The Secret Life Of The Memory Manager
The Secret Life Of The Memory Manager
RICHARD CLARK
The Macintosh Memory Manager has changed in some subtle ways since it was
documented in Inside Macintosh . This, combined with the difficulty of observing
what the Memory Manager actually does, has led to a general misunderstanding of
how the Memory Manager works. This article first discusses some common myths
about the Memory Manager, then describes some ways to avoid memory-related
errors and control fragmentation without sacrificing execution speed.
Few parts of the Macintosh operating system raise as many questions as the Memory Manager. Since
the contents of RAM change dynamically, it's hard to really examine the Memory Manager's
behavior. This, combined with the unusual concept of relocatable blocks and the fact that the
Memory Manager is used by most of the operating system, has left many Macintosh programmers
confused about the behavior of the Memory Manager and, more important, about the impact of this
behavior on their applications.
MYTHS ABOUT THE MEMORY MANAGER
Several myths have grown up around the Memory Manager, serving to increase the confusion about
its real behavior. Three of the most prevalent--but mistaken-- beliefs are that (1) the Memory
Manager will move and delete blocks, and otherwise mangle the heap, at random; (2) using
nonrelocatable blocks will cause serious heap fragmentation; and (3) if you use Virtual Memory you
don't need to worry about the Memory Manager. We'll demolish each of these myths in turn.
MYTH 1: THE MEMORY MANAGER WILL MOVE AND DELETE BLOCKS,
AND OTHERWISE MANGLE THE HEAP, AT RANDOM
This simply isn't so. The Memory Manager is in fact quite predictable. It only moves blocks under
these circumstances:
- When your application calls a routine that allocates new blocks
or enlarges existing ones, when you request that blocks be moved, or when your
application calls a routine that in turn calls a ROM routine that may trigger block
relocation. Appendix A of the Inside Macintosh XRef lists all routines
defined in Inside Macintosh that may cause blocks to move.
- When the called routine is in a different segment from the code that makes the
call, or when the called routine is in the same
segment as the caller, but the called routine calls a routine or
routines in a different segment. If a called routine lies in a
different code segment, the Segment Loader may need to call
the code segment in from disk and/or move it to the top of the heap. Either of
these actions can cause blocks to move.
MYTH 2: USING NONRELOCATABLE BLOCKS WILL CAUSE
SERIOUS MEMORY FRAGMENTATION
This is a half-truth at best. The Memory Manager actually does a good job of allocating nonrelocat-
able blocks, but can fragment the heap when these blocks are deallocated and new ones allocated.
Similar problems can happen when you start locking relocatable blocks.
This myth actually has a basis in reality, as the earliest versions of the Memory Manager did a poor
job of allocating nonrelocatable blocks. Before the 128K ROMs (introduced with the Macintosh
512Ke and Macintosh Plus), the Memory Manager would not move a relocatable block around a
nonrelocatable block in its quest to allocate a new nonrelocatable block. This made the heap into a
patchwork of relocatable and nonrelocatable blocks, and caused general fragmentation problems, as
illustrated in Figure 1.
Figure 1. Fragmentation of Free Space
But that has long since changed, as NewPtr
will now move a
relocatable block around a nonrelocatable block when allocating
memory. This tends to partition the heap into two active areas, with
all of the nonrelocatable blocks at the bottom of the heap, and the
relocatable blocks located immediately above. (See the sidebar "How
the Memory Manager Allocates Heap Space" for further details.)
On the other hand, for all of the improvements in allocation of nonrelocatable blocks, there is still a
problem withde allocation of these blocks. Since the Memory Manager uses a "find the first free block
that fulfills the request" strategy (as opposed to "find a block that fits the request exactly"), if you
allocate a subsequent block that is smaller than the block you just deleted, the heap will become
fragmented and the amount of usable memory will likely decrease, as illustrated in Figure 3.
Figure 3. The Effect of Deallocating and Reallocating a Nonrelocatable Block
Locking too many relocatable blocks can cause the same kind of fragmentation problems as
deallocating and reallocating nonrelocatable blocks. A well-trained programmer uses the callMoveHHi
to move a relocatable block to the top of the heap before locking it. This has the effect of partitioning
the heap into four areas, as shown in Figure 4. The idea of usingMoveHHi
is to keep the contiguous
free space as large as possible. However,MoveHHi
will only move a block upward until it meets either a
nonrelocatable block or a locked relocatable block. UnlikeNewPtr
(andResrvMem
),MoveHHi
will not
move a relocatable block around one that is not relocatable.
