Subversion Repositories openrisc_me

                Memory allocation package - Implementation Notes
                ------------------------------------------------



Made with loving care by Jonathan Larmour (jlarmour@redhat.com)
Initial version: 2000-07-03
Last updated:    2000-07-03



Meta
----

This document describes some interesting bits and pieces about the memory
allocation package - CYGPKG_MEMALLOC. It is intended as a guide to
developers, not users. This isn't (yet) in formal documentation format,
and probably should be.


Philosophy
----------

The object of this package is to provide everything required for dynamic
memory allocation, some sample implementations, the ability to plug in
more implementations, and a standard malloc() style interface to those
allocators.

The classic Unix-style view of a heap is using brk()/sbrk() to extend the
data segment of the application. However this is inappropriate for an
embedded system because:

- you may not have an MMU, which means memory may be disjoint, thus breaking
  this paradigm

- in a single process system there is no need to play tricks since there
  is only the one address space and therefore heap area to use.

Therefore instead, we base the heap on the idea of fixed size memory pools.
The size of each pool is known in advance.


Overview
--------

Most of the infrastructure this package provides is geared towards
supporting the ISO standard malloc() family of functions. A "standard"
eCos allocator should be able to plug in to this infrastructure and
transparently work. The interface is based on simple use of C++ - nothing
too advanced.

The allocator to use is dictated by the
CYGBLD_MEMALLOC_MALLOC_IMPLEMENTATION_HEADER option. Choosing the
allocator can be done by ensuring the CDL for the new allocator
has a "requires" that sets the location of the header to use when that
allocator is enabled. New allocators should default to disabled, so they
don't have to worry about which one is the default, thus causing CDL
conflicts. When enabled the new allocator should also claim to implement
CYGINT_MEMALLOC_MALLOC_ALLOCATORS.

The implementation header file that is set must have a special property
though - it may be included with __MALLOC_IMPL_WANTED defined. If this
is the case, then this means the infrastructure wants to find out the
name of the class that is implemented in this header file. This is done
by setting CYGCLS_MEMALLOC_MALLOC_IMPL. If __MALLOC_IMPL_WANTED is defined
then no non-preprocessor output should be generated, as this will be included
in a TCL script in due course. An existing example from this package would
be:

#define CYGCLS_MEMALLOC_MALLOC_IMPL Cyg_Mempool_dlmalloc

// if the implementation is all that's required, don't output anything else
#ifndef __MALLOC_IMPL_WANTED

class Cyg_Mempool_dlmalloc
{
[etc.]

To meet the expectations of malloc, the class should have the following
public interfaces (for details it is best to look at some of the
examples in this package):

- a constructor taking arguments of the form:

  ALLOCATORNAME( cyg_uint8 *base, cyg_int32 size );

  If you want to be able to support other arguments for when accessing
  the allocator directly you can add them, but give them default values,
  or use overloading

- a destructor

- a try_alloc() function that returns new memory, or NULL on failure:

    cyg_uint8 *
    try_alloc( cyg_int32 size );

- a free() function taking one pointer argument that returns a boolean
  for success or failure:

    cyg_bool
    free( cyg_uint8 *ptr );

  Again, extra arguments can be added, as long as they are defaulted.


- resize_alloc() which is designed purely to support realloc(). It
  has the prototype: 
    cyg_uint8 *
    resize_alloc( cyg_uint8 *alloc_ptr, cyg_int32 newsize,
                  cyg_int32 *oldsize );

  The idea is that if alloc_ptr can be adjusted to newsize, then it will
  be. If oldsize is non-NULL the old size (possibly rounded) is placed
  there. However what this *doesn't* do (unlike the real realloc()) is
  fall back to doing a new malloc(). All it does is try to do tricks
  inside the allocator. It's up to higher layers to call malloc().

- get_status() allows the retrieval of info from the allocator. The idea
  is to pass in the bitmask OR of the flags defined in common.hxx, which
  selects what information is requested. If the request is supported by
  the allocator, the approriate structure fields are filled in; otherwise
  unsupported fields will be left with the value -1. (The constructor for
  Cyg_Mempool_Status initializes them to -1). If you want to reinitialize
  the structure and deliberately lose the data in a Cyg_Mempool_Status
  object, you need to invoke the init() method of the status object to
  reinitialize it.

    void
    get_status( cyg_mempool_status_flag_t flags, Cyg_Mempool_Status &status );

  A subset of the available stats are exported via mallinfo()


Cyg_Mempolt2 template
---------------------

If using the eCos kernel with multiple threads accessing the allocators,
then obviously you need to be sure that the allocator is accessed in a
thread-safe way. The malloc() wrappers do not make any assumptions
about this. One helpful approach currently used by all the allocators
in this package is to (optionally) use a template (Cyg_Mempolt2) that
provides extra functions like a blocking alloc() that waits for memory
to be freed before returning, and a timed variant. Other calls are
generally passed straight through, but with the kernel scheduler locked
to prevent pre-emption.

