Re: [RESEND] [PATCH] VFS: make file->f_pos access atomic on 32bit arch

[Eric Dumazet <dada1@cosmosbay.com>]

Peter Zijlstra a écrit :
> On Wed, 2008-10-08 at 13:48 +0900, Hisashi Hifumi wrote:
> 
>> Simultaneous access by two or more writer can corrupt file content,
>> so this case needs some locks(flock or fcntl) to preserve synchronization
>> of file content. This is responsibility of user-space application.
>> But file->f_pos race issue can occur even if multiple threads just read 
>> simultaneously. I think this is not responsibility of user-space application. 
>> To avoid this currently, an application needs some locks to protect file offset
>> even if it just read a file. So I think f_pos race should be fixed.
> 
> Just to add to all those who already said you're wrong :-)
> 
> You're wrong, if two threads would like to read the same file they
> either dup() the fd or open() the file twice. There is absolutely no
> valid reason to have two threads read from the same fd without
> synchronising their access to it - never.
> 

About dup() syscall, it wont help, since old and new descriptor points to
the same "struct file", definitly sharing file position, since first Unixes.

To quote the fine manual :

       After successful return of dup or dup2, the old and new descriptors may
       be used interchangeably. They share locks, file position  pointers  and
       flags;  for example, if the file position is modified by using lseek on
       one of the descriptors, the position is also changed for the other.


pread()/pwrite() are used my multi-threaded applications that want to share
a single "struct file". Or they must use some form of synchronization around
regular read()/write()/lseek() calls.

There is no generic f_pos race, only buggy applications.

A far more interesting problem is the "tail -f logfile" problem that raised
recently in lkml, when file is NFS mounted, where reader can get nul bytes...
(Subject : blocks of zeros (NULLs) in NFS files in kernels >= 2.6.20 )