Even if you succeed in moving a relocatable block to the top of the heap, your problems are far from
over. Unlocking or deleting locked blocks can also cause fragmentation, unless they are unlocked
beginning with the lowest locked block. In the case illustrated in Figure 4, unlocking and deleting
blocks in the middle of the locked area has resulted in heap fragmentation. The relocatable blocks
thus trapped in the middle won't be moved until the locked block below them is unlocked.
Figure 4. The Effect of Unlocking Locked Blocks
MYTH 3: IF YOU USE VIRTUAL MEMORY, YOU DON'T NEED TO WORRY ABOUT THE MEMORY MANAGER
Many people believe that the wide availability of Virtual Memory will remove the need for careful
memory management. Wrong! The Virtual Memory system is based on a series of "pages" of
memory that can be swapped to and from the disk, rather than on individual blocks of memory. If you
fragment RAM, you also "fragment" the contents of the swap file and gain nothing. In fact, Virtual
Memory makes careful memory management even more critical, for two reasons. First, fragmenting
the swap file will degrade system performance worse than fragmenting physical memory will, since
disk access speeds are obviously slower than the RAM access speed. Second, the combination of
Virtual Memory and MultiFinder encourages users to run more programs at the same time than they
used to, and users often reduce the partition sizes of their applications to squeeze in "one more
program."
THE EXPERT'S GUIDE TO MEMORY MANAGEMENT
Now you know that the Memory Manager moves blocks of memory only at certain well-defined
times; that nonrelocatable blocks can be allocated without causing serious fragmentation in the heap,
although deallocation and reallocation of these blocks, and locking too many relocatable blocks, can
cause problems; and that use of Virtual Memory makes careful memory management even more
important. It's time to put this knowledge into action. In this section, you'll learn how you can work
cooperatively with the Memory Manager to increase the efficiency and robustness of your
applications.
TO AVOID DANGLING POINTERS
As every programmer learns early on, the gravest side effect of the Memory Manager's penchant for
moving blocks of memory is the peril of dangling pointers. (For a refresher on how these come
about, see the sidebar entitled "A Primer on Handles and Their Pitfalls" in Curt Bianchi's article
"Using Objects Safely in Object Pascal" in this issue.) And the best defense against having to spend
hours--or days--debugging errors caused by dangling pointers is to anticipate situations in which
block movement might occur, and if it does occur, will throw a monkey wrench into the works. In
these situations, much grief can be saved by using a temporary local or global variable to store a
duplicate of the relocatable block. (Note, though, that this trick only works properly if the block can
stand on its own--that is, it's not part of a linked list.)
Some of the situations that might get you into trouble are well documented, such as the use of the
WITH statement in Pascal. Other dangerous situations are less obvious, so we'll explore them here.
Be careful when evaluating expressions. There are times when
evaluating a seemingly innocent expression might have serious side
effects. For example, look at the following code:
TYPE
windowInfoHdl = ^windowInfoPtr;
windowInfoPtr = ^windowInfo;
windowInfo = RECORD
aControlHdl: ControlHandle;
aWindowPtr: WindowPtr;
END;
VAR
myHandle : windowInfoHdl;
BEGIN
myHandle := windowInfoHdl(NewHandle(sizeof(windowInfo)));
{ The next 2 statements have problems. }
myHandle^^.aWindowPtr := GetNewWindow(1000, NIL, WindowPtr(-1));
myHandle^^.aControlHdl :=
GetNewControl(1000, myHandle^^.aWindowPtr);
END;
In Pascal, the above statements would probably cause a run-time
error. The problem is in the expression " myHandle^^.something :=
"
as the compiler evaluates expressions from left to right and calculates
the address on the left side of the assignment statement before making
the toolbox call. When GetNewWindow
is called, myHandle^^
is moved
(we passed in NII
to force a call to NewPtr
) and the address on the left-
hand side is no longer valid! This means that the returned WindowPtr
will be written into the wrong area of memory, and the program will probably crash.
