Version 10 (modified by 10 years ago) (diff) | ,
---|
The user_lock library
The user_lock.c and user_lock.h files define two types of services:
- user-level atomic access functions, to increment a shared counter.
- user-level spin-locks, to obtain exclusive access to a shared resource in a parallel multi-tasks application.
The proposed spin-lock implement a waiting queue service, to enforce fairness and avoid live-lock situations. Each lock (giet_lock_t object) occupies a complete 64 bytes cache line to avoid false sharing.
The user_lock_t being shared by several tasks, should be defined as a global variable in the application code. It is possible to control precisely the placement of a specific lock on a specific physical memory bank by defining a specific vseg in the application mapping.
Distributed version are available in user_sqt_lock.c and user_sqt_lock.h files.
Access functions
unsigned int atomic_increment( unsigned int* shared, unsigned int increment)
This function uses a LL/SC to atomically increment a shared variable.
- ptr : pointer on the shared variable
- increment : increment value.
void lock_init( user_lock_t * lock )
This function should be called by one single task.
- lock : pointer on the user_lock_t structure.
void lock_acquire( user_lock_t * lock )
This blocking function returns only when the lock has been successfully taken.
- lock : pointer on the user_lock_t structure.
void lock_release( user_lock_t * lock )
This function releases the lock. It must be called by a task after a successful lock_acquire().
- lock : pointer on the user_lock_t structure.
Distributed functions
The sqt_lock can be used in multi-clusters architectures, and is implemented as a - physically distributed - Synchronisation-Quad-Tree (SQT). The SQT topology is completely defined by the (x_size / y_size) parameters, with the following constraints:
- The involved clusters form a mesh(x_size * y_size).
- The lower left involved cluster is cluster(0,0).
- All involved clusters must contain a heap_x_y vseg declared in the mapping.
- The number of involved tasks in a given cluster is the same for all involved clusters.
void sqt_lock_init( sqt_lock_t * lock )
This function should be called by one single task.
- lock : pointer on the sqt_lock_t structure.
void sqt_lock_acquire( sqt_lock_t * lock )
This blocking function returns only when all locks of the distributed tree have been successfully taken.
- lock : pointer on the sqt_lock_t structure.
void sqt_lock_release( sqt_lock_t * lock )
This function releases the lock. It must be called by a task after a successful lock_acquire().
- lock : pointer on the sqt_lock_t structure.