You don't have to use this facility to fit into the infrastructure though,
and thread safety is not a prerequisite for the rest of the infrastructure.
And indeed certain allocators will be able to do scheduling at a finer
granularity than just locking the scheduler every time.

The odd name is because of an original desire to keep 8.3 filenames, which
was reflected in the class name to make it correspond to the filename.
There used to be an alternative Cyg_Mempoolt template, but that has fallen
into disuse and is no longer supported.


Automatic heap sizing
---------------------

This package contains infrastructure to allow the automatic definition
of memory pools that occupy all available memory. In order to do this
you must use the eCos Memory Layout Tool to define a user-defined section.
These sections *must* have the prefix "heap", for example "heap1", "heap2",
"heapdram" etc. otherwise they will be ignored.

The user-defined section may be of fixed size, or of unknown size. If it
has unknown size then its size is dictated by either the location of
the next following section with an absolute address, or if there are
no following sections, the end of the memory region. The latter should
be the norm.

If no user-defined sections starting with "heap" are found, a fallback
static array (i.e. allocated in the BSS) will be used, whose size can
be set in the configuration.

It is also possible to define multiple heap sections. This is
necessary when you have multiple disjoint memory regions, and no MMU
to join it up into one contiguous memory space. In which case
a special wrapper allocator object is automatically used. This object
is an instantiation of the Cyg_Mempool_Joined template class,
defined in memjoin.hxx. It is instantiated with a list of every heap
section, which it then records. It's sole purpose is to act as a go
between to the underlying implementation, and does the right thing by
using pointer addresses to determine which memory pool the pointer
allocator, and therefore which memory pool instantiation to use.

Obviously using the Cyg_Mempool_Joined class adds overhead, but if this
is a problem, then in that case you shouldn't define multiple disjoint
heaps!


Run-time heap sizing
--------------------

As a special case, some platforms support the addition of memory in the
field, in which case it is desirable to automatically make this
available to malloc. The mechanism for this is to define a macro in
the HAL, specifically, defined in hal_intr.h:

HAL_MEM_REAL_REGION_TOP( cyg_uint8 *regionend )

This macro takes the address of the "normal" end of the region. This
corresponds with the size of the memory region in the MLT, and would
be end of the "unexpanded" region. This makes sense because the memory
region must be determined by the "worst case" of what memory will be
installed.

This macro then returns a pointer which is the *real* region end,
as determined by the HAL at run-time.

By having the macro in this form, it is therefore flexible enough to
work with multiple memory regions.

There is an example in the ARM HAL - specifically the EBSA285.


How it works
------------

The MLT outputs macros providing information about user-defined sections
into a header file, available via system.h with the CYGHWR_MEMORY_LAYOUT_H
define. When the user-defined section has no known size, it determines
the size correctly relative to the end of the region, and sets the SIZE
macro accordingly.

A custom build rule preprocesses src/heapgen.cpp to generate heapgeninc.tcl
This contains TCL "set"s to allow access to the values of various
bits of configuration data. heapgen.cpp also includes the malloc
implementation header (as defined by
CYGBLD_MEMALLOC_MALLOC_IMPLEMENTATION_HEADER) with __MALLOC_IMPL_WANTED
defined. This tells the header that it should define the macro
CYGCLS_MEMALLOC_MALLOC_IMPL to be the name of the actual class. This
is then also exported with a TCL "set".

src/heapgen.tcl then includes heapgeninc.tcl which gives it access to
the configuration values. heapgen.tcl then searches the LDI file for
any sections beginning with "heap" (with possibly leading underscores).
It records each one it finds and then generates a file heaps.cxx in the
build tree to instantiate a memory pool object of the required class for
each heap. It also generates a list containing the addresses of each
pool that was instantiated. A header file heaps.hxx is then generated
that exports the number of pools, a reference to this list array and
includes the implementation header.