While both statements suffer from the same basic problem, the first
one is more likely to cause a crash than the second one and is therefore
easier to debug. Why is this?
The statement containing GetNewWindow
will make a call to NewPtr
to
allocate a nonrelocatable block at the bottom of the heap, forcing
relocatable blocks upward in the process. The other statement,
containing GetNewControl
, allocates a relocatable block, which
usually appears above the existing blocks, with block movement
happening only if a compaction is required.
While this problem occurs most frequently in Pascal, C programs are
not immune. Most C compilers on the Macintosh evaluate the right-
hand side of an assignment before the left-hand side--which avoids
this problem entirely--but the order of evaluation is not guaranteed by
the ANSI standard.
This problem can be solved easily by using a temporary variable. The
following code avoids the problem:
VAR
myHandle: windowInfoHdl;
aWindowPtr: WindowPtr; { This is allocated on the }
{ stack, so it won't move. }
aControlHandle: ControlHandle; { Also on the stack. }
BEGIN
myHandle := windowInfoHdl(NewHandle(sizeof(windowInfo)));
{ Copy the result into a temporary variable, then copy }
{ that into the relocatable block. }
aWindowPtr := GetNewWindow(1000, NIL, WindowPtr(-1));
myHandle^^.aWindowPtr := aWindowPtr;
aControlHandle := GetNewControl(1000, aWindowPtr);
myHandle^^.aControlHdl := aControlHandle;
END;
Be careful when using callback routines. When you pass pointers to your
routines, say as a ROM callback routine, and your routines are in
multiple segments, you need to be careful.
The following code is fine now, but we'll soon edit it to demonstrate
the problem:
{$S Main }
PROCEDURE MyCallback(ctl: ControlHandle; part: INTEGER);
{ This represents a callback routine used for continuous }
{ tracking in controls. }
BEGIN
{ Do whatever you need to do. }
END;
PROCEDURE HandleMyControl(theControl: ControlHandle;
pt: Point);
BEGIN
part := TrackControl(theControl, pt, @MyCallback);
END;
The expression @MyCallback
pushes the address of the callback
routine onto the stack before calling TrackControl
. If the two
routines are in the same segment, as in the preceding example, all is
fine. The segment is locked in memory when @MyCallback
is both
evaluated and used; therefore, the address is valid. If the two routines
are in different segments, this also works, as the compiler takes the
address of the jump table entry for MyCallback
.
In some cases, and especially in C, you may choose to set up a table
of procedure addresses. But if you store the address of the routine
into a variable, strange things may happen. Take a look at the
following code:
{ ----------------------------- }
{ For an example, we'll place the addresses of two control }
{ tracking routines into an array, then use them. }
VAR
gCallbackArray: ARRAY [1..2] OF ProcPtr;
{ ----------------------------- }
{$S Segment1 }
PROCEDURE MyVScrollCallback(theControl: ControlHandle;
part: INTEGER);
BEGIN
{ This will get called if our control is a vertical }
{ scrollbar. }
END;
PROCEDURE MyVScrollCallback(theControl: ControlHandle;
part: INTEGER);
BEGIN
{ This will get called if our control is a horizontal }
{ scrollbar. }
END;
PROCEDURE InitCallbackArray;
{ Fill in the addresses in the global "Callback" array. }
BEGIN
{ Problem: Since we're in the same segment, these aren't }
{ addresses of the jump table entries, but are absolute }
{ locations in RAM! If the segment moves (i.e., if }
{ UnloadSeg is called), the addresses will be invalid. }
gCallbackArray[1] := @MyVScrollCallback;
gCallbackArray[2] := @MyHScrollCallback;
END;
END.
