Writing in a file from multiple threads

问题:

I'm writing a download manager in Objective-C which downloads file from multiple segments at the same times in order to improve the speed. Each segement of the file is downloaded in a thread.

At first, I thought to write each segment in a different file and to put together all the files at the end of the download. But for many reasons, it's not a good solution.

So, I'm searching a way to write in a file at a specific position and which is able to handle multiple thread because in my application, each segment is downloaded inside a thread. In Java, I know that FileChannel does the trick perfectly but I have no idea in Objective-C.

回答1:

Never forget, Obj-C bases on normal C and thus I would just write an own class, that handles file I/O using standard C API, which allows you to place the current write position anywhere within a new file, even far beyond the current file size (missing bytes are filled with zero bytes), as well as jumping forward and backward as you wish. The easiest way to achieve thread-safety is using a lock, this is not necessary the fastest way but in your specific case, I bet that the bottleneck is certainly not thread-synchronization. The class could have a header like this:

@interface MultiThreadFileWriter : NSObject
{
    @private
        FILE * i_outputFile;
        NSLock * i_fileLock;
}
- (id)initWithOutputPath:(NSString *)aFilePath;
- (BOOL)writeBytes:(const void *)bytes ofLength:(size_t)length
    toFileOffset:(off_t)offset;
- (BOOL)writeData:(NSData *)data toFileOffset:(off_t)offset;
- (void)close;
@end

And an implementation similar to this one:

#import "MultiThreadFileWriter.h"

@implementation MultiThreadFileWriter

- (id)initWithOutputPath:(NSString *)aFilePath
{
    self = [super init];
    if (self) {
        i_fileLock = [[NSLock alloc] init];
        i_outputFile = fopen([aFilePath UTF8String], "w");
        if (!i_outputFile || !i_fileLock) {
            [self release];
            self = nil;
        }
    }
    return self;
}

- (void)dealloc
{
    [self close];
    [i_fileLock release];
    [super dealloc];
}

- (BOOL)writeBytes:(const void *)bytes ofLength:(size_t)length
    toFileOffset:(off_t)offset
{
    BOOL success;

    [i_fileLock lock];
    success = i_outputFile != NULL
        && fseeko(i_outputFile, offset, SEEK_SET) == 0
        && fwrite(bytes, length, 1, i_outputFile) == 1;
    [i_fileLock unlock];
    return success;
}

- (BOOL)writeData:(NSData *)data toFileOffset:(off_t)offset
{
    return [self writeBytes:[data bytes] ofLength:[data length]
        toFileOffset:offset
    ];
}

- (void)close
{
    [i_fileLock lock];
    if (i_outputFile) {
        fclose(i_outputFile);
        i_outputFile = NULL;
    }
    [i_fileLock unlock];
}
@end

The lock could be avoided in various way. Using Grand Central Dispatch and Blocks to schedule the seek + write operations on a Serial Queue would work. Another way would be to use UNIX (POSIX) file handlers instead of standard C ones (open() and int instead of FILE * and fopen()), duplicate the handler multiple times (dup() function) and then placing each of them to a different file offset, which avoids further seeking operations on each write and also locking, since POSIX I/O is thread-safe. However, both implementations would be somewhat more complicating, less portable and there would be no measurable speed improvement.

回答2:

The answers given thus far have some clear disadvantages:

• File i/o using system calls definitely has some disadvantages regarding locking.
• Caching parts in memory leads to serious issues in a memory constrained environment. (i.e. any computer)

A thread safe, efficient, lock free approach would be to use memory mapping, which works as follows:

• create the result file of (at least) the total length needed
• open() the file for read/write
• mmap() it to some place in memory. The file now "lives" in memory.
• write received parts in memory at the right offset in the file
• keep track if all pieces have been received (e.g. by posting some selector on the main thread for every piece received and stored)
• munmap() the memory and close() the file

The actual writing is handled by the kernel - your program will never issue a write system call of any form. Memory mapping generally has little downsides and is used extensively for things like shared libraries.

update: a piece of code says more than 1000 words... This is the mmap version of Mecki's lock-based multi-thread file writer. Note that writing is reduced to a simple memcpy, which cannot fail(!!), so there is no BOOL success to check. Performance is equivalent to the lock based version. (tested by writing 100 1mb blocks in parallel)

Regarding a comment on "overkill" of an mmap based approach: this uses less lines of code, doesn't require locking, is less likely to block on writing, requires no checking of return values on writing. The only "overkill" would be that it requires the developer to understand another concept than good old read/write file I/O.

The possibility to read directly into the mmapped memory region is left out, but is quite simple to implement. You can just read(fd,i_filedata+offset,length); or recv(socket,i_filedata+offset,length,flags); directly into the file.

@interface MultiThreadFileWriterMMap : NSObject
{
@private
    FILE * i_outputFile;
    NSUInteger i_length;
    unsigned char *i_filedata;
}

- (id)initWithOutputPath:(NSString *)aFilePath length:(NSUInteger)length;
- (void)writeBytes:(const void *)bytes ofLength:(size_t)length
      toFileOffset:(off_t)offset;
- (void)writeData:(NSData *)data toFileOffset:(off_t)offset;
- (void)close;
@end

#import "MultiThreadFileWriterMMap.h"
#import <sys/mman.h>
#import <sys/types.h>

@implementation MultiThreadFileWriterMMap

- (id)initWithOutputPath:(NSString *)aFilePath length:(NSUInteger)length
{
    self = [super init];
    if (self) {
        i_outputFile = fopen([aFilePath UTF8String], "w+");
        i_length = length;
        if ( i_outputFile ) {
            ftruncate(fileno(i_outputFile), i_length);
            i_filedata = mmap(NULL,i_length,PROT_WRITE,MAP_SHARED,fileno(i_outputFile),0);
            if ( i_filedata == MAP_FAILED ) perror("mmap");
        }
        if ( !i_outputFile || i_filedata==MAP_FAILED ) {
            [self release];
            self = nil;
        }
    }
    return self;
}

- (void)dealloc
{
    [self close];
    [super dealloc];
}

- (void)writeBytes:(const void *)bytes ofLength:(size_t)length
      toFileOffset:(off_t)offset
{
    memcpy(i_filedata+offset,bytes,length);
}

- (void)writeData:(NSData *)data toFileOffset:(off_t)offset
{
    memcpy(i_filedata+offset,[data bytes],[data length]);
}

- (void)close
{
    munmap(i_filedata,i_length);
    i_filedata = NULL;
    fclose(i_outputFile);
    i_outputFile = NULL;
}

@end

回答3:

Queue up the segment-objects as they are received to a writer-thread. The writer-thread should keep a list of out-of-order objects so that the actual disk-writing is sequential. If a segment download fails, it can be pushed back onto the downloading thread pool for another try, (perhaps an internal retry-count should be kept). I suggest a pool of segment-objects to prevent one or more failed download of one segment resulting in runaway memory use as later segments are downloaded and added to the list.