June 96 - KON & BAL's Puzzle Page: New World Order
KON & BAL's Puzzle Page:
New World Order
CAMERON ESFAHANI and ALEX ROSENBERG
See if you can solve this programming puzzle, presented in the form of a dialog
between Cameron Esfahani (cam) and Alex Rosenberg. The dialog gives clues to
help you. Keep guessing until you're done; your score is the number to the left
of the clue that gave you the correct answer. Even if you never run into the
particular problems being solved here, you'll learn some valuable debugging
techniques that will help you solve your own programming conundrums. And you'll
also learn interesting Macintosh trivia.
Alex Hey cam, Marathon crashed in a weird manner when I tried to play it under
an early version of Mac OS 8.
cam Working hard, eh? I suppose you'll tell me this was compatibility
Alex Yeah, well, it makes a good demo.
cam Hey, wait a minute! If it's an early version of Mac OS 8, Puzzle Page
readers won't ever find this bug, and they'll write nasty letters to the editor
Alex Well, they should know that they'll learn some valuable debugging
techniques that they can apply to their own programming conundrums. Don't they
read the intro to the Puzzle Page?
cam I guess not. Anyway, back to your problem. So, in what way does it fail?
Alex Just after launch, the machine freezes. This happens every time; it's
100% reproducible. I can't seem to get into MacsBug.
cam Philistine! We use the one true debugger, the Macintosh Debugger
for PowerPC. Mac OS 8 debugging is generally done from a second machine over a
serial cable. You're probably frozen because the program has crashed and the
debugger has halted the machine, waiting to start a debugging session.
Alex What? I thought I gave that up with my Lisa. This is a Macintosh, after
cam Kernel-based operating systems are typically developed with two-machine
debuggers. Besides, think of the wonderful third-party opportunity!
Alex Um, yeah. Anyway, you've hooked up the serial cable and are running
the debugger on the second machine. After watching a progress bar for a while,
you see a dialog that says "Access Fault."
cam An access fault is caused by an attempt to access an illegal address. The
PC is at the instruction that caused the fault.
Alex There's a Show PC command in the debugger's Extras menu. It puts me at
cam Right. We need to isolate whether this fault occurred in application code
or in the system. Choose Show Fragment Info from the Views menu and type that
address into it.
Alex I can't type anything; the machine is crashed. Oh, I get it: I have to
keep switching my head back and forth between machines like a spectator at a
tennis match. What fun. So, how long does this barber pole thingy spin for,
anyway? Hey look, the Fragment Info window highlighted the Marathon code
fragment. The PC is in Marathon's code.
cam What does the code around the PC look like?
Alex It looks like this:
626FDE34 mflr r0
626FDE38 stw r0,0x0008(SP)
626FDE3C stwu SP,-0x0038(SP)
626FDE40 lwz r4,0x0000(r3)
626FDE44 lha r3,0x0000(r4)
626FDE48 bl _eGetDCtlEntry
626FDE4C lwz RTOC,0x0014(SP)
626FDE50 lwz r12,0x0000(r3)
626FDE54 lbz r3,0x0028(r12)
626FDE58 extsb r3,r3
626FDE5C lwz r0,0x0040(SP)
626FDE60 addic SP,SP,56
626FDE64 mtlr r0
R3 is the return value from the function call to _eGetDCtlEntry and
apparently contained a bad address.
cam Choose Show Registers from the Views menu. This will show all the
registers of the current process.
Alex It looks like the return value for _eGetDCtlEntry was 0. The lwz
instruction is dereferencing R3 and putting the result in R12.
cam If you select the R12 register and choose Show Memory from the Views menu,
you can see the memory at that address.
Alex That entire area of memory is full of 0xEEEEEEEE's.
cam That's unmapped memory. Is _eGetDCtlEntry the internal name of the routine
Alex Yes, the debugger is able to pick up that name by using a "trace-back
table," which is the PowerPC equivalent of MacsBug symbols. I guess the next
step would be to figure out why GetDCtlEntry is returning nil. What is it
supposed to be doing?
cam According to Inside Macintosh: Devices, GetDCtlEntry returns the device
control entry for the device specified by the value passed in refNum.
