March 93 - MACINTOSH DEBUGGING: THE BELLY OF THE BEAST REVISITED
MACINTOSH DEBUGGING: THE BELLY OF THE BEAST REVISITED
FRED HUXHAM AND GREG MARRIOTT
ADAPTED FROM THEIR TALK AT THE WWDC BY DAVE JOHNSON
This is a supplement to the article "Macintosh Debugging: A Weird Journey Into
the Belly of the Beast" in Issue 8 of develop. It presents
a few debugging tools that were discussed at Apple's Worldwide Developers
Conference in May 1992. Like those discussed in the previous article, these tools are
designed to help you force the nasty, subtle bugs in your code to show their hideous
little faces immediately, rather than lying in wait and biting you when you least
expect it.
People often ask us, "How can I be a totally awesome, godlike debugging stud [or studette] like you?"
Unfortunately, the big truth from the Issue 8 debugging article is just as true now as it was then: debugging is hard. That's just the way it is. The only way to get better at it is to practice. Now that
we've got that straight and before we get into describing the new debugging tools, here are three
pearls of wisdom to guide you in your practice.
First of all, it helps to know a lot about the operating system. The better sense you have of how the
Macintosh works, the better off you'll be trying to track down a nasty bug. Dare to delve into the
bowels of the OS. Read and rereadInside Macintosh; take it with you to bed, to the bathroom, out to
dinner, and on dates. (You might want to invest in a sturdy wheelbarrow, especially with the new Inside Macintosh
volumes proliferating like rabbits.) For that matter, read every Macintosh
programming book ever written (especially those listed at the end of this article) and every Technical
Note, Snippet, piece of Sample Code, and issue ofdevelop, as well as every word on the AppleLink
Discussion boards. Also, spend lots of time in debuggers, watching the system do its thing. If you're
not dreaming in hex, you're not spending enough time in MacsBug.
Second, get slammed a lot. The people who are the best at debugging are usually the ones who've
had to track down the most bugs and therefore have an encyclopedic knowledge of them. If you have
a really nasty bug in your code that crashes the machine on a seemingly random basis and takes you
three days to find and squash, then by jove you'll remember that bug the next time you see it. Simply
put, the more bugs you find, the better you'll be at finding bugs.
Last, use good tools, and use them all. Reread the Issue 8 article. Turn on those tools and stress your
code. Bend, fold, staple, and mutilate it. Show no mercy.
These things will help you on your way to becoming a primo bug stomper, but debugging is like any
complex skill in that it can't really be taught past a certain point. You simply have to do it a lot, andover time you'll get better. Tools and techniques such as the ones presented here can help
enormously, especially by forcing hidden bugs to the surface, but they can never do the whole job for
you.
This time there are only four new tools to talk about -- Double Trouble, Dispose Resource, Blat,
and Smart Friends -- so this article is much shorter than the last one. The tools are available on theDeveloper CD Seriesdisc, as well as on AppleLink and elsewhere. We're doing this backward from the
last time: first we'll present a buggy code sample, then we'll talk about the tool that would find the
bug.
DOUBLE TROUBLE
Can you find the bug in this code sample?
myHandle = NewHandle(100);
if (myHandle) {
AddResource(myHandle, 'dumb', 10, "\p");
if (resError()) HandleTheError();
CloseResFile(outputFileRef);
DisposeHandle(myHandle);
}
OK, time's up. This one's not too hard. The problem is that during CloseResFile the Resource
Manager disposes of all the resources in memory. The DisposeHandle call afterward is unnecessary
and is actually potentially disastrous. Normally you'll just get an error and DisposeHandle will do
nothing, but occasionally the data structures in the Memory Manager will conspire to really screw
you.
Here's how: Master pointers are allocated in clumps called master pointer blocks, which are
nonrelocatable blocks in your application's heap. The master pointers that are currently free for use
are kept in a linked list by the Memory Manager. The list is LIFO, like a stack: when you allocate a
new handle, the Memory Manager uses the first master pointer in the free list, and when you dispose
of a handle the freed master pointer is returned to the beginning of the list.
Now the plot thickens. If the first master pointer in the free list also happens to be the first master
pointer in its master pointer block (so that the master pointer and the master pointer block have the
same address) and then you dispose of a handle twice by mistake,very bad thingswill happen. On the
first dispose, everything is fine: the Memory Manager frees the block the master pointer points to
and returns the master pointer to the start of the free list. At this time, the master pointerstill points
to a valid block of memory, but now it's the master pointer block itself! So on the second,
unintentional dispose, when the Memory Manager dutifully frees the block for reuse, you're set up
for disaster. Subsequent memory use will likely result in writing over many master pointers, which
will of course trash you one way or another.
