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 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 /* */ /* This is a generated file - do not edit! */ #define CYGMEM_HEAP_COUNT 1 #include extern Cyg_Mempool_dlmalloc *cygmem_memalloc_heaps[ 2 ]; #endif /* EOF */ 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####