If we look at the rest of the code in this function, right before we call
GetDCtlEntry, we seem to be getting the refNum from the first 16 bits of some
"handle" (or some other kind of pointer to a pointer), which is getting passed
into this function.
Alex All right, but we're going to have to restart. Any information passed
into this function has been lost because we're after the call to
cam To restart we'll need to remember the offset into the Marathon fragment
where the fault will occur, because the Marathon fragment could be loaded in a
different address range.
Alex The offset can be calculated by subtracting the faulting address from
the beginning address for the fragment, which was shown in the Fragment Info
window. For this address, the offset is 0x4832C.
cam Right, but we'd like to get control a little before the actual crash. The
offset to the beginning of that function is 0x48310. Restart the system, and
hold down the Control key when you relaunch Marathon. On a debugging system,
this will break into the debugger after it has completely loaded the
application but just before it begins to execute it.
Alex All right. The machine seems to have stopped at that point. The new start
of the Marathon code fragment is 0x6337D6A0. Adding in the offset of 0x48310,
we get an address of 0x633C59B0.
cam Bring up the Show Instructions window and enter 0x633C59B0 as the address.
It will be exactly the same code as what we saw before. Set a breakpoint at the
first instruction in this function -- the mflr instruction -- and run.
Alex We've reached that breakpoint.
cam Do a stack crawl and see who called us. Head over to the ever useful Views
menu; there's a Show Stack Crawl command.
Alex All right. Apparently the caller is address 0x633988E0.
cam OK, let's step through this function and see what they end up passing to
GetDCtlEntry for the refNum.
Alex It looks like they're passing in 0 for the refNum.
cam Well, there's your problem: 0 is not a valid refNum. It seems that they're
getting an invalid refNum from some part of the system and passing that to
GetDCtlEntry. GetDCtlEntry is returning nil and we're crashing by dereferencing
Alex Uh, that's great, but I still can't play Marathon.
cam Where does the caller of this function, 0x633988E0, get the "handle"
Alex I'll bring up an instruction window at that address:
63398778 mflr r0
633987B0 lwz r24,-0x0218(RTOC)
633988D8 lwz r3,0x0000(r24)
633988DC addic r3,r3,0x001C
633988E0 bl $+0x2D340 ; 0x633C5C20
633988E8 stw r3,0x0000(r30)
R3 seems to be loaded from a global. Let's figure out where this global gets
initialized. The "handle" lives inside a structure that's pointed to by the
global at -0x218(RTOC). This pointer has the "handle" stored at offset 0x1C
cam We could try to track down where the field at offset 0x1C gets
initialized. A pointer wouldn't move around, so we wouldn't have to worry about
relocation. We can use the Data Breakpoint window feature of the Macintosh
Debugger. The PowerPC 601 has a special register that allows you to stop
execution whenever a specified address is read or written to. It's like
hardware support for our old friend step spy.
Alex Sounds like a plan. So, I bring up the Data Breakpoint window and will
break whenever someone writes to that address.
cam Of course, you realize that just as the code fragment could be loaded in a
different place each time it's launched, the RTOC could have a different value
Alex Good point. I'll be sure to use the new RTOC value when I restart
cam What happens after we set up the data breakpoint?
Alex It's kind of strange. We seem to be stopping a lot, but people aren't
writing to offset 0x1C in this structure; they seem to be writing 32 bits to
offset 0x1A and overwriting 0x1C.
cam I don't understand. The routine that called the crashing routine was
passing in a value at offset 0x1C.
Alex Apparently we calculated something wrong.
cam I don't know where we could have gone wrong. Wait a second. Look at
address 0x633988E0; it says we're branching to address 0x633C5C20, but the
routine we're crashing in is at 0x633C59B0.
