OBJECTIVES (1 of 2)
Download
Report
Transcript OBJECTIVES (1 of 2)
Windows Memory
Management, MemoryMapped Files, and DLLs
OBJECTIVES (1 of 2)
Upon completion of this Chapter you will be able to:
–2–
Describe the Windows memory management architecture and
the role of heaps and memory-mapped files
Use multiple independent heaps in applications requiring
dynamic memory management
Use Structured Exception Handling to respond to memory
allocation errors
Use memory-mapped files
OBJECTIVES (2 of 2)
–3–
Determine when to use the independent heaps and when to use
memory-mapped files and to describe the advantages and
disadvantages of each
Describe Windows dynamic link libraries (DLLs)
Describe the difference between static, implicit, and explicit
linking
Describe the advantages and disadvantages of each
Use DLLs to load different implementations of the same
function
OVERVIEW (1 of 2)
32-bit operating system, so pointers are 4-byte objects
Win64 provides 64-bit pointers
Processes have a private 4GB virtual address space
Half (2GB) is available to a process
Remainder allocated to shared data and code
Win64 enlarges VA space; required for many applications
Programs can create independent memory “heaps”
Processes can map files to memory
–4–
Processes can share memory through a mapped file
Fast and convenient for some file processing
OVERVIEW (2 of 2)
Dynamic Link Libraries with Monolithic Programs
Gather all the source code, including commonly used Chapters
such as utility functions
Put all the source code in a single project
Build, test, debug, and use the program
Inefficiency
Recompile same code in all projects
All executables include the same object code
Waste of disc space and physical memory at run
time
–5–
Maintenance complexity as shared code changes
Part I
Memory Management and Heaps
Memory Management Architecture
Windows Program
C library: malloc, free
Heap API: HeapCreate,
HeapDestroy,
HeapAlloc, HeapFree
MMF API:
CreateFileMapping,
CreateViewOfFile
Virtual Memory API
Windows Kernel with
Virtual Memory Manager
Physical
Memory
–7–
Disc &
File
System
HEAPS (1 of 2)
Pools of memory within the process virtual address
space
Every process has a default process heap
A process may have more than one heap. Benefits of
separate heaps include:
–8–
Fairness (between threads and between uses)
Allocation efficiency (fixed size blocks in each heap)
Deallocation efficiency (you can deallocate a complete data
structure with one call)
Locality of reference efficiency
HEAPS (2 of 2)
Every process has a process heap
Every heap has a handle
The programmer can use the process heap or create
new ones
HANDLE GetProcessHeap (VOID)
Return: The handle for the process’ heap; NULL on
failure
–9–
MEMORY MGT. IN MULTIPLE HEAPS
Virtual Address Space
Program
Not allocated
ProcHeap = GetProcessHeap ( );
pRoot = HeapAlloc (ProcHeap);
Process
Heap
Not allocated
Record
RecHeap
Record
Record
Not allocated
Node
NodeHeap
Node
Node
Not allocated
– 10 –
· · ·
RecHeap = HeapCreate ( );
NodeHeap = HeapCreate ( );
· · ·
while ( ) {
pRec = HeapAlloc (RecHeap);
pNode = HeapAlloc (NodeHeap);
· · ·
}
· · ·
HeapFree (RecHeap, 0, pRec);
HeapFree (NodeHeap, 0, pNode);
HeapDestroy (RecHeap);
HeapDestroy (NodeHeap);
HEAP MANAGEMENT (1 of 2)
HANDLE HeapCreate (DWORD flOptions,
DWORD dwInitialSize, DWORD
dwMaximumSize)
Return: A heap handle or NULL on failure
dwMaximumSize — How large the heap can become
0 — “growable heap”; no fixed limit
non-zero — “non-growable heap”
The entire block is allocated from the virtual
address space
But only the initial size is committed in the paging
file
– 11 –
HEAP MANAGEMENT (2 of 2)
