rt_malloc.txt

rtai-3.1-test3的源代码(Real-Time Application Interface )

RTAI Dynamic Memory Management Module.
======================================

The dynamic memory module for RTAI gives real time application programs
the ability to be able to dynamically create and free memory using the
standard UNIX programming API calls.  Before this module real time
applications which needed dynamic memory management had to use the
standard Linux kernel calls: kmalloc and kfree.  This is potentially very
dangerous as these calls can block, and if this were to occur from a real
time task the result is usually a total system lock up.  The situation is
made worse as this can lead to intermittent bugs as real time applications
can appear to work using these calls, but fail under varying load
conditions and circumstances.

The dynamic memory manager module pre-allocates blocks (chunks) of memory
from the Linux kernel which is available for use by real time tasks using
a "UNIX like" API: rt_malloc and rt_free.  The manager allocates and frees
memory from these blocks of memory, and also monitors the amount of free
memory that is available in these blocks.  When the amount of available
memory falls below a low water mark a request for another block of memory
is made pending.  Similarly when the amount of available memory is greater
than a high water mark a request to free a block of memory is made
pending.  These pending requests are carried out using either vmalloc and
vfree or kmalloc and kfree calls at a safe time, ie when the real time system
becomes idle, just before control is handed back to Linux.  Using this
mechanism the memory manager balances the memory requirements of the real
time application with the need to keep as much memory as possible
available to the Linux kernel.

Some of the reasons for using vmalloc are:
  - Simpler to share allocated buffers with user space.
  - Doesn't have the size restriction of kmalloc. (128KB)

and for using kmalloc:
  - Faster.
  - Contiguous buffer address, needed by DMA controllers which don't have
    scatter/gather capability.

The dynamic memory manager can be configured for the number of memory
blocks that are kept available, and the size of the blocks.  This means
that the module can be configured to meet the specific requirements and
operating conditions of a real time application, allowing the application
to be programmed using the flexibility of dynamic memory allocation,
whilst minimizing the memory resource burden on the Linux kernel.  Another
key feature provided by the module is the ability to create real time
threads from other real time threads which is an essential feature for
many applications. For this purpose, RT_TASK *rt_alloc_dynamic_task(void)
has been added to the schedulers.

Dynamic memory allocation for real time tasks is supported by the
implementation of the following functions:

1/ void *rt_malloc(unsigned int size);

rt_malloc allocates size bytes and returns a pointer to the allocated
memory.  If the allocation request fails a NULL is returned.


2/ void rt_free(void *ptr);

rt_free frees the memory space pointed by ptr which must have been
returned by a previous call to rt_malloc.  rt_free returns no value.

The default configuration of the dynamic memory manager is:

Memory block size: 64 KBytes
Number of free blocks kept available: 2

These parameters can be changed if required by using the following module
parameters:

Memory block size:  granularity
Number of free blocks kept available: low_chk_ref

For example to change the size of the memory blocks to 32 Kbytes and the
number of free blocks kept available for allocation to 4:

insmod rtai_sched_up.o granularity=32768 low_chk_ref=4



RTAI C++ Support Built Into The Module
======================================

Real time C++ support is provided with the implementation of the operators:

	void* operator new(size_t);
	void* operator new [](size_t);
	void  operator delete(void*);
	void  operator delete [](void*);

These operators use the rt_malloc() and rt_free() primitives and thus
make it possible to execute real time C++ written modules. The LXRT directory
provides an example on how to do this. Notice that C++ support is limited
as programs must be compiled with the -fno-exception g++ option. Also, 
an abstract base class that defines pure virtual functions should implement
an empty function otherwise the compiler will generate the infamous 
__pure_virtual() call which will result in an unresolved symbol at insmod
time: 

class Wathever {
public:
	virtual void foo() = 0; // This generates a __pure_virtual() call.
	virtual void Better {}  // The empty function will not.
};

Alternatively you could choose to implement your own __pure_virtual()
function with something like this:

extern "C" void __pure_virtual();
void __pure_virtual()
{
	RT_TASK *t;
	rt_printk( "%X calling a pure virtual\n", t = rt_whoami());
	rt_task_suspend(t);
}

Notice that if you attemp to compile with the -fPIC option you will
see another infamous unresolved symbol: _GLOBAL_OFFSET_TABLE_.

Also, you cannot instantiate global objects because nothing actually
does the global initialisation. This is normally done before the program
enters main and you need to link with the library files crtbegin.o and
crtend.o to do that. However, linking with those two files will introduce
you to two other infamous unresolved symbols: __register_frame_info and
__deregister_frame_info.

To resume, C++ support is limited in that exception handling and global
instantiation services are not available in the kernel due to lack of 
library support. It is possible to trick the compiler in order to avoid 
the __pure_virtual unresolved symbol.