{ ----------------------------- }
{$S Main }
PROCEDURE HandleAScrollbar(theControl: ControlHandle;
pt: Point);
{ We'll call this if the user clicks in our scrollbar (except }
{ if she clicks in the thumb, which uses a different kind of }
{ callback.) If it's a vertical scrollbar, use one callback; }
{ if horizontal, use the other. }
VAR
part: INTEGER;
theCallback: ProcPtr;
isVertical: Boolean;
aRect: Rect;
cntlWidth: INTEGER;
BEGIN
aRect := theControl^^.cntrlRect;
cntlWidth := aRect.right - aRect.left;
isVertical := cntlWidth = 16;
IF isVertical THEN
part := TrackControl(theControl, pt, gCallbackArray[1])
ELSE
part := TrackControl(theControl, pt, gCallbackArray[2])
{ The TrackControl calls will probably crash if }
{ Segment1 has been unloaded since the table was built. }
{ You'll have a wonderful time trying to find the bug! }
END;
When setting up a table of such procedure addresses, or even a single
global variable, you should do one of the following things: (1) make
sure that the setup procedure is in a different segment from the
procedures being called, thus insuring that you get the address of a
jump table entry; (2) keep everything in one segment and never unload
it; or (3) always load the segment and build the table before using any
of the addresses (and make sure that the segment doesn't get
unloaded in the meantime).
Be careful when passing parameters. Another problem area occurs when
you pass parameters to routines that allocate or move memory. Can
you spot the problem in the following code?
PROCEDURE ValidateControl(theControl: ControlHandle);
BEGIN
ValidRect(theControl^^.contrlRect);
END;
ValidRect
receives the address of a rectangle, which is pushed onto
the stack before the trap is called. The problem is that beforeValidRect
uses the rectangle's address, it often allocates memory of
its own, which can cause theControl^^
to move and therefore
invalidate the rectangle's address.
This problem happens when you pass (1) any parameter larger than
four bytes, or
(2) any VAR
parameter. Again, the solution requires a temporary
variable:
PROCEDURE ValidateControl(theControl: ControlHandle);
VAR
r : Rect; { r is stack-based, so it doesn't move. }
BEGIN
r := theControl^^.contrlRect;
ValidRect(r);
END;
Pascal compilers often avoid this problem for user-defined functions by making a local copy of
non-VAR
parameters that are passed by address. The ROM doesn't make such a copy, so you need to be
careful. This is discussed at length by Scott Knaster in How to Write Macintosh Software , 2nd ed.
(Hayden Books, 1988).
TO CONTROL HEAP FRAGMENTATION
As you will recall, heap fragmentation can be caused by (1) deallocating and reallocating
nonrelocatable blocks, and (2) locking too many relocatable blocks.
To keep heap fragmentation under control, follow a few simple rules.
Use nonrelocatable blocks sparingly. To avoid the potential problems that deallocation and reallocation
of nonrelocatable blocks can cause, you should theoretically use relocatable blocks for everything.
However, in practice, there are areas where you must use nonrelocatable blocks, such as forGrafPorts
and WindowRecords
. In light of this reality, here are three suggestions to help you control
fragmentation.
First, remember that you should not choose to use nonrelocatable blocks lightly. Use them only
when the Macintosh operating system requires them, or when you can demonstrate a severe
performance penalty for using relocatable blocks.
Second, avoid allocating nonrelocatable blocks unless they will never be deleted. If you know about
such blocks ahead of time, then you can allocate them at program start-up. This works well if you'll
have a single large "image buffer" or the like, or a limit on the number of available windows. In
these cases, allocating your large fixed blocks at start-up time will avoid potential fragmentation
problems.
Third, if you must allocate and deallocate nonrelocatable blocks on demand, you can add some
additional memory management code of your own. When you want to deallocate a block of RAM,
you can add it to a linked list of free blocks (that you maintain), and then check this list for a free
block of the exact size you need each time you want to allocate a new block. Of course, this works
best if the range of block sizes you support is limited, and you still have to decide what to do if the
block you want doesn't fit any of the free blocks exactly. If you have to allocate a large number of
nonrelocatable blocks, or have other special needs, you should consider allocating a large block of
memory and doing your own memory management within that. Donald Knuth's bookThe Art of
Computer Programming, volume 2, 2nd ed. (Addison-Wesley, 1973) contains a useful overview of
memory management techniques under the heading "Dynamic Storage Allocation" (pp. 435-55 and
460-61).
Note that this strategy of reusing nonrelocatable blocks works best under the 128K ROM (and later)
Memory Manager, since that version does the best job of allocating nonrelocatable blocks. If you
plan to write software under the 64K ROMs (Macintosh 128K or 512K), you should consult Scott
Knaster's How to Write Macintosh Software , which describes a strategy that does a better job with the
old Memory Manager than this strategy does.