flOptions is a combination of two flags:
HEAP_GENERATE_EXCEPTIONS
HEAP_NO_SERIALIZE
By generating exceptions, you can avoid explicit tests
after each heap management call
– 12 –
HEAPS
BOOL HeapDestroy (HANDLE hHeap)
hHeap — a heap generated using HeapCreate
Do not destroy the process’ heap (obtained using
GetProcessHeap)
Benefits of HeapDestroy:
No data structure traversal code
No need to deallocate each individual data
structure element, which can be time-consuming
– 13 –
MANAGING HEAP MEMORY (1 of 4)
LPVOID HeapAlloc (HANDLE hHeap, DWORD dwFlags,
DWORD dwBytes)
Return: A pointer to the allocated memory block (of size
dwBytes) or NULL on failure (unless exception
generation is specified)
hHeap — Handle from GetProcessHeap or HeapCreate
dwFlags — A combination of:
HEAP_GENERATE_EXCEPTIONS
HEAP_NO_SERIALIZE
HEAP_ZERO_MEMORY — Allocated memory initialized to zero
– 14 –
MANAGING HEAP MEMORY (2 of 4)
BOOL HeapFree (HANDLE hHeap, DWORD dwFlags,
LPVOID lpMem)
dwFlags — Should be zero (or HEAP_NO_SERIALIZE)
lpMem — Should have a value returned by HeapAlloc or
HeapReAlloc
hHeap — Should be the heap that lpMem was allocated
from
– 15 –
MANAGING HEAP MEMORY (3 of 4)
LPVOID HeapReAlloc (HANDLE hHeap, DWORD dwFlags,
LPVOID lpMem, DWORD dwBytes)
Return: Pointer to the reallocated block. Failure returns
NULL or causes exception.
dwFlags — Some essential control options:
HEAP_GENERATE_EXCEPTIONS and HEAP_NO_SERIALIZE
HEAP_ZERO_MEMORY — Only newly allocated memory is
initialized
HEAP_REALLOC_IN_PLACE_ONLY — Do not move the block
lpMem — Existing block in hHeap to be reallocated
dwByte — New block size
– 16 –
MANAGING HEAP MEMORY (4 of 4)
DWORD HeapSize (HANDLE hHeap, DWORD dwFlags,
LPVOID lpMem)
Return: The size of the block or zero on failure.
– 17 –
HEAP FLAGS (1 of 2)
HEAP_NO_SERIALIZE
Specified in HeapCreate, HeapAlloc, and other functions
Performance gain (about 15% in tests) as functions do not
provide mutual exclusion to threads accessing the heap
Can safely be used if (BUT, BE CAREFUL):
Your process uses only a single thread
Each thread has its own heap(s) that no other
– 18 –
thread can access
You provide your own mutual exclusion
mechanism to prevent concurrent access to a
heap by several threads
You use HeapLock and HeapUnlock
HEAP FLAGS (2 of 2)
HEAP_GENERATE_EXCEPTIONS
– 19 –
Allows you to avoid error tests after each allocation
OTHER HEAP FUNCTIONS
HeapValidate
Determine whether a heap has been corrupted
HeapCompact
Combine adjacent free blocks; decommit large free blocks
HeapWalk
– 20 –
Determine all blocks allocated within a heap
EXAMPLE-A
Write a program, sortHP, which reads fixed-size
records from a file into a memory-allocated buffer in a
heap, where the first 8 characters are a birth date
(CCYYMMDD format). The rest of the record is a line of
text.
– 21 –
Enter each date in an array, along with a file position. Each
array element will contain the date and the file position of the
record (which is not fixed length).
Sort the array using the C library qsort function.
Print out the complete file sorted by birth date.
Repeat the process for each file on the command line. Before
each new file, destroy the heaps from the preceding file.
EXAMPLE-A
The TestData directory contains two text files with 64byte records that can be used to test your program. Or,
use the RandFile program to generate sortable files of
any size.
– 22 –
EXAMPLE-A
Modify the sort program to create sortBT, which enters
the records in to a binary search tree and then scans
the tree to display the records in order.
– 23 –
Allocate the tree nodes and the data in separate heaps.
