is achieved by using an
+ * intermediate state (IN_WAKEUP).
+ * - UNDO values are stored in an array (one per process and per
+ * semaphore array, lazily allocated). For backwards compatibility, multiple
+ * modes for the UNDO variables are supported (per process, per thread)
+ * (see copy_semundo, CLONE_SYSVSEM)
+ * - There are two lists of the pending operations: a per-array list
+ * and per-semaphore list (stored in the array). This allows to achieve FIFO
+ * ordering without always scanning all pending operations.
+ * The worst-case behavior is nevertheless O(N^2) for N wakeups.
*/
#include <linux/slab.h>
OpenPOWER on IntegriCloud