Lock selectively and consider alternatives. Fear of dangling pointers often drives new programmers to
lock down everything in sight, quickly fragmenting the heap and impeding the application's
performance. More experienced programmers try to avoid locking relocatable blocks, preferring
instead to predict when the Memory Manager will move blocks of memory and then only locking a
relocatable block if they must. If done infrequently, locking has a negligible impact on your
application.
If you must lock a relocatable block, you should unlock it as soon as possible.
This will lessen the probability of another block being moved in underneath (by MoveHHi
) and
locked. Also, if you move and lock several blocks together, you should unlock all of them together,
or at least in the reverse of the order in which they were moved high. This will help ensure that the
free area is kept together in the heap.
As an alternative to locking relocatable blocks, consider using temporary variables. We've already
seen the use of temporary variables for such small items as window pointers and rectangles, but this
approach can also be used for entire structures. Using temporary variables can simplify your code by
removing the need for HLock
and HUnlock
calls. For example, many programs use a window's
reference constant (RefCon
) field to hold a handle to a data structure. Programs that do so look
something like this:
TYPE
windowInfoHdl = ^windowInfoPtr;
windowInfoPtr = ^windowInfo;
windowInfo = RECORD
rectArray: ARRAY [1..10] OF Rect;
END;
PROCEDURE UpdateWindow(wp: WindowPtr);
{ The window's RefCon contains a handle to the data structure shown }
{ above. The rectArray field contains an array of rectangles that }
{ we want to draw. }
VAR
myHandle: windowInfoHdl;
count: INTEGER;
BEGIN
{ Get the window information, then lock the block so that it }
{ doesn't move while drawing the rectangles. }
myHandle := windowInfoHdl(GetWRefCon(wp));
MoveHHi(Handle(myHandle));
HLock(Handle(myHandle));
BeginUpdate(wp);
FOR count := 1 TO 10 DO
{ Working with the heap-based window information, draw each }
{ rectangle. }
FrameRect(myHandle^^.rectArray[count]);
EndUpdate(wp);
HUnlock(Handle(myHandle));
END;
Notice that we had to perform several type casts, and use MoveHHi
,HLock
, and HUnlock
. Now, let's see how this would look using a
temporary variable:
{ Type declarations omitted for brevity. }
PROCEDURE GetWindowInfo (wp: WindowPtr; VAR info: windowInfo);
{ Utility routine to make a copy (usually stack-based) of our }
{ window information structure. }
VAR
myHandle: windowInfoHdl;
BEGIN
{ First, do a little error checking. }
IF (wp <> NIL) THEN BEGIN
myHandle := windowInfoHdl(GetWRefCon(wp));
{ You can incorporate extra error checking here. For example, }
{ this is a good place to compare the handle's size to the }
{ window information structure's size, or to verify that the }
{ contents of the block are legal values. }
{ }
{ Next, go ahead and copy the contents of the relocatable }
{ block to the specified location. We don't have to lock }
{ things down, since BlockMove won't cause compaction. }
BlockMove(Ptr(myHandle^), @info, sizeof(windowInfo));
END;
END;
PROCEDURE UpdateWindow(wp: WindowPtr);
VAR
info: windowInfo; { This storage is on the stack, therefore it }
{ won't move. }
BEGIN
GetWindowInfo(wp, info); { Get a copy of the window information. }
BeginUpdate(wp);
FOR count := 1 TO 10 DO
{ Working with the stack-based copy of the window information, }
{ draw each rectangle. }
FrameRect(info.rectArray[count]);
EndUpdate(wp);
END;
This approach has two major advantages: safety and code simplification. If you have one central
routine that gets the window information (and another similar one to set it), you can add quite a bit
of error checking and catch a large number of potential errors. Speed shouldn't be a problem, as the
single BlockMove
operation is generally faster than the corresponding MoveHHi
since the latter may
need to move an old relocatable block out of the way first.
Of course, you have to beextremely careful when using this technique, as it is easy to exceed the stack
size limit when using recursive or heavily nested procedures. If you have a series of nested procedures
that all use the window information structure, you can get the structure in the topmost procedure and
pass the block down as aVAR
parameter (pass-by-address in C) so that an extra copy of the data
structure isn't made.