Destroy the heaps before sorting the next file, rather than
freeing individual tree nodes and data elements.
Test the program with and without heap serialization and
determine whether there is a detectable performance difference.
Part II
Memory-Mapped Files
MEMORY-MAPPED FILES
Advantages to mapping your virtual memory space
directly to normal files rather than the paging file:
– 25 –
You never need to perform direct file I/O
Data structures you create are saved in the file
You can use in-memory algorithms (string processing, sorts,
search trees) to process data even though the file may be much
larger than available physical memory
There is no need to manage buffers and the file data they
contain
Multiple processes can share memory (this is the only way), and
the file views will be coherent
There is no need to consume space in the paging file
PROCESS ADDRESS SPACE
MAPPED TO A FILE
Program
File
fH = CreateFile ( );
mH = CreateFileMapping (fH);
· · ·
while (
pRecA
pRecB
pRecB
Process
Address
Space
– 26 –
) {
= MapViewOfFile (mH);
= MapViewOfFile (mH);
-> Data = pRecA -> Data;
· · ·
UnmapViewOfFile (pRecA);
UnmapViewOfFile (pRecB);
}
CloseHandle (mH);
CloseHandle (fH);
FILE-MAPPING OBJECTS (1 of 4)
HANDLE CreateFileMapping (HANDLE hFile,
LPSECURITY_ATTRIBUTES lpsa,
DWORD dwProtect, DWORD dwMaximumSizeHigh,
DWORD dwMaximumSizeLow, LPCTSTR lpMapName)
Return: A file mapping handle or NULL
– 27 –
FILE-MAPPING OBJECTS (2 of 4)
Parameters
hFile — Open file handle; protection flags compatible with
dwProtect
LPSECURITY_ATTRIBUTES — NULL for now
dwProtect — How you can access the mapped file:
PAGE_READONLY
— Pages in the mapped region are
read only
— Full access if hFile has both
GENERIC_READ and GENERIC_WRITE access
PAGE_WRITECOPY — When you change mapped
memory, a copy is written to the paging file
PAGE_READWRITE
– 28 –
FILE-MAPPING OBJECTS (3 of 4)
– 29 –
dwMaximumSizeHigh and dwMaximumSizeLow — Specify the
size of the mapping object; 0 for current file size. The file is
extended if the current file size is smaller than the map size.
lpMapName — Names the mapping object, allowing other
processes to share the object
FILE-MAPPING OBJECTS (4 of 4)
You can also obtain a file-mapping handle by specifying
an existing mapping object name
HANDLE OpenFileMapping (DWORD dwDesiredAccess,
BOOL bInheritHandle, LPCTSTR lpNameP)
Return: A file mapping handle or NULL
CloseHandle destroys mapping handles
– 30 –
MAPPING PROCESS ADDRESS
SPACE (1 of 3)
LPVOID MapViewOfFile (HANDLE hMapObject,
DWORD dwAccess, DWORD dwOffsetHigh,
DWORD dwOffsetLow, DWORD cbMap)
Return: The starting address of the block (file view) or
NULL on failure
hMapObject — Identifies a file-mapping object
dwAccess — Must be compatible with mapping object’s
access:
– 31 –
FILE_MAP_WRITE
FILE_MAP_READ
FILE_MAP_ALL_ACCESS
MAPPING PROCESS ADDRESS
SPACE (2 of 3)
dwOffsetHigh and dwOffsetLow
Starting location of the mapped file region
Must be a multiple of 64K
Zero offset to map from beginning of file
cbMap — Size in bytes of the mapped region
– 32 –
Zero indicates entire file
Note: The map size is limited by the 32-bit address
MAPPING PROCESS ADDRESS
SPACE (3 of 3)
MapViewOfFileEx is similar, but you can specify an
existing address
BOOL UnmapViewOfFile (LPVOID lpBaseAddress)
– 33 –
To release file views
FILE-MAPPING LIMITATIONS
– 34 –
Disparity between Windows’s 64-bit file system and 32-bit
addressing
With a large file (greater than 4GB) you cannot map everything
into virtual memory space
Process data space is limited to 2GB
You cannot use all 2GB; available contiguous blocks will be
smaller
When dealing with large files, you must create code that
carefully maps and unmaps file regions as you need them
BASED POINTERS (1 of 2)
If you use pointers in a mapped file region, they should
be of type _based
– 35 –
A conventional pointer refers to the virtual address
This address base will almost certainly be different the next
time that file is mapped or a new view is created of the same
region
The pointer should be based on the view address
BASED POINTERS (2 of 2)
– 36 –
int *pi;
int __based(pi) *bpi, i;
...