Alex Well, maybe it's just a call to a different code fragment, a
"cross-TOC" call, and it has to use some indirection to get to the crashing
cam No, it can't be a cross-TOC call, for two reasons: first, there's no TOC
reload after the function call, and second, the routine we crashed in is in the
Marathon fragment. You saw the Fragment Info results.
Alex I'll buy that. Now let's set our breakpoint just before this function
call to the unknown address. We can step through that code.
cam All right. After we hit the breakpoint, step into that function.
45 Alex Holy cow! It's some totally different piece of code.
cam Look through the instruction disassembly of this new routine. Is there
anywhere in there where they call the crashing function?
633C5C98 bl 0x6337E150
633C5C9C lwz RTOC,0x0014(SP)
633C5CA0 ori r31,r3,0x0000
633C5CDC ori r3,r31,0x0000
* 633C5CE0 bl 0x633C59B0
Alex Yeah. At address 0x633C5CE0, they call our crashing function with a
parameter obtained from R31. Working further back in the instruction
disassembly, we see that a function call is made and the result of that is put
in R31. This occurs at 0x633C5C98. It calls a routine at address 0x6337E150.
cam And looking at that, it appears to be a cross-TOC call. I restart Marathon
and step into the routine at 0x6337E150.
Alex It appears to be cross-TOC glue:
6337E150 lwz r12,-0x0A90(RTOC)
6337E154 stw RTOC,0x0014(SP)
6337E158 lwz r0,0x0000(r12)
6337E15C lwz RTOC,0x0004(r12)
6337E160 mtctr r0
cam So apparently Marathon is calling another library to get this mystical
Alex Yep. Whenever you're going to go from one library context to another,
you need to save and restore the TOC. That's one of the things this glue code
does. As you can see, R12 is loaded from -0x0A90(RTOC). R12 will contain a
pointer to a transition vector, which contains an address of a routine and a
new TOC value. The transition vector is imported from the library we're linking
cam So we should be able to plop the transition vector address into the
Fragment Info window and figure out which library that comes from, right?
Alex Good idea. Dumping the address at -0x0A90(RTOC) we get the following:
0200CC0C: 01FC9D98 01FC9DA4 01FC9DB0 01FC9DBC
0200CC1C: 01FC9DE0 01FC9DEC 01FC9D80 01FC9D74
cam I use the Fragment Info window to find out which fragment contains the
Alex It seems to live in the QuickDraw data section, which makes sense,
since a transition vector is data.
cam Aha! QuickDraw! That figures. And you wondered why they call it KON &
BAL's Puzzle Page. I use the Show Exports button in the Fragment Info window to
list all of the exports of the QuickDraw library.
Alex I was wondering when we were going to use that button. You end up getting
a long list of all the routines exported by QuickDraw, sorted alphabetically.
cam But I have an address I want to match. If you click on the Address column
title, that list will get resorted by address. I search through the list for
Alex That address is the address of the GetDeviceList transition vector.
cam That makes sense. This "handle" we've been worrying about has a refNum in
the first 16 bits of the structure, and a GDHandle has the refNum of the
associated driver stored in the first field in the structure. I bring up a
memory window and examine the device list stored in low memory to see if the
gdRefNum is 0.
Alex It is. Who's responsible for initializing the GDevice record?
cam NewGDevice and InitGDevice. NewGDevice will pass the refNum to
InitGDevice. Let's disassemble the code for NewGDevice.
Alex Apparently it does a NewHandleClear to allocate the GDHandle, and never
initializes the refNum.
cam Whoops. Ah, the joys of pre-alpha software. Well, it should be reasonably
easy to get one of the QuickDraw engineers to fix this bug. I install a fixed
version of the QuickDraw shared library. We should be rockin' now!
Alex Not so fast! When I restart Marathon, I crash. If I do a stack crawl and
examine the code, I seem to be crashing in exactly the same place. GetDCtlEntry
still seems to return nil.
cam Just another day at the salt mines. OK, it's time to step through
GetDCtlEntry. I put a breakpoint just before we call it.
