Self Modifying Code
Volume Number: | | 8
|
Issue Number: | | 4
|
Column Tag: | | Article Rebuttal
|
Self Modifying code is a No-No!
A better way to do an event patch without self-modifying code or Assembly
By Scott T. Boyd, Apple Computer, Inc. and Mike Scanlin, MacTutor Regular Contributing Author
Note: Source code files accompanying article are located on MacTech CD-ROM or source code disks.
Mike Scanlins article Rotten Apple INIT for April Fools brings up a minor but essential point. The GetNextEvent patch looks like:
;1
@first
Lea @exitAddress, A0
Move.L (SP)+,(A0)
Lea @eventRecPtr,A0
Move.L (SP),(A0)
Pea @tailPatch
DCJmpInstruction
@origTrap
NOP
NOP
The NOPs get replaced with the original GetNextEvent trap address by the installation code:
;2
Lea @origTrap,A1
Move.L A0,(A1)
Now, consider the processor instruction cache. Its a piece of the processor which remembers whats at a set of memory locations. It does this so the CPU wont have to do a memory access for recently referenced instructions. This is designed to save time. Its a neat hardware feature.
However, the Macintosh system software doesnt make a distinction between code and data. Thats different from OSs like Unix, which keep code and data in separate address spaces. When Mikes code installs the original trap address with the Move.L A0,(A1), its putting an address into the middle of a piece of code. Unfortunately, the cache which records the new value is the data cache.
The instruction cache has no clue that instructions just changed. This is one way of doing whats commonly called self-modifying code.
Self-modifying code is, in general, a bad thing to write. Apple has long discouraged, and continues to discourage self-modifying code.
This is bad in this case because the processor, if it were to execute this code right away, might believe (for some unspecified reason) that those memory locations were cached in the instruction cache. If it did, it would pull whatever had been in that location the last time code was executed from that spot, then try to execute whatever was there. The odds that the instruction cache held the value you just put into the data cache are not in your favor.
Contrast that approach with this approach:
;3
@first
...
Move.L @origTrap,A0
JMP.L (A0)
...
/* branch around this, or put it somewhere else, but dont let the PC run through
here */
@origTrap
NOP
NOP
While this approach still stores the old address into a piece of code, its never referenced as code by the processor. Its treated specifically as data. The instruction cache never comes into play since the original address is moved as data.
Yet another approach, which also saves a register:
;4
@first
...
Move.L@origTrap,-(SP)
RTS
...
TN #261: Cache As Cache Can discusses this topic in more detail, especially with regard to moving whole chunks of code around.
As it happens, the caches are almost certainly flushed before this particular eight bytes ever get loaded for execution, but thats a happy coincidence, and not something you should rely on. Whats happening is that we have made several traps flush the caches (guaranteeing that there wont be any misunderstanding about something being in the instruction cache when its not), but we may change our minds about which traps should flush, and when. You shouldnt count on any given traps current cache-flushing behavior.
One final consideration. Putting data into code does not work if code is ever write-protected, and that may happen one day. So where can you put something when you cant allocate any global storage (e.g., PC-relative data or low-memory globals with a fixed address)? You can use NewGestalt to register a new selector. When you call Gestalt, it can return a value which is actually a pointer (or handle) to your global data. This technique wont work well if you cant afford to make the trap call (like from some time-sensitive routine youve patched), but it works nicely if you have the time and you want to avoid putting data into your patch code.
Scott T Boyd, Apple Computer, Inc.
Mike Scanlin Says
Scott's point about stale code in the instruction cache is well taken and I deserve a thumping for having written it. I made the poor judgement call that it wouldn't matter in this case because I expected the instruction cache to be flushed between the time the patch installation code finished and the first time the patch code was executed. I hang my head in shame.
As partial retribution (and to satisfy a few requests for a non-assembly version) I have written a trap patching shell in C that doesn't use any self-modifying code (see listing below). It obeys all of the rules except for the one about storing data into a code segment (Scott's solutions have this problem, too, as he mentions). Until we have write-protected code segments, this will not be a problem.
Mike Scanlin
/*********************************************************
* PatchGNE.c:
* This INIT installs a patch on GetNextEvent and
* SystemEvent that intercepts keyDown and autoKey events.
* For this example, the intercepted key events are
* converted to lower case if both the capsLock key and the
* shiftKey are down (thus making the Mac keyboard behave
* like an IBM keyboard). However, you can use this shell to
* do generalized event intercepting as well as generalized
* trap patching (with no asm and no self-modifying code).
* If your patches need globals, put them in the
* PatchGlobals struct and initialize them in main.
* In Think C, set the Project Type to Code Resource, the
* File Type to INIT, the Creator to anything, the Type to
* INIT, the ID to something like 55 (55 will work but it
* doesn't have to be 55), turn Custom Header ON and Attrs
* to 20 (purgeable) and Multi Segment OFF.
*
* Mike Scanlin. 16 May 1992.
*********************************************************/
#include "Traps.h"
/**********************************************************
* typedefs
*********************************************************/
typedef pascal short (*GNEProcPtr)(short eventMask,
EventRecord *theEvent);
typedef pascal short (*SEProcPtr)(EventRecord *theEvent);
typedef struct PatchGlobals {
GNEProcPtr pgOldGNE;
SEProcPtrpgOldSE;
} PatchGlobals, *PatchGlobalsPtr;
/**********************************************************
* prototypes
*********************************************************/
void main(void);
void StartPatchCode(void);
pascal short MyGetNextEvent(short eventMask, EventRecord
*theEvent);
pascal short MySystemEvent(EventRecord *theEvent);
void CheckKeyCase(EventRecord *theEvent);
void EndPatchCode(void);
/**********************************************************
* main:
* Gets some memory in the system heap and installs the GNE
* and SE patches (as well as allocating and initializing
* the patc 8.4 Self Modifying Codeutine that gets
* executed at startup time (by the INIT mechanism).