pi = MapViewOfFile (...);
*pi = 3;
bpi = pi;
i = *bpi;
...
EXAMPLE-B
Rewrite the atou (ASCII to UNICODE) program to create
atouMM
– 37 –
Use memory mapping only; do not use ReadFile and
WriteFile
You do not need to change the main function in atou.c.
Instead, change the asc2un.c function to create asc2unMM.c.
EXAMPLE-B
Rewrite the sort program of the previous section to
create sortMM, so that key records (in the array) are
mapped to a “key” file
– 38 –
Do not use the file pointers; instead, use based pointers to
address in a view of the original file
As part of the test of _based pointers, have a program option to
simply use the saved key file to produce a sorted listing without
actually performing a sort. The next slide shows diagrams the
operation.
This is a difficult exercise!
sortMM OPERATION
sortMM
MyFile
Ki: Key
Si: String
Pi: Based Pointer
MyFile.idx
MyFile
K0
K0
S0
Ki
– 39 –
P0
K1
K1
Pi
P1
S1
Kj
K2
P2
K2
S2
···
Pj
Kk
Pk
···
qsort
···
Part III
Dynamic Link Libraries
STATIC LIBRARIES
Build one or more libraries as “static libraries”
Link the libraries with each project as needed
Advantages
Simplifies and expedites project building
Disadvantages
Disc and memory space issues
Maintenance requires relinking and redistribution
Different programs may use different library
versions
Programs cannot use alternate utility
implementations for different situations
– 41 –
DYNAMIC LINK LIBRARIES (1 of 4)
DLLs solve these and other problems very neatly
Library functions are linked at:
Program load time — implicit linking
Program run time — explicit linking
Program image can be much smaller
It does not include the library functions
Multiple programs can share a single DLL
Only a single copy will be loaded into memory
All programs map their process address space to
DLL code
Each thread will have its own copy of non-shared
storage on the stack
– 42 –
DYNAMIC LINK LIBRARIES (2 of 4)
New versions or alternate implementations:
Supplying a new version of the DLL
All programs can use the new version without modification
Explicit linking:
Program decides at run time which library version to use
Different libraries may be alternate implementations of the same
function
May carry out totally different tasks
Just as separate programs do
– 43 –
The library will run in the same process and thread as the
calling program
DYNAMIC LINK LIBRARIES (3 of 4)
DLLs are used in nearly every operating system
– 44 –
Including UNIX and Windows 3.1
Windows (all versions) uses DLLs to implement the OS
interfaces, among other things
Windows 3.1 DLLs run at the same address space for all
processes
Windows DLLs run in the process’ virtual address space
DYNAMIC LINK LIBRARIES (4 of 4)
Multiple Windows processes can share DLL code
Code, when called, runs as part of the calling process
and thread
Library can use the calling process’ resources (file handles, ...)