Alex The refNum from the GDHandle is -51.
cam That looks like a valid refNum. Step into GetDCtlEntry.
Alex We first go through the cross-TOC glue and eventually get into
cam What does GetDCtlEntry look like?
Alex It seems fine, but when you step through the routine and actually fetch
the DCtlHandle from the unit table, it ends up being nil.
626FC79C cntlzw r3,r3
626FC7A0 srwi r3,r3,5
cam That obviously shouldn't happen. There's a driver entry in the table and
the refNum seems valid. What is the code at 0x626FC79C doing?
Alex It's performing a logical NOT operation on R3. Groovy, huh?
cam But the bitwise NOT of the refNum should be used as the index into the
unit table, not the logical NOT!
Alex So, you claim that GetDCtlEntry is looking at the wrong place in the unit
table to get the DCtlHandle. Are you sure?
cam Let's go to the source. What does Inside Macintosh: Devices say?
Alex It says, "The device reference number is the one's complement (logical
NOT) of the unit number." But the logical NOT isn't the one's complement; the
bitwise NOT is.
cam Um, OK, what does Inside Macintosh Volume II say?
Alex It claims that the unit number is "equal to -1 * (refNum + 1)."
cam And that's a bitwise NOT. So it seems that Inside Macintosh: Devices is
wrong. Weird, wacky stuff. But I still don't understand why the stack crawl we
did earlier pointed us to the wrong place.
Alex We did the stack crawl when we had just entered the crashing function,
even before it executed the mflr instruction. The debugger, when it does a
stack crawl, is going to look for stack frames to see where the callers are.
Since we hadn't allocated a stack frame in the crashing function yet, we were
still using the caller's stack frame. So that was the stack frame the debugger
started from. If we had stepped a few more instructions in and allocated our
stack frame, the debugger would have figured it out. It would be interesting to
see if the debugger could actually detect the case of a nonexistent stack frame
and use the link register to work back to the caller.
cam That also explains why the data breakpoint stuff didn't work. We thought
that the data structure we were watching held the GDHandle. It didn't; it
contained something totally unrelated, which was passed into the function that
called the crashing routine.
Alex So, because NewGDevice didn't initialize the gdRefNum and GetDCtlEntry
was returning the wrong entry from the unit table, I don't get to teach the
Pfhor about large caliber, high-velocity rounds.
Your performance compares to these memorable screen roles:
80-100 Chuck Heston in El Cid
55-75 Anne Parillaud in La Femme Nikita
30-50 Chow Yun-Fat in Hard Boiled
5-25 Richard Roundtree in Shaft*
CAMERON ESFAHANI (firstname.lastname@example.org, AppleLink DIRTY) SWM,
24, 5' 7", 180 pounds, brown hair, brown eyes. Apple engineer. Loves movies and
music. Plays golf and tennis, rollerblades, ice-skates, and is learning to ski
again. Enjoys life but has a serious side. Likes cooking, reading, shopping.
Once a dog guy, now a definite cat man. Believes the American musical of the
'50s and '60s to be the second greatest invention of the twentieth century.
Favorites include West Side Story, The Sound of Music, Music Man, and Singin'
in the Rain.*
ALEX ROSENBERG (email@example.com) Alex's left brain works on everything from
communications software to the latest 3D graphics tricks. His right brain is
constantly thinking up interesting T-shirts that Apple's Marketing folks don't
approve of. While working as a member of the Mac OS 8 "Ministry of
Information," he experimented with optimization for PowerPC, worked closely
with Apple's compiler team, and contributed to IBM's The PowerPC Compiler
Writer's Guide. Now one of the minions at Bungie Software, Alex recently
decided that eating is overrated.*
Thanks to Tom Dowdy, Wayne Meretsky, Mike Neil, Tom Saulpaugh, KON (Konstantin
Othmer), and BAL (Bruce Leak) for reviewing this column.*