Short: AllocP32 - BetterAlloc (AllocMem patch for PPC) V2.01 Author: thor@math.tu-berlin.de Uploader: thor math tu-berlin de Type: util/boot Version: 2.01 Architecture: m68k-amigaos (with the friendly permission of the original author, Andreas Kleinert) _____________________________________________________________________________ New in release 2.01: - Added a workarount for stupid "SaferPatches" clones. _____________________________________________________________________________ New in release 2.00: - Rewritten in assembly language, much shorter file - Tries a partial flush first, instead of removing all libraries on failure. - Removed unnecessary AllocVec patch since it calls AllocMem anyways. - Added workarounds for bugs in the layers.lib _____________________________________________________________________________ AllocP32 will be part of a major memory defragmentizer project still to be published, called PoolMem. Please stay tuned! _____________________________________________________________________________ WARNING: THIS PATCH ISN'T WITHOUT QUIRKS! READ THIS FILE CAREFULLY BEFORE INSTALLING ALLOCP32. REALLY, I MEAN IT! _____________________________________________________________________________ AllocP32 is identical to AllocP EXCEPT that it allocates memory in 32 byte wide blocks to avoid conflicts between PPC and MC68K microprocessors. If you don't have a PPC board, this patch is most likely unnecessary and will only slow down your computer and waste memory. Use AllocP instead. This program is still in an "experimental" phase, even though it has been tested now. No guarantee that this works with a PPC board, I don't have one for testing! (However, it's welcome as a gift.... ;-) _____________________________________________________________________________ Sometimes programs fail with a "not enough memory" error, but after calling "avail flush" the same operation does succeed without problems. If AllocMem routine in the ROMs did not find enough memory, it tries to flush disk based libraries and devices and, afterwards, tries again to reallocate the memory. However, due to a design flaw of the AllocMem() routine, this memory flushing does not have the desired result some- times - even though the libraries have been informed to remove themselves, the memory is not available directly afterwards. The AllocMem() call will fail anyways, EVEN THOUGH the requiested memory will be available immediatly after completion of AllocMem(). The reasons for this strange behaiviour are rather technical and explained below, for the interested user. This patch does ensure that AllocMem/AllocVec won't fail unless there's really no memory available, even by flushing. It does ALSO, and in addition to AllocP, ensure that the allocated memory is contained in 32 byte memory blocks and not on boundaries of these blocks, so match in one cache line of either the PPC or the MC68K microprocessor, not in both. This avoids that parts of the allocated memory is in different caches of the two CPUs and not up-to-date. The patch DOES NOT ensure that the memory is 32 byte aligned! (The current implemenation returns always memory eight bytes on top of a 32 byte boundary, but don't take this for sure!) The patch does NOT apply to structures allocated by the layer library so take care! The reason is a design flaw (avoiding to call this a serious bug) in the layers library, which causes also some trouble for MungWall and MemSniff. See below for what this bug is about. WARNING: Due to this workaround of the layers.libary bug, it is currently NOT allowed to install an AllocMem() patch ON TOP of this one or it will CRASH! If a second patch of AllocMem is unavoidable, run AllocP32 LAST! To find out which patches are installed in your system, get the "SaferPatches" package from my home page at http://www.tu-berlin.de/~thor/thor/index.html or from the AmiNet. Because of the memory alignment, AllocP32 must allocate more memory than requested and will "round up" the memory size to fit entirely in 32 byte blocks. This DOES, of course, waste memory! Don't use this patch without good reason! (good reason = a PPC board!) Note: Works now for all operating systems and all CPUs, is no longer restricted to V37 or MC'20. Does not require a PPC, but if you don't have one, try AllocP instead - it's much less memory consuming. Usage: Try starting in the Shell/CLI. If it does run stable, copy it into your C: directory and add it somewhere into your s:user-startup AllocP32 You use this patch at your own risk. No guarantee for anything. Source code in assembly language included, requires the DevPac assembler and my macro package at dev/asm/DvPkMacros.lha in the AmiNet. All mentioned trademarks are subject to their owners. _____________________________________________________________________________ The design flaw in AllocMem(): When looking closely at the ROM routine of AllocMem(), you'll notice that a memory flush is TRIED if the first allocation failed. Why does AllocMem then fail, even though if enough memory is available? Consider the following situation: A library has launched a sub-task for control of some of its features. This sub-task could be used for disk I-O, for example. If this hypothetical library has to be expunged, it can't do this on its own because this will release memory of the sub-task which is still running. This sub-task must be informed somehow that its memory must go. While this is in principle simple to do - just send a message to the sub task - the library can't wait for an answer of the sub-task because this would brake the Forbid() state, which is illegal at this point of the operation. The only solution for the library is to leave the memory flush to the sub task and to return to the low-memory handler of exec WITHOUT trying to remove itself - leaving this to its sub-task. On the other hand, since multitasking is still forbidden, there is NO chance that this subtask will actually get a chance to remove the library, leaving it in memory and causing the allocation failure. EVEN THOUGH the library was informed to get removed, there is no chance of doing this while flushing. However, as soon as exec leaves the AllocMem routine, multi- tasking will be turned on again, thus causing a task switch to the subtask and NOW causing the removal of the library - to late to have any effect. AllocP tries to avoid this situation by calling AllocMem a second time. _____________________________________________________________________________ The bug in the layers.library: The layers.library allocates memory for tempory region information in big chunks and cuts this memory apart in smaller pieces, freeing the snippets individually. This is according to the RKRM standards ILLEGAL and has caused a lot of problems (including crashes of earlier versions of AllocP32) not only for AllocP32. Please DO NOT share layers.library structures between different CPUs until this has found a solution in a bugfixed OS version. If this can't be avoided, flush the data caches! _____________________________________________________________________________ The THOR-Software Licence This License applies to the computer programs known as "AllocP32". The "Program", below, refers to such program. The programs and files in this distribution are freely distributable under the restrictions stated below, but are also Copyright (c) Thomas Richter. Distribution of the Program by a commercial organization without written permission from the author to any third party is prohibited if any payment is made in connection with such distribution, whether directly (as in payment for a copy of the Program) or indirectly (as in payment for some service related to the Program, or payment for some product or service that includes a copy of the Program "without charge"; these are only examples, and not an exhaustive enumeration of prohibited activities). However, the following methods of distribution involving payment shall not in and of themselves be a violation of this restriction: (i) Posting the Program on a public access information storage and retrieval service for which a fee is received for retrieving information (such as an on-line service), provided that the fee is not content-dependent (i.e., the fee would be the same for retrieving the same volume of information consisting of random data). (ii) Distributing the Program on a CD-ROM, provided that the files containing the Program are reproduced entirely and verbatim on such CD-ROM, and provided further that all information on such CD-ROM be redistributable for non-commercial purposes without charge. Everything in this distribution must be kept together, in original and unmodified form. Limitations. THE PROGRAM IS PROVIDED TO YOU "AS IS," WITHOUT WARRANTY. THERE IS NO WARRANTY FOR THE PROGRAM, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT OF THIRD PARTY RIGHTS. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION. IF YOU DO NOT ACCEPT THIS LICENCE, YOU MUST DELETE ALL FILES CONTAINED IN THIS ARCHIVE. _____________________________________________________________________________ Thomas, October 1997