open-source-search-engine/BigFile.h

346 lines
11 KiB
C
Raw Normal View History

2013-08-03 00:12:24 +04:00
// Copyright Sep 2000 Matt Wells
// . this class makes the 2gb file limit transparent
// . you can actually set you own file size limit to whatever you want as
2014-11-11 01:45:11 +03:00
// int32_t as it's greater than zero
2013-08-03 00:12:24 +04:00
// . this is usually 2gb
// . TODO: fix the O_SYNC option
// . TODO: provide ability to pass flags
#ifndef _BIGFILE_H_
#define _BIGFILE_H_
#include "File.h"
ssize_t gbpwrite(int fd, const void *buf, size_t count, off_t offset);
// . use 512 disk megs per part file
// . at a GB_PAGE_SIZE of 16k this is 32k RdbMap pages (512k) per part file
// . let's use 2GB now to conserve file descriptors
//#define MAX_PART_SIZE (2000LL*1000LL*1000LL-1414144LL)
#define MAX_PART_SIZE (1920LL*1024LL*1024LL)
// debug define
//#define MAX_PART_SIZE (32LL*1024LL*1024LL)
// have enough part files to do a 2048gig file
#define MAX_PART_FILES (((2048LL*1000LL*1000LL*1000LL)/MAX_PART_SIZE)+1LL)
// debug define
//#define MAX_PART_FILES 100
// use this state class for doing non-blocking reads/writes
2014-06-08 01:11:12 +04:00
#ifdef ASYNCIO
2013-08-03 00:12:24 +04:00
#include <aio.h> // TODO: use kaio, uses only 4 threads
2014-06-08 01:11:12 +04:00
#endif
2013-08-03 00:12:24 +04:00
class FileState {
public:
// this is where we go after the thread has exited
//void (*m_threadDone) ( void *state ) ;
// callback must be top 4 bytes of the state class we give to g_loop
// callback must be first X bytes
class BigFile *m_this;
//struct aiocb m_aiostate;
char *m_buf;
2014-11-11 01:45:11 +03:00
int32_t m_bytesToGo;
2014-10-30 22:36:39 +03:00
int64_t m_offset;
2013-08-03 00:12:24 +04:00
// . the original offset, because we set m_offset to m_currentOffset
// if the original offset specified is -1
// . we also advance BigFile::m_currentOffset when done w/ read/write
2014-10-30 22:36:39 +03:00
//int64_t m_origOffset;
2013-08-03 00:12:24 +04:00
bool m_doWrite;
2014-11-11 01:45:11 +03:00
int32_t m_bytesDone;
2013-08-03 00:12:24 +04:00
void *m_state ;
void (*m_callback) ( void *state ) ;
// goes from 0 to 1, the lower the niceness, the higher the priority
2014-11-11 01:45:11 +03:00
int32_t m_niceness;
2013-08-03 00:12:24 +04:00
// . if signal is still pending we need to know if BigFile got deleted
// . m_files must be NULL terminated
//class BigFile **m_files;
// . we get our fds before starting the read thread to avoid
// problems with accessing m_files since RdbMerge may call unlinkPart
// from the main thread while we're trying to get these things
// . no read should span more than 2 file descriptors
2014-11-11 01:45:11 +03:00
int32_t m_filenum1;
int32_t m_filenum2;
2013-08-03 00:12:24 +04:00
int m_fd1 ;
int m_fd2 ;
// hold the errno from the threaded read/write here
2014-11-11 01:45:11 +03:00
int32_t m_errno;
2013-08-03 00:12:24 +04:00
// just for flagging unlinked/renamed thread ops
2014-11-11 01:45:11 +03:00
int32_t m_errno2;
2013-08-03 00:12:24 +04:00
// when we started for graphing purposes (in milliseconds)
2014-10-30 22:36:39 +03:00
int64_t m_startTime;
int64_t m_doneTime;
2013-08-03 00:12:24 +04:00
// this is used for calling DiskPageCache::addPages() when done
// with the read/write
class DiskPageCache *m_pc;
// this is just used for accessing the DiskPageCache, m_pc, it is
// a "virtual fd" for this whole file
2014-11-11 01:45:11 +03:00
int32_t m_vfd;
2013-08-03 00:12:24 +04:00
// test parms
2014-11-11 01:45:11 +03:00
//int32_t m_osize;
2013-08-03 00:12:24 +04:00
//char *m_obuf;
// for avoiding unlink/reopens while doing a threaded read
2014-11-11 01:45:11 +03:00
int32_t m_closeCount1 ;
int32_t m_closeCount2 ;
int32_t m_vfd1;
int32_t m_vfd2;
2013-08-03 00:12:24 +04:00
//char m_baseFilename[32];
2014-11-11 01:45:11 +03:00
int32_t m_flags;
2013-08-03 00:12:24 +04:00
// when we are given a NULL buffer to read into we must allocate
// it either in DiskPageCache.cpp or in Threads.cpp right before the
// thread is launched. this will stop us from having 19000 unlaunched
// threads each hogging up 32KB of memory waiting to read tfndb.