Figure 1 illustrates this scenario. On the left is the top part of a master pointer block that resides in
the heap at address 80. The heap's free list is a standard linked list (each entry contains the next
entry's address) beginning at hFstFree. Note that the first entry in the heap's free list is also the first
master pointer in the block. This is the first step to trouble.
Figure 1 How Disposing of the Same Handle Twice Can Spell Disaster
Now we call DisposeHandle on the master pointer at 81. DisposeHandle looks at the block pointed
to by the master pointer (in this case the block at 144, not shown), determines that it is indeed a valid
block, marks it as free for reuse, and adds the newly freed master pointer to the front of the free list.
So far so good. Now the master pointer block looks like the one on the right in the figure.
Then we call DisposeHandle on 81 again by mistake. DisposeHandle looks at the block pointed to by
the master pointer (now it's the block at 80, our master pointer block!), determines that it is indeed a
valid block (uh oh), marks it as free for reuse (yikes!), and adds the newly freed master pointer to the
front of the free list -- and the heap is now hosed for good. This Memory Manager bug is subtle and
rare, but oh so nasty.
Even if you're lucky enough to avoid this particular sequence of events, a double disposal is definitely
a bug. Double Trouble is a system extension that watches calls to DisposeHandle to make sure it's
not being called on something in the free list. If it is, Double Trouble drops into the debugger with a
suitable warning.
We'll be the first to admit that Double Trouble is far from perfect. It infers the existence of heap
zones by watching InitZone and then trying to figure out when a heap isn't a heap anymore. The
possibility exists that it will guess wrong and cause a bus error when trying to walk a free list that's no
longer a free list. Furthermore, in some cases Double Trouble can noticeably slow down parts of the
system. (After playing a long QuickTime movie, for instance, the machine may freeze for almost a
minute.)
But despite Double Trouble's shortcomings, we do still recommend running it all the time. Just try
to remember that it's running so you don't chase your tail trying to find the cause of occasional
mysterious slowdowns.
DISPOSE RESOURCE
Here's the code. What's the bug?
myPicture = GetPicture(kPicID);
if (myPicture) {
DrawPicture(myPicture, &myRect);
DisposeHandle(myPicture);
}
That's right, you should never call DisposeHandle on a resource handle. If you do, the Memory
Manager will free it just fine, but the Resource Manager has another reference to it, stored in the
resource map, that will be left dangling. Later on, since the Resource Manager doesn't know the
handle was disposed of, it may try some manipulation with the handle. The results may not crash you
immediately, or at
all -- it depends on what the operation is and what's in the handle -- but they're certainly not what
was intended. Instead of DisposeHandle, you should always call ReleaseResource on resource
handles. ReleaseResource will properly dispose of the handleandwill update the resource map. (Note
that KillPicture won't do the right thing here either; it's intended for pictures created via
OpenPicture, not for PICT resources.)
Dispose Resource is another extension a lot like Double Trouble. It also watches DisposeHandle
calls, this time looking to see if the handle being disposed of is a resource handle. If so, you'll drop
into the debugger with a suitable warning.
Dispose Resource has one idiosyncrasy you should know about: it's been known to indicate "false
positives." Some parts of the system (we haven't been able to track down which ones yet) seem to
save a resource handle's state, detach the resource, and then restore the state of the handle (restoring
the resource bit!). Use Dispose Resource. It will ensure that you don't make the same mistake.
BLAT
This time the code's in assembler:
; Offset the rect by 128 pixels in each direction.
PEA theRect(A6)
MOVE.W $0080, -(SP)
MOVE.W $0080, -(SP)
_OffsetRect
If you have "iron man" syndrome and insist on programming in assembly language, this can happen
to you. We forgot to type a # in front of each $0080. As a result, instead of moving the number
$0080 (128) onto the stack twice in preparation for the OffsetRect call, we're moving the contents of
memory location $0080. Often this kind of bug is immediately obvious, but not always. If you're
moving a Boolean, for instance, you have a fifty-fifty chance of getting the right value, even though
you're getting it from some random spot in memory. It's those cases that will give you debugging
headaches.