*
* The block of memory that main allocates will look like
* this when main has finished:
*
* +--------------------+
* | PatchGlobals |
* +--------------------+
* | StartPatchCode() |
* GNE trap addr -> +--------------------+
* | MyGetNextEvent() |
* SE trap addr -> +--------------------+
* | MySystemEvent() |
* +--------------------+
* | CheckKeyCase() |
* +--------------------+
* | EndPatchCode() |
* +--------------------+
*
*********************************************************/
void main()
{
Ptr patchPtr;
PatchGlobalsPtr pgPtr;
long codeSize, offset;
/* try and get some memory in the system heap for code
and globals */
codeSize = (long) EndPatchCode - (long) StartPatchCode;
patchPtr = NewPtrSys(codeSize + sizeof(PatchGlobals));
if (!patchPtr)
return; /* out of memory -- abort patching */
/* initialize the patch globals at the beginning
of the block */
pgPtr = (PatchGlobalsPtr) patchPtr;
pgPtr->pgOldGNE = (GNEProcPtr)
GetTrapAddress(_GetNextEvent);
pgPtr->pgOldSE = (SEProcPtr)
GetTrapAddress(_SystemEvent);
/* move the code into place after the globals */
BlockMove(StartPatchCode, patchPtr +
sizeof(PatchGlobals), codeSize);
/* set the patches */
patchPtr += sizeof(PatchGlobals);
offset = (long) MyGetNextEvent - (long) StartPatchCode;
SetTrapAddress((long) patchPtr + offset, _GetNextEvent);
offset = (long) MySystemEvent - (long) StartPatchCode;
SetTrapAddress((long) patchPtr + offset, _SystemEvent);
}
/**********************************************************
* StartPatchCode:
* Dummy proc to mark the beginning of the code for the
* patches. Make sure all of your patch code is between
* here and EndPatchCode.
*********************************************************/
void StartPatchCode()
{
}
/*********************************************************
* MyGetNextEvent:
* Tail patch on GetNextEvent.
*
* The reason this returns a short instead of a Boolean is
* because we need to make sure the low byte of the top word
* on the stack is zero because some programs do a Tst.W
* (SP)+ when this returns instead of Tst.B (SP)+ like they
* should (which is technically their bug but, we might as
* well work around it since it's not hard).
*
* If you want to eat the event and not pass it on to the
* caller then set returnValue to zero.
*********************************************************/
pascal short MyGetNextEvent(short eventMask,
EventRecord *theEvent)
{
PatchGlobalsPtr pgPtr;
short returnValue;
/* find our globals */
pgPtr = (PatchGlobalsPtr) ((long) StartPatchCode -
sizeof(PatchGlobals));
/* call original GNE first */
returnValue = (*pgPtr->pgOldGNE)(eventMask, theEvent);
/* do some post-processing */
CheckKeyCase(theEvent);
/* return to original caller */
return (returnValue);
}
/**********************************************************
* MySystemEvent:
* Tail patch on SystemEvent.
*
* The reason this returns a short instead of a Boolean is
* because we need to make sure the low byte of the top word
* on the stack is zero because some programs do a Tst.W
* (SP)+ when this returns instead of Tst.B (SP)+ like they
* should (which is technically their bug but, we might as
* well work around it since it's not hard).
*
* We need this patch as well as the one on GetNextEvent
* because of desk accessories. If you don't patch
* SystemEvent then the patch will not apply to events that
* are sent to DAs.
*
* If you want to eat the event and not pass it on to the
* caller then set returnValue to zero.
*********************************************************/
pascal short MySystemEvent(EventRecord *theEvent)
{
PatchGlobalsPtr pgPtr;
short returnValue;
/* find our globals */
pgPtr = (PatchGlobalsPtr) ((long) StartPatchCode -
sizeof(PatchGlobals));
/* call original GNE first */
returnValue = (*pgPtr->pgOldSE)(theEvent);
/* do some post-processing */
CheckKeyCase(theEvent);
/* return to original caller */
return (returnValue);
}
/*********************************************************
* CheckKeyCase:
* If theEvent was a keyDown or autoKey event, this checks
* if both the shiftKey and the capsLock key were down. If
* so, it changes theEvent to be a lowercase letter. If not,
* nothing is changed. Also, if either the optionKey or
* cmdKey is down then nothing is changed.
********************************************************/
void CheckKeyCase(EventRecord *theEvent)
{
register long theMods, theMessage;
register char theChar;
if (theEvent->what == keyDown ||
theEvent->what == autoKey) {
theMods = theEvent->modifiers;
theMods &= shiftKey | alphaLock |
optionKey | cmdKey;
theMods ^= shiftKey | alphaLock;
if (!theMods) {
theMessage = theEvent->message;
theChar = theMessage & charCodeMask;
if (theChar >= 'A' && theChar <= 'Z') {
theMessage &= ~charCodeMask;
theMessage |= theChar + 'a' - 'A';
theEvent->message = theMessage;
}
}
}
}
/*********************************************************
* EndPatchCode:
* Dummy proc to mark the end of the code for the patches.
* Make sure all of your patch code is between here and
* StartPatchCode.
*********************************************************/
void EndPatchCode()
{
}