// m_allocBuf points to what we allocated.
char *m_allocBuf;
2014-11-11 01:45:11 +03:00
int32_t m_allocSize;
2013-08-03 00:12:24 +04:00
// m_allocOff is offset into m_allocBuf where we start reading into
// from the file
2014-11-11 01:45:11 +03:00
int32_t m_allocOff;
// do not call pthread_create() for every read we do. use async io
// because it should be much much faster
#ifdef ASYNCIO
struct aiocb m_aiocb[2];
#endif
2013-08-03 00:12:24 +04:00
};
class BigFile {
public:
~BigFile();
BigFile();
// . set a big file's name
// . we split little files that make up this BigFile between
// "dir" and "stripeDir"
bool set ( char *dir , char *baseFilename , char *stripeDir = NULL );
// self explanatory
bool doesExist ( ) ;
// does file part #n exist?
2014-11-11 01:45:11 +03:00
bool doesPartExist ( int32_t n ) ;
2013-08-03 00:12:24 +04:00
// . does not actually open any part file we have
// . waits for a read/write operation before doing that
// . if you set maxFileSize to -1 we set it to BigFile::getFileSize()
// . if you are opening a new file for writing, you need to provide it
// if you pass in a DiskPageCache ptr
bool open ( int flags ,
class DiskPageCache *pc = NULL ,
2014-10-30 22:36:39 +03:00
int64_t maxFileSize = -1 ,
2013-08-03 00:12:24 +04:00
int permissions =
S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH );
int getFlags() { return m_flags; };
2014-11-11 01:45:11 +03:00
void setBlocking ( ) { m_flags &= ~((int32_t)O_NONBLOCK); };
2013-08-03 00:12:24 +04:00
void setNonBlocking ( ) { m_flags |= O_NONBLOCK ; };
// . return -2 on error
// . return -1 if does not exist
// . otherwise return the big file's complete file size (can b >2gb)
2014-10-30 22:36:39 +03:00
int64_t getFileSize ( );
int64_t getSize ( ) { return getFileSize(); };
2013-08-03 00:12:24 +04:00
// use the base filename as our filename
char *getFilename() { return m_baseFilename; };
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . otherwise, returns 1 if the read was completed
// . decides what 2gb part file(s) we should read from
bool read ( void *buf ,
2014-11-11 01:45:11 +03:00
int32_t size ,
2014-10-30 22:36:39 +03:00
int64_t offset ,
2013-08-03 00:12:24 +04:00
FileState *fs = NULL ,
void *state = NULL ,
void (* callback)(void *state) = NULL ,
2014-11-11 01:45:11 +03:00
int32_t niceness = 1 ,
2013-08-03 00:12:24 +04:00
bool allowPageCache = true ,
bool hitDisk = true ,
2014-11-11 01:45:11 +03:00
int32_t allocOff = 0 );
2013-08-03 00:12:24 +04:00
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . IMPORTANT: if returns -1 it MAY have written some bytes
// successfully to OTHER parts that's why caller should be
// responsible for maintaining current write offset
bool write ( void *buf ,
2014-11-11 01:45:11 +03:00
int32_t size ,
2014-10-30 22:36:39 +03:00
int64_t offset ,
2013-08-03 00:12:24 +04:00
FileState *fs = NULL ,
