I recently had to migrate a large multi-user-access database system from Solaris to Linux. I took care of big/little-endian by means of C++ data-types called xlong/xshort (See: http://www.stanford.edu/dept/its/support/uspires/xlong/ ). But buffered I/O no longer worked as it did in Solaris, where buffers where maintained at the system level. Any changes made to records by one user were immediately known to other users. But in Linux, buffers were maintained locally in each user's workspace. So if user A altered a control block in A's buffer, it was NOT known in B's buffer. Each time B tried to read a control block, B got the same (old) information. None of A's changes appeared in B's block.
The solution was to recode the file I/O to use "read" and "write" and "lseek" instead of "fread" and "fwrite" and "ftell", along with a few other functional differences. The files could still be opened and closed with "fopen" and "fclose", yielding a FILE di (device identifier), but I/O had to be done with a file identifier: fid = fileno(di).
What follows is a summary of all the if-tested logic added to the program to allow it to work on both Solaris (buffered) and Linux (unbuffered):
/*- Use -DEMFBUFF to define buffered I/O everywhere -*/
#define Fsync(x) fsync(fileno(x))
---
if ((mydi = curDev->di) != 0) /*- device assigned -*/
{ /*- Determine where file ends -*/
#ifdef EMFBUFF
if (!fseek(mydi, 0L, SEEK_END)) /*- Position -*/
llen = ftell(mydi); /*- Get EOF position -*/
#else
lseekpos = lseek(fileno(mydi), 0L, SEEK_END);
if (lseekpos > 0) llen = (long)lseekpos; /*- Position -*/
#endif
clearerr(mydi); /*- Everything is error free -*/
}
---
{ /*--- Process READ and WRITE for files ---*/
#ifdef EMFBUFF
long initpos; /*- Initial position -*/
long position; /*- Position in file -*/
#else
off_t initpos; /*- Initial position -*/
off_t position; /*- Position in file -*/
off_t lseekpos; /*- lseek position -*/
#endif
long blklen; /*- length of I/O data -*/
long blknum; /*- block to read/write -*/
byte *blkloc; /*- location of buffer -*/
#ifdef EMFBUFF
int excess;
#else
ssize_t excess;
int fid; /*- Unbuffered file identifier -*/
#endif
int blksiz, loop;
FILE *mydi; /*- Buffered file stream -*/
---
#ifndef EMFBUFF
fid = fileno(mydi);
#endif
---
#ifdef EMFBUFF
if (fseek(mydi, position, SEEK_SET))
#else
if (lseek(fid, position, SEEK_SET) < 0)
#endif
---
#ifdef EMFBUFF
{ if ((excess = fread(blkloc, 1, blksiz, mydi)) == 0)
#else
{ if ((excess = read(fid, blkloc, (size_t)blksiz)) <= 0)
#endif
---
#ifdef EMFBUFF
{ excess = fwrite(blkloc, 1, blksiz, mydi);
if (excess == 0)
#else
{ excess = write(fid, blkloc, (size_t)blksiz);
if (excess <= 0)
#endif
---
#ifdef EMFBUFF
if (fflush(mydi))
{ setError(curDev);
return;
}
#else
Fsync(mydi);
#endif
---
#ifdef EMFBUFF
clearerr(mydi); /*- Clear all errors -*/
blklen = ftell(mydi) - initpos;
#else
if (code != 247) Fsync(mydi);
clearerr(mydi); /*- Clear all errors -*/
lseekpos = lseek(fid, 0L, SEEK_CUR);
if (lseekpos <= 0) blklen = 0;
else blklen = (long)(lseekpos - initpos);
#endif
===========
I've supplied this if-tested code to give everyone an opportunity to see the transformation. There are subtle differences in how things like "fread" and "read" operate, "fwrite" and "write" operate, and how seeking and flushing blocks operate. Buffered I/O still works fine for non-shared files, like standard input and output associated with each individual user. But shares files require Unbuffered I/O, which appears to force any "buffers" to be maintained at the system-level where everyone has immediate access to any changes.
If you're having troubles with buffered I/O because of simultaneous access by multiple users, then switch to unbuffered I/O.
|
Submit Your Article
Latest Articles
Editor's Choice
Article Categories
Join Linux Forums Today
Win Great Prizes
User Score
Editor Score
Tools »
Bookmark
Share
Comment