FINAL WORDS OF ADVICE
In this article, we've taken a quick look inside the Memory Manager, but we have not been able to
cover everything. If you want to have a fuller understanding of Macintosh memory management,
there are a few things you can do. First, reread chapter 3 of
Inside Macintosh , volume I, and chapter 1
of
Inside Macintosh , volume II. Next, take a look at Scott Knaster's
How to Write Macintosh Software ,
mentioned earlier, which has an excellent discussion of memory management. (In fact, I recommend
the book highly to anybody who wants a better understanding of developing and debugging
Macintosh software.) Finally, examine the Memory Manager's behavior in real-life situations.develop, the disc, contains the source and object code for the Heap Demo application, which sets
up a small heap independent of the main application heap and allows you to manipulate handles and
pointers in that environment. If you do these things, you'll be well on the way to mastering the
Memory Manager.
HOW THE MEMORY MANAGER ALLOCATES HEAP SPACE
The Memory Manager uses two basic techniques to create space for blocks on the heap: compaction and
reservation. It uses compaction to create space for new relocatable blocks, and reservation to create space
for new nonrelocatable blocks.
When your application (or the operating system) calls NewHandle
to allocate a new relocatable block, the
Memory Manager first looks for a large enough space to hold a block of the requested size. If a large
enough space is found (and it need not be a perfect fit), the block is allocated. If there is not enough free
space to satisfy the request, compaction takes place--relocatable blocks are moved downward (toward low
memory) to make space for the new block. As a rule, the Memory Manager allocates new relocatable blocks as low in the heap as possible without
compaction. If the heap must be compacted, the Memory Manager begins with the lowest blocks and
gradually works its way upward until it has created a large enough free space to accommodate the new
relocatable block or until the entire heap has been compacted.
On the other hand, when your application (or the operating system) calls NewPtr
to allocate a new
nonrelocatable block, the Memory Manager calls ResrvMem
to create an empty space at the bottom of the
heap for the new nonrelocatable block. This technique is
known as reservation (after the call), although you won't
find this term anywhere in Inside Macintosh.
The Memory Manager always allocates nonrelocatable blocks as low as possible on the heap, even if it
means that other blocks have to be moved. In the case shown in Figure 2, the Memory Manager has to
move a relocatable block twice when the user allocates two nonrelocatable blocks. Note that each time the
4KB relocatable block is moved, it leaves a 4KB space behind. This is a result of the way the Memory
Manager reserves memory. It first moves the block upward into the first free area above its former position
large enough to hold it, then uses the old space for the new block.
Figure 2. The Effect of Allocating Nonrelocatable Blocks
In summary, allocating a new nonrelocatable block is likely to move other (relocatable) blocks upward, while
allocating a new relocatable block may cause compaction, which moves relocatable blocks downward.
ABOUT LOCAL AND GLOBAL VARIABLES
While in this article we're primarily interested in information stored on the heap, there are actually three
places you can store information in memory: in a relocatable or nonrelocatable block on the heap, in a local
variable, or in a global variable. In terms of storage efficiency, relocatable blocks are your best bet. But if
you need to store information in an area that will not move, you can use local or global variables.
Local variables are allocated on the machine's stack, and only exist as long as the enclosing procedure is
running. Global variables are stored in a special block above the top of the application's stack and heap, and
exist as long
as the program is running. Both of these areas share one disadvantage: limited space. You can only allocate
32KB of global variables, and the maximum available stack space typically varies between 8KB and 24KB,
depending on the machine, the operating system version, and whether or not the application has requested a
larger stack when launched.
RICHARD "TIGGER" CLARK wears brightly colored clothes, writes odd graffiti, tells horrible puns, and is amazingly graceful
when running for the bus. He earned a BS in social science (which he says is a hybrid psychology/computer science
degree) from the University of California-Irvine in 1985. When he's not teaching at Developer U, you can find him
stunt-kite flying (sometimes indoors), mountain climbing (sometimes indoors), or collecting Disney memorabilia. An avid reader,
he has totally worn out his copy of Winnie the
Pooh in Latin. In his time he has been a Valley Boy, a Macintosh repairman, a software developer,
King Henry VIII's head steward, a Renaissance bishop, and probably a few other things he won't tell us about. But hey, he's from southern
California. *