void *state = NULL ,
void (* callback)(void *state) = NULL ,
2014-11-11 01:45:11 +03:00
int32_t niceness = 1 ,
2013-08-03 00:12:24 +04:00
bool allowPageCache = true );
// unlinks all part files
bool unlink ( );
// . renames ALL parts too
// . doesn't change directory, just the base filename
// . use m_dir if newBaseFilenameDir is NULL
bool rename ( char *newBaseFilename , char *newBaseFilenameDir=NULL) ;
2013-12-08 09:12:48 +04:00
bool move ( char *newDir );
2013-08-03 00:12:24 +04:00
// . returns false and sets g_errno on failure
// . chop only parts LESS THAN "part"
2014-11-11 01:45:11 +03:00
bool chopHead ( int32_t part );
2013-08-03 00:12:24 +04:00
// . these here all use threads and call your callback when done
// . they return false if blocked, true otherwise
// . they set g_errno on error
bool unlink ( void (* callback) ( void *state ) ,
void *state ) ;
bool rename ( char *newBaseFilename ,
void (* callback) ( void *state ) ,
void *state ) ;
2014-11-11 01:45:11 +03:00
bool chopHead ( int32_t part ,
2013-08-03 00:12:24 +04:00
void (* callback) ( void *state ) ,
void *state ) ;
// closes all part files
bool close ();
// just close all the fds of the part files, used by RdbMap.cpp.
bool closeFds ( ) ;
2014-10-30 22:36:39 +03:00
//int getfdByOffset ( int64_t offset );
2013-08-03 00:12:24 +04:00
// what part (little File) of this BigFile has offset "offset"?
2014-10-30 22:36:39 +03:00
int getPartNum ( int64_t offset ) { return offset / MAX_PART_SIZE; };
2013-08-03 00:12:24 +04:00
// . opens the nth file if necessary to get it's fd
// . returns -1 if none, >=0 on success
2014-11-11 01:45:11 +03:00
int getfd ( int32_t n , bool forReading , int32_t *vfd = NULL );
2013-08-03 00:12:24 +04:00
// public for wrapper to call
//bool readwrite_r ( FileState *fstate );
2014-10-30 22:36:39 +03:00
int64_t m_currentOffset;
2013-08-03 00:12:24 +04:00
DiskPageCache *getDiskPageCache ( ) { return m_pc; };
2014-11-11 01:45:11 +03:00
int32_t getVfd ( ) { return m_vfd; };
2013-08-03 00:12:24 +04:00
// WARNING: some may have been unlinked from call to chopHead()
2014-11-11 01:45:11 +03:00
int32_t getNumParts ( ) { return m_numParts; };
2013-08-03 00:12:24 +04:00
2014-11-11 01:45:11 +03:00
File *getFile ( int32_t n ) { return m_files[n]; };
2013-08-03 00:12:24 +04:00
// makes the filename of part file #n
void makeFilename_r ( char *baseFilename ,
char *baseFilenameDir ,
2014-11-11 01:45:11 +03:00
int32_t n ,
2013-08-03 00:12:24 +04:00
char *buf );
2014-11-11 01:45:11 +03:00
void removePart ( int32_t i ) ;
2013-08-03 00:12:24 +04:00
// don't launch a threaded rename/unlink if one already in progress
// since we only have one callback, m_callback
2014-11-11 01:45:11 +03:00
int32_t m_numThreads;
2013-08-03 00:12:24 +04:00
void (*m_callback)(void *state);
void *m_state;
// is the threaded op an unlink? (or rename?)
