Important: This document is part of the Legacy section of the ADC Reference Library. This information should not be used for new development.
Current information on this Reference Library topic can be found here:
|
IntroductionThe introduction of the PowerPC Macintosh also introduces a new Memory Manager. Many of the splendid features of this new implementation have already been discussed in tech note Memory 13. Weary travelers may have noted some strange behavior though, and this note attempts to answer most of the remaining big questions you may have about it. Bus Error HandlersWhen 32-bit QuickDraw was introduced around the Macintosh IIci, it became
immediately possible that frame buffers would exceed the maximum address space
available within a Macintosh NuBus slot with 24-bit addressing. (As you may
remember, NuBus slots have only 1/16th of the addressable space when accessed
in 24-bit mode). This problem was initially solved by making a 32-bit
addressing mode that could be enabled via When the 32-bit clean Memory Manager was introduced, a (then-helpful) side
effect of the bus error handlers was to filter bad addresses and gracefully
return an error code instead of crashing. Unfortunately, bus error handlers
are still there watching out for 24-bit handles. Programmers may not be
passing in bad handles any more, but the bus error handlers watch to make sure
that all the major dereferences to find the zone do not fail. Because the
handlers were there protecting you from yourself (much like the USA does with
seat belt laws), you didn't notice some faulty address calculations in your
code actually cause a bus error in the Memory Manager. All your code would
have seen is a -111, Commonly, these faulty calculations come from assuming the size and shapes of block and heap headers. This is a bad thing. The sizes of block and heap headers have both changed with the Modern Memory Manager, and are bound to change again in the future. Don't rely on undocumented features, including features such as structure sizes that you deduced from the documentation but were not explicitly documented. If you have concerns whether you will break in the future, it is best to contact DTS now. What Was DoneUnfortunately, the bus error protection provided by the Memory Manager was very time consuming, and with the introduction of future operating systems that protect these vectors, it was not going to get any faster. Most importantly, now that addresses are always real logical addresses, the handlers are just extra baggage. The best option was to completely remove the handlers, and this was tried unsuccessfully in early versions of the Modern Memory Manager shipping with PowerPC. Unfortunately, there are still too many important clients that rely on the side effects provided to simply remove them. Out of the ashes of this came a happy medium. On the PowerPC machines with the
Modern Memory Manager, bus error handlers are installed, but the first stop on
the exception journey is to the PowerPC debugger, not directly to the Memory
Manager exception handler. Should you be one of the lucky abusers of the
Memory Manager, you will see the debugger stop in the Memory Manager code with
an "Access Fault." Unlike using Remember when you propagate the exception, your chance to debug the problem is lost. Fixing these problems in your own source early on is the best way to feel confident that late in testing you will not have to propagate exceptions, only to find that it wasn't the Memory Manager, and that there was no exception handler to catch your fall. There are some lessons to be learned here. The first is if you administrate systems as a hobby or know of someone who does, be sure that they do not put the PowerPC Debugger Nub on end-user PowerPC machines. Users will be confused as their machines freeze for 15-30 seconds at a time, then suddenly continue at the blazing speed PowerPC's are known and loved for. The second lesson is that if you are developing on a PowerPC and you see an access fault, try to track down who it is. If it is in your code, fix it. If it is in someone else's code, try to contact them and get them to fix it. If you like the set of Extensions, Control Panels, and Applications you are currently using and they cause access faults, you will be mighty unhappy when the safety net is removed and they don't all work in the next release. Your letter or phone call to the developer might make the difference. And if it is your code, it is pure suicide not to test at this point on a PowerPC. The bus error handlers, even though they are implemented better than their ancestors, are not cheap. They will be removed at the next speed release. Bus Error Return ValuesOn the subject of the bus error handlers, the Modern Memory Manager no longer
returns both -111, The difference between these errors was the by product of the need to know which
address needed to be passed to Free Block MiscellaneaWhile on the subject of the new Memory Manager, it is important to remind you that disposing blocks is hazardous to their integrity. Random data scattered throughout the newly freed block are targets for our gratuitous and spiteful clobbering. This has been documented before but it never hurts to mention it again. While this point was undoubtedly well taken, only top scorers on KON & BAL's Puzzle Page figured out that there are other side effects that will bring this behavior to life. While this is no attempt to list all of them, the general ideas here will be enough to seed your imagination, and maybe even get a better score on the next Puzzle Page. The first is closing out a resource file while the data in the resources
belonging to that file are still in use. Inside Mac tells you that
A similar situation occurs when there is a purgeable handle (including
purgeable resources) around and you expect to use it after calling something
that allocates memory. This has always been a problem, but can be a problem in
different ways now that the integrity of freed blocks is guaranteed nil and the
dynamics of the memory manager are different, including different algorithms
and different block sizes causing the different algorithms to act differently
(see, things really are different). One bug we saw recently was that You can find these pretty easily though. Using the A5 World Problems and Heap Callback ProceduresMany Memory Manager savvy applications use Grow Zone and Purge procedures to
indicate when they should clean up or release memory so that the system can
continue to function. A common technique is to register a grow zone procedure
(using Unfortunately, Since this problem is so rare, we did not come across it until just recently.
Options for those truly stuck by this or not interested in finding this is a
problem later include tracking your own A5 via ConclusionWith the exception of the Heap Callback problems, the gray area of where programs work but are not really correct in the eyes of the Memory Manager is getting smaller. This may seem like a burden, but all of these points are problems waiting to happen, whether you are using the Traditional or Modern Memory Manager. Correcting them now will ensure that as the rules get stricter, you are not left in the compatibility doghouse with your customers. ReferencesInside Macintosh, Designing Cards and Drivers for Macintosh Technical Note Memory #13 Downloadables
|