• Jeff Dike's avatar
    [PATCH] uml: fix I/O hang · 53b17332
    Jeff Dike authored
    Fix a UML hang in which everything would just stop until some I/O happened
    - a ping, someone whacking the keyboard - at which point everything would
    start up again as though nothing had happened.
    
    The cause was gcc reordering some code which absolutely needed to be
    executed in the order in the source.  When unblock_signals switches signals
    from off to on, it needs to see if any interrupts had happened in the
    critical section.  The interrupt handlers check signals_enabled - if it is
    zero, then the handler adds a bit to the "pending" bitmask and returns.
    unblock_signals checks this mask to see if any signals need to be
    delivered.
    
    The crucial part is this:
    	signals_enabled = 1;
    	save_pending = pending;
    	if(save_pending == 0)
    		return;
    	pending = 0;
    
    In order to avoid an interrupt arriving between reading pending and setting
    it to zero, in which case, the record of the interrupt would be erased,
    signals are enabled.
    
    What happened was that gcc reordered this so that 'save_pending = pending'
    came before 'signals_enabled = 1', creating a one-instruction window within
    which an interrupt could arrive, set its bit in pending, and have it be
    immediately erased.
    
    When the I/O workload is purely disk-based, the loss of a block device
    interrupt stops the entire I/O system because the next block request will
    wait for the current one to finish.  Thus the system hangs until something
    else causes some I/O to arrive, such as a network packet or console input.
    
    The fix to this particular problem is a memory barrier between enabling
    signals and reading the pending signal mask.  An xchg would also probably
    work.
    
    Looking over this code for similar problems led me to do a few more
    things:
    
    - make signals_enabled and pending volatile so that they don't get cached
      in registers
    
    - add an mb() to the return paths of block_signals and unblock_signals so
      that the modification of signals_enabled doesn't get shuffled into the
      caller in the event that these are inlined in the future.
    Signed-off-by: default avatarJeff Dike <jdike@addtoit.com>
    Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
    Signed-off-by: default avatarAndrew Morton <akpm@osdl.org>
    Signed-off-by: default avatarLinus Torvalds <torvalds@osdl.org>
    53b17332
barrier.h 184 Bytes