bool m_isUnlink;
2014-11-11 01:45:11 +03:00
int32_t m_part; // part # to unlink (-1 for all)
2013-08-03 00:12:24 +04:00
// number of parts remaining to be unlinked/renamed
2014-11-11 01:45:11 +03:00
int32_t m_partsRemaining;
2013-08-03 00:12:24 +04:00
// rename stores the new name here so we can rename the m_files[i]
// after the rename has completed and the rename thread returns
char m_newBaseFilename [256];
// if first char in this dir is 0 then use m_dir
char m_newBaseFilenameDir [256];
// store our base filename here
char m_baseFilename [256];
// ptrs to the part files
File *m_files [ MAX_PART_FILES ];
// private:
// . wrapper for all reads and writes
// . if doWrite is true then we'll do a write, otherwise we do a read
// . returns false and sets errno on error, true on success
bool readwrite ( void *buf,
2014-11-11 01:45:11 +03:00
int32_t size,
2014-10-30 22:36:39 +03:00
int64_t offset,
2013-08-03 00:12:24 +04:00
bool doWrite,
FileState *fstate ,
void *state ,
void (* callback) ( void *state ) ,
2014-11-11 01:45:11 +03:00
int32_t niceness ,
2013-08-03 00:12:24 +04:00
bool allowPageCache ,
bool hitDisk ,
2014-11-11 01:45:11 +03:00
int32_t allocOff );
2013-08-03 00:12:24 +04:00
// . returns false if blocked, true otherwise
// . sets g_errno on error
bool unlinkRename ( char *newBaseFilename ,
2014-11-11 01:45:11 +03:00
int32_t part ,
2013-08-03 00:12:24 +04:00
bool useThread ,
void (* callback) ( void *state ) ,
void *state ,
char *newBaseFilenameDir = NULL ) ;
// . add all parts from this directory
// . called by set() above for normal dir as well as stripe dir
bool addParts ( char *dirname ) ;
2014-11-11 01:45:11 +03:00
bool addPart ( int32_t n ) ;
2013-08-03 00:12:24 +04:00
2014-11-11 01:45:11 +03:00
//bool unlinkPart ( int32_t n , bool block );
2013-08-03 00:12:24 +04:00
// if part file not created, will create it
2014-11-11 01:45:11 +03:00
File *getPartFile ( int32_t n ) { return m_files[n]; };
2013-08-03 00:12:24 +04:00
// . put a signal on the queue to do reading/writing
// . we call readwrite ( FileState *) when we handle the signal
void addSig ( FileState *fstate ) ;
bool reset ( );
// store our base filename here
char m_dir [256];
char m_stripeDir [256];
2014-11-11 01:45:11 +03:00
int32_t m_permissions;
int32_t m_flags;
2013-08-03 00:12:24 +04:00
// determined in open() override
int m_numParts;
// maximum part #
2014-11-11 01:45:11 +03:00
int32_t m_maxParts;
2013-08-03 00:12:24 +04:00
class DiskPageCache *m_pc;
2014-11-11 01:45:11 +03:00
int32_t m_vfd;
2013-08-03 00:12:24 +04:00
bool m_vfdAllowed;
// prevent circular calls to BigFile::close() with this
char m_isClosing;
2014-10-30 22:36:39 +03:00
int64_t m_fileSize;
2013-08-03 00:12:24 +04:00
// oldest of the last modified dates of all the part files
time_t m_lastModified;
time_t getLastModifiedTime();
};
2014-11-11 01:45:11 +03:00
extern int32_t g_unlinkRenameThreads;
2013-08-03 00:12:24 +04:00
2014-10-30 22:36:39 +03:00
extern int64_t g_lastDiskReadStarted;
extern int64_t g_lastDiskReadCompleted;
2013-08-03 00:12:24 +04:00
extern bool g_diskIsStuck;
extern void *readwriteWrapper_r ( void *state , class ThreadEntry *t ) ;
#endif