One easy (and recommended) way to avoid the problem in this example is to write in a higher-level
language. But we realize that's not always possible, and besides, this is really a wholeclassof
problems: reads and writes from places in memory you didn't intend. The best way to catch this wild
memory reference kind of problem is, naturally, with memory protection, something that -- sadly --
the Macintosh normally lacks. In the last article we mentioned Jasik's implementation, but now
there's something else you should know about. Bo3b Johnson has written a dcmd called Blat that uses
the MMU to protect memory locations 0-255 from both reads and writes.
Blat has been tested and works well on the Macintosh IIfx, IIx, and SE/30. Because its operation is so
hardware dependent, it's hard to predict whether it will work on a given machine. Some basic
guidelines are that it requires an MMU and won't work with 68040 machines or with most
configurations of machines with the IIci ROM (IIci, IIsi, LC). For further details, see the release
notes and the source code, thoughtfully provided by Bo3b along with the dcmd itself.
SMART FRIENDS
This bug is subtle, so pay close attention:
#pragma parameter __d0 GetA0
Ptr GetA0(void) = {0x2008}; // MOVE.l A0,D0
void MyCompletionRoutine()
{
long saveA5;
HooHahPtr myHooHah;
myHooHah = (HooHahPtr)GetA0();
saveA5 = SetA5(myHooHah->myA5);
gSomething[0].flag = true; // Set a flag in a global array.
SetA5(saveA5);
}
This code really tries hard to do everything right. As the name implies, it's a completion routine, so
it could be called at interrupt time. First a pointer to the data is retrieved from A0, and then A5 is set
to a previously saved value, thus allowing the routine to access its global variables. Once A5 is set up,
the global reference can be made safely. Finally, A5 is restored to its previous value to clean up.
Sounds great, right? The only problem is, it doesn't work.
Here's why: the MPW C compiler will actually set up the global referencebeforethe SetA5 call, so
accessing the global accesses some unknown part of memory. This is legal compiler optimization
behavior! If GetA0 and SetA5 were functions or traps, the bug would disappear, but since they're
declared inline the compiler doesn't feel compelled to delay the evaluation of the global array
reference. The solution is to set up A5, then call a different routine that does the global reference.
Now in this case, how do you think we -- the debugging gods -- figured out the bug? We tried the
first few things we could think of; but then when we weren't making headway after a few probes, we
didn't just sit there and suffer in silence, banging our heads against the proverbial wall. We called in
some Smart Friends! The veil of illusion was torn from our eyes, and we were shown the heart of the
truth (in other words, one of them had seen this bug before). The point is that in debugging, two (or
more) heads are far, far better than one. Bugs are not like germs: when you share them, everyone
benefits. Maybe your very own Smart Friends have had a similar bug before, so they'll recognize
immediately what's going on. Or maybe they'll think of something different to try. At the very least,
they'll temporarily divert you from your frustration, maybe make you feel less stupid, and then you
can all go out for pie together.
THAT'S IT!
Add these tools to your arsenal of bug sprays and foggers. Use them all and use them well, and you,
your code, and your customers will be far better off.
FURTHER READING
Bedside books for the serious student of debugging:
- How to Write Macintosh Software, 3rd ed., by Scott Knaster and Keith Rollin (Addison-Wesley, 1992).
- Macintosh Programming Secrets, 2nd ed., by Scott Knaster and Keith Rollin (Addison-Wesley, 1992).
- Debugging Macintosh Software with MacsBug by Konstantin Othmer and Jim Straus (Addison-Wesley,
1991).
- MC68000 Family Programmer's Reference Manual (Motorola, Inc.).
FRED HUXHAM (AppleLink FRED) was born and raised in California. He used to be a tremendous athlete, know bazillions
of babes, and go to wild parties in New York and California with people like Andy Warhol and Keith Haring. Now he's
15 pounds heavier, knows only one babe (his wife), and thinks a day spent sitting on his roof deck watching boats go
through the Golden Gate is really exciting.*
GREG MARRIOTT (AppleLink GREG) is a SWM, 28, 6'0", 195 lbs., brown hair and eyes, sincere, hardworking, good
sense of humor. Enjoys music, romantic walks, quiet evenings, and good books. Seeks nice woman for friendship and
more. Send photo.*
In Bo3b's name, the "3" is silent. *
THANKS TO OUR TECHNICAL REVIEWERS Jim Reekes and Bryan Stearns*