• Andi Kleen's avatar
    vfs: do (nearly) lockless generic_file_llseek · ef3d0fd2
    Andi Kleen authored
    The i_mutex lock use of generic _file_llseek hurts.  Independent processes
    accessing the same file synchronize over a single lock, even though
    they have no need for synchronization at all.
    
    Under high utilization this can cause llseek to scale very poorly on larger
    systems.
    
    This patch does some rethinking of the llseek locking model:
    
    First the 64bit f_pos is not necessarily atomic without locks
    on 32bit systems. This can already cause races with read() today.
    This was discussed on linux-kernel in the past and deemed acceptable.
    The patch does not change that.
    
    Let's look at the different seek variants:
    
    SEEK_SET: Doesn't really need any locking.
    If there's a race one writer wins, the other loses.
    
    For 32bit the non atomic update races against read()
    stay the same. Without a lock they can also happen
    against write() now.  The read() race was deemed
    acceptable in past discussions, and I think if it's
    ok for read it's ok for write too.
    
    => Don't need a lock.
    
    SEEK_END: This behaves like SEEK_SET plus it reads
    the maximum size too. Reading the maximum size would have the
    32bit atomic problem. But luckily we already have a way to read
    the maximum size without locking (i_size_read), so we
    can just use that instead.
    
    Without i_mutex there is no synchronization with write() anymore,
    however since the write() update is atomic on 64bit it just behaves
    like another racy SEEK_SET.  On non atomic 32bit it's the same
    as SEEK_SET.
    
    => Don't need a lock, but need to use i_size_read()
    
    SEEK_CUR: This has a read-modify-write race window
    on the same file. One could argue that any application
    doing unsynchronized seeks on the same file is already broken.
    But for the sake of not adding a regression here I'm
    using the file->f_lock to synchronize this. Using this
    lock is much better than the inode mutex because it doesn't
    synchronize between processes.
    
    => So still need a lock, but can use a f_lock.
    
    This patch implements this new scheme in generic_file_llseek.
    I dropped generic_file_llseek_unlocked and changed all callers.
    Signed-off-by: default avatarAndi Kleen <ak@linux.intel.com>
    Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
    ef3d0fd2
file.c 24.2 KB