Uses the calling thread’s stack
DLLs must be thread-safe
DLLs can also export variables as well as function entry
points
– 45 –
IMPLICIT LINKING (1 of 2)
Implicit, or load-time, linking is the easiest of the two
techniques
Steps:
– 46 –
Collect and built function source as a DLL
Build process constructs a .LIB library file
“stub” for the actual code
Place .LIB in project library directory
Build process also constructs a .DLL file
IMPLICIT LINKING (2 of 2)
– 47 –
Contains the actual executable image
Placed in the same directory as the application that uses it
The current working directory is the secondary location
Then system directory, Windows directory, PATH
The program loads the DLL during its initialization
You must “export” the function interfaces in the DLL source
EXPORTING AND IMPORTING
INTERFACES (1 of 3)
DLL entry point must be declared
Microsoft C, using the _declspec (dllexport) storage
modifier:
_declspec (dllexport)
DWORD MyFunction (...);
Calling program declares the function is to be imported
– 48 –
Use the _declspec (dllimport) storage modifier
EXPORTING AND IMPORTING
INTERFACES (2 of 3)
Standard technique in include file
Use a preprocessor variable such as “MYPROJ_EXPORTS“
“MYPROJ” is the project name
#ifdef MYPROJ_EXPORTS
#define LIBSPEC _declspec (dllexport)
#else
#define LIBSPEC _declspec (dllimport)
#endif
LIBSPEC DWORD MyFunction (...);
– 49 –
EXPORTING AND IMPORTING
INTERFACES (3 of 3)
The DLL project defines MYPROJ_EXPORTS
Calling application leaves MYPROJ_EXPORTS undefined
You can export and import variables as well as function
entry points
– 50 –
EXAMPLE–C
Build one or more of the ASCII to Unicode functions
(asc2un) as a DLL. Export the entry point.
Chapter 2 version is straightforward file I/O. It can be made
faster with larger buffers, sequential scan flags, etc.
Chapter 4 version uses memory mapping
The two versions exhibit different performance characteristics
depending on the file system type (NTFS or FAT)
Rebuild the atou calling program (Chapter 2) so that it
implicitely links to a asc2un DLL
– 51 –
EXPLICIT LINKING (1 of 4)
Explicit (run-time) linking requires:
– 52 –
Program loads a DLL be loaded — LoadLibrary
Finds the address of the entry point(s) — GetProcAddress
Cast the address pointer to the function type
Call the function using the pointer
Optionally free the library — FreeLibrary
NOTE: The function is not declared in the calling program; you
declare a variable as a pointer to a function.
Therefore, there is no need for the .LIB file at link time
EXPLICIT LINKING (2 of 4)
HINSTANCE LoadLibrary (LPCTSTR lpLibFileName)
Returned handle is NULL on failure
HINSTANCE, rather than a conventional HANDLE
– 53 –
It contains different information
EXPLICIT LINKING (3 of 4)
BOOL FreeLibrary (HINSTANCE hLibModule)
You are done with the library or want a different version
LoadLibraryEx is similar
– 54 –
Several flags for specifying alternate search paths and loading
the library as a data file
EXPLICIT LINKING (4 of 4)
To obtain the entry point:
FARPROC GetProcAddress (HMODULE hModule,
LPCSTR lpProcName)
hModule is an instance produced by LoadLibrary
Or GetModuleHandle (not described here)
lpProcName is the entry point name
Cannot be Unicode
NULL in case of failure
FARPROC, like “long pointer,” is an anachronism
– 55 –
Loading a DLL
BOOL (*Asc2Un)(LPCTSTR, LPCTSTR, BOOL);
FARPROC pA2U;
/* Load the ASCII to Unicode function DLL file
*/
hDLL = LoadLibrary (argv [LocDLL]);
/* Get the entry point address */
pA2U = GetProcAddress (hDLL, "Asc2Un");
/* Convert to a function pointer */
Asc2Un = (BOOL (*)(LPCTSTR, LPCTSTR, BOOL))
pA2U;
/* Call the function */
Asc2Un (argv [LocFileIn], argv [LocFileOut],
FALSE);
– 56 –
EXAMPLE–D (1 of 2)
Modify the atou program so that the name of the DLL
file is on the command line
– 57 –
Then load the DLL and call it, as illustrated on the previous
slide
EXAMPLE–D (2 of 2)
Here are some suggested extensions, not included in
the lab solution:
– 58 –
Modify the calling program so that it sequentially loads and
calls several alternative implementations
Modify the calling program so that it determines the operating
environment (OS version, file system types, etc.) and then loads
the most efficient implementation for the environment. Use
your knowledge from previous performance experiments to
determine the best implementation for a given situation.