Custom build rules then copy the heaps.hxx into the include/pkgconf
subdir of the install tree, and compile the heaps.cxx.

To access the generated information, you must #include <pkgconf/heaps.hxx>
The number of heaps is given by the CYGMEM_HEAP_COUNT macro. The type of
the pools is given by CYGCLS_MEMALLOC_MALLOC_IMPL, and the array of
instantiated pools is available with cygmem_memalloc_heaps. For example,
here is a sample heaps.hxx:

#ifndef CYGONCE_PKGCONF_HEAPS_HXX
#define CYGONCE_PKGCONF_HEAPS_HXX
/* <pkgconf/heaps.hxx> */
 
/* This is a generated file - do not edit! */
 
#define CYGMEM_HEAP_COUNT 1
#include <cyg/memalloc/dlmalloc.hxx>
 
extern Cyg_Mempool_dlmalloc *cygmem_memalloc_heaps[ 2 ];
 
#endif
/* EOF <pkgconf/heaps.hxx> */

The array has size 2 because it consists of one pool, plus a terminating
NULL.

In future the addition of cdl_get() available from TCL scripts contained
within the CDL scripts will remove the need for a lot of this magic.


dlmalloc
--------

A port of dlmalloc is included. Far too many changes were required to make
it fit within the scheme above, so therefore there was no point
trying to preserve the layout to make it easier to merge in new versions.
However dlmalloc rarely changes any more - it is very stable.

The version of dlmalloc used was a mixture of 2.6.6 and the dlmalloc from
newlib (based on 2.6.4). In the event, most of the patches merged were
of no consequence to the final version.

For reference, the various versions examined are included in the
doc/dlmalloc subdirectory: dlmalloc-2.6.4.c, dlmalloc-2.6.6.c,
dlmalloc-newlib.c and dlmalloc-merged.c (which is the result of merging
the changes between 2.6.4 and the newlib version into 2.6.6). Note it
was not tested at that point.            


Remaining issues
----------------

You should be allowed to have different allocators for different memory
regions. The biggest hurdle here is host tools support to express this.

Currently the "joined" allocator wrapper simply treats each memory pool
as an equal. It doesn't understand that some memory pools may be faster
than others, and cannot make decisions about which pools (and therefore
regions and therefore possibly speeds of memory) to use on the basis
of allocation size. This should be (configurably) possible.


History
-------


A long, long time ago, in a galaxy far far away.... the situation used to
be that the kernel package contained the fixed block and simple variable
block memory allocators, and those were the only memory allocator
implementations. This was all a bit incongruous as it meant that any code
wanting dynamic memory allocation had to include the whole kernel, even
though the dependencies could be encapsulated. This was particularly silly
because the implementation of malloc() (etc.) in the C library didn't use
any of the features that *did* depend on the kernel, such as timed waits
while allocating memory, etc.

The C library malloc was pretty naff then too. It used a static buffer
as the basis of the memory pool, with a hard-coded size, set in the
configuration. You couldn't make it fit into all of memory.

Jifl
2000-07-03

//####ECOSGPLCOPYRIGHTBEGIN####
// -------------------------------------------
// This file is part of eCos, the Embedded Configurable Operating System.
// Copyright (C) 1998, 1999, 2000, 2001, 2002 Red Hat, Inc.
//
// eCos is free software; you can redistribute it and/or modify it under
// the terms of the GNU General Public License as published by the Free
// Software Foundation; either version 2 or (at your option) any later version.
//
// eCos is distributed in the hope that it will be useful, but WITHOUT ANY
// WARRANTY; without even the implied warranty of MERCHANTABILITY or
// FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
// for more details.
//
// You should have received a copy of the GNU General Public License along
// with eCos; if not, write to the Free Software Foundation, Inc.,
// 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
//
// As a special exception, if other files instantiate templates or use macros
// or inline functions from this file, or you compile this file and link it
// with other works to produce a work based on this file, this file does not
// by itself cause the resulting work to be covered by the GNU General Public
// License. However the source code for this file must still be made available
// in accordance with section (3) of the GNU General Public License.
//
// This exception does not invalidate any other reasons why a work based on
// this file might be covered by the GNU General Public License.
//
// Alternative licenses for eCos may be arranged by contacting Red Hat, Inc.
// at http://sources.redhat.com/ecos/ecos-license/
// -------------------------------------------
//####ECOSGPLCOPYRIGHTEND####