[Pvfs-users] pvfs_read beyond EOF

[Rob Ross]

Hi Wei-keng,

PVFS2 isn't doesn't hold to the POSIX standard for all cases.  In the 
past we have cleared the entire read buffer in order to solve the 
problem of sparse file data access (servers with no data don't return 
anything).  In the general case (noncontiguous buffers and I/O regions 
that might be read in any order), this is a hard problem to solve. 
That's why you're seeing the behavior that you are seeing.

However, I think that this behavior has changed recently and is likely 
closer to what you are expecting.  Sam, what's the latest behavior in 
this case?  Wei-keng, what version of PVFS2 are you using?

For my own edification, I found a copy of the POSIX spec here:
   http://www.opengroup.org/onlinepubs/009695399/toc.htm

Here's the closest text that I found describing EOF behavior in that spec:

   No data transfer shall occur past the current end-of-file. If the
   starting position is at or after the end-of-file, 0 shall be returned.
   If the file refers to a device special file, the result of subsequent
   read() requests is implementation-defined.

That's ambiguous.  "data transfer" is vague.  They probably intended 
your interpretation, but they should have specified the condition of the 
buffer explicitly, perhaps in the rationale section.  I wouldn't count 
on that behavior on other systems either.

Anyway, Sam will fill us in on the current semantic.  Given how much 
trouble it has been to get to where we are, I honestly doubt we'll 
adjust that code again in the near future.

Regards,

Rob

Wei-keng Liao wrote:
> 
> What will happen if a read (pvfs_read) reads beyond the current end of 
> file (EOF)? According to POSIX, read() should stop at EOF and return the 
> number of bytes read to the buffer. The buffer beyond EOF should not be 
> touched. (Please correct me if this is wrong.)
> 
> But, I found the pvfs_read will set the buffer contents to null for the 
> beyond part. Here is the example code I tested: 1) create a new file, 2) 
> initialize read buffer, 3) read 10 bytes, 4) print out buffer. (I am 
> using PVFS 1.6.3) Is this the way PVFS behaves for reading beyond EOF?
> 
> Wei-keng Liao
> ---------------------------------------------------------------------
> 
> 
> #include <string.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <fcntl.h>
> #include <pvfs.h>
> 
> #define PVFS_FILE_NAME "/pvfs/wkliao/testfile.dat"
> #define UNIX_FILE_NAME "testfile.dat"
> 
> #define TEST_FILE(filename,ofunc,cfunc,rfunc)    {               \
>     fd = ofunc(filename, flag, 0600, 0, 0);                      \
>                                                                  \
>     readsize = rfunc(fd, buf, 10);                               \
>     cfunc(fd);                                                   \
>                                                                  \
>     printf("readsize for %s = %d bytes\n", filename, readsize);  \
>     printf("buf[*]= ");                                          \
>     for (i=0; i<10; i++) printf(" '%c'", buf[i]);                \
>     printf("\n");                                                \
> }
> 
> int main(int argc, char **argv) {
>     int  i, readsize, flag;
>     int  fd;
>     char buf[10];
> 
>     flag = O_TRUNC | O_CREAT | O_RDWR;
> 
>     /* test pvfs file -------------------------------------------*/
>     for (i=0; i<10; i++) buf[i] = '1';
>     TEST_FILE(PVFS_FILE_NAME, pvfs_open, pvfs_close, pvfs_read);
> 
>     /* test unix file -------------------------------------------*/
>     for (i=0; i<10; i++) buf[i] = '1';
>     TEST_FILE(UNIX_FILE_NAME, open, close, read);
> 
>     return 1;
> }