doc/locking.txt
56c2dabd
 # $Id$
 #
 # History:
 #---------
 #  2003-03-11  created by andrei
d3da8467
 #  2006-04-04  minor archs updates, added lock_try(..)  (andrei)
56c2dabd
 
 
 SER locking interface
 
 
 1. Why use it?
    ----------
 
  The main reason in creating it was to have a single transparent interface to various locking methods. For example right now ser uses the following locking methods, depending on their availability on the target system:
d3da8467
  FAST_LOCK - fast inline assembly locks, defined in fast_lock.h. They are currently available for x86, x86_64, sparc, sparc64, arm , armv6 (no smp mode supported yet), ppc, ppc64, mips, mips64 and alpha . In general if the assembly code exists for a given arhitecture and the compiler knows inline assembly (for example sun cc does not) FAST_LOCK is prefered. The main advantage of using FAST_LOCK is very low memory overhead and extremely fast lock/unlock operations (like 20 times faster then SYSV semaphores on linux & 40 times on solaris). The only thing that comes close to them are pthread mutexes (which are about 3-4 times slower).
56c2dabd
  PTHREAD_MUTEX - uses pthread_mutex_lock/unlock. They are quite fast but they work between processes only on some systems (they do not work on linux).
d3da8467
  POSIX_SEM  - uses posix semaphores (sem_wait/sem_post). They are slower then the previous methods but still way faster then SYSV sempahores. Unfortunately they also do not work on all the systems (e.g. linux).
56c2dabd
  SYSV_SEM - this is the most portable but also the slowest locking method. Another problem is that the number of semaphores that can be alocated by a process is limited. One also has to free them before exiting.
 
 
 2. How to use it?
 -----------------
 
 First of all you have to include locking.h. Then when compiling the code one or all of FAST_LOCK, USE_PTHREAD_MUTEX, USE_PTHREAD_SEM or USE_SYSV_SEM must be defined (the ser Makefile.defs takes care of this, you should need to change it only for new arhitectures or compilers).
 
 locking.h defines 2 new types: gen_lock_t and lock_set_t.
 
 
 3. Simple locks
 ---------------
 
 The simple locks are simple mutexes. The type is gen_lock_t.
 WARNING: do not make any assumptions on gen_lock_t base type, it does not have to be always an int.
 
 Allocation & initialization:
 ----------------------------
 
 The locks are allocated with:
    gen_lock_t* lock_alloc();
 and initialized with:
    gen_lock_t* lock_init(gen_lock_t* lock);
 
 Both function return 0 on failure.
 The locks must be initialized before use.
 
 A proper alloc/init sequence looks like:
 
 gen_lock_t* lock;
 
 lock=lock_alloc();
 if (lock==0) goto error;
 if (lock_init(lock)==0){
    lock_dealloc(lock);
    goto error; /* could not init lock*/
 }
 ...
 
 Lock allocation can be skipped in some cases: if the lock is already in shared memory you don't need to allocate it again, you can initialize it directly, but keep in mind that the lock MUST be in shared memory.
 
 Example:
 
 struct s {
     int foo;
     gen_lock_t lock;
 } bar;
 
 bar=shm_malloc(sizeof struct s); /* we allocate it in the shared memory */
 if (lock_init(&bar->lock)==0){
  /* error initializing the lock */
  ...
 }
 
 
 Destroying & deallocating the locks:
 ------------------------------------
 
   void    lock_destroy(gen_lock_t* lock);
   void    lock_dealloc(gen_lock_t* lock);
  
  The lock_destroy function must be called first. It removes the resources associated with the lock, but it does not also free the lock shared memory part. Think of sysv rmid.
  Please don't forget to call this function, or you can leave allocated resources in some cases (e.g sysv semaphores). Be carefull to call it in your module destroy function if you use any global module locks.
 
 Example:
 lock_destroy(lock);
 lock_dealloc(lock);
 
 
 Of course you don't need to call lock_dealloc if your lock was not allocated with lock_alloc.
 
 
 Locking & unlocking:
 --------------------
 
 void    lock_get(gen_lock_t* lock);      - lock (mutex down)
 void    lock_release(gen_lock_t* lock);  - unlock (mutex up)
d3da8467
 int     lock_try(gen_lock_t* lock);      - tries to lock and returns 0 
                                            if succesfull, -1 if not (this is
                                            a non-blocking lock_get())
56c2dabd
 
 
 
 4. Lock Sets
 ------------
 
 The lock sets are kind of sysv semaphore sets equivalent. The type is lock_set_t.
 Use them when you need a lot of mutexes. In some cases they waste less system resources than arrays of gen_lock_t (e.g. sys v semaphores).
 
 Allocating & initializing:
 --------------------------
 
 lock_set_t* lock_set_alloc(int no);
 lock_set_t* lock_set_init(lock_set_t* set);
 
 Both functions return 0 on failure.
 
 WARNING: expect the allocation function to fail for large numbers. It depends on the locking method used & the system available resources (again the sysv semaphores example).
 
 Example:
 
 lock_set_t *lock_set;
 
 lock_set=lock_set_alloc(100);
 if (lock_set==0) goto error;
 if (lock_set_init(lock_set)==0){
    lock_set_dealloc(lock_set);
    goto error;
 }
 
7f5a5c30
 or
  if ((lock_set=lock_set_alloc(100))==0) || (lock_set_init(lock_set)==0)){
    if (lock_set) lock_set_dealloc(lock_set);
    goto error;
  }
 
56c2dabd
 
 Destroying & deallocating:
 --------------------------
   void lock_set_destroy(lock_set_t* s);
   void lock_set_dealloc(lock_set_t* s);
 
 Again don't forget to "destroy" the locks.
 
 
 Locking & unlocking:
 --------------------
 
 void lock_set_get(lock_set_t* s, int i);
 void lock_set_release(lock_set_t* s, int i);
d3da8467
 int  lock_set_try(lock_set_t* s, int i);      - tries to lock the i-th lock
                                                 from the set. If succesfull
                                                 returns 0, if not -1.
56c2dabd
 
 Example:
 
 lock_set_get(lock_set, 2);
 /* do something */
 lock_set_release(lock_set, 2);
 
 
7f5a5c30
 When to use lock_set_t & when to use gen_lock_t
 -----------------------------------------------
 If you use lots of semaphores and GEN_LOCK_T_PREFERED is undefined then use
  lock_set_t. If GEN_LOCK_T_PREFERED is defined you can safely use gen_lock_t 
  arrays instead.