• Alexander Viro's avatar
    [PATCH] fix check_disk_change() deadlocks · 951f4be9
    Alexander Viro authored
    Small, but tricky: fix for check_disk_change() deadlocks.
    What we do is
    	a) opening block device shifted from check_partition() to
    	   grok_partitions(); check_partitions() takes opened
    	   struct block_device.
    	b) all callers of check_disk_change() fall in two groups -
    	   ones that are called only from some ->open() and ones
    	   that are _never_ called from ->open().  There is no
    	   middle ground.  We split the thing in two functions -
    	   check_disk_change() for the first class and full_check_....
    	   for the second.  The former (ones inside ->open()) doesn't
    	   touch partition tables but marks the bdev as "had been
    	   invalidated".  In the end of do_open() we check if
    	   bdev is marked and call wipe_partitions()/check_partition()
    	   if it is - at that point bdev is fully set up and ready.
    	c) ->bd_part_sem kludge is gone - we use ->bd_sem instead.
    	   That is, do_open() on a partition grabs ->bd_sem on entire
    	   disk and picks partition data while under it; do_open() on
    	   entire disk rereads partition if needed before dropping
    	   ->bd_sem (right before dropping it); BLKRRPART does
    	   trylock on ->bd_sem and then checks ->bd_part_count -
    	   same logics as before, except that we use ->bd_sem instead
    	   of ->bd_part_sem.
    
    That kills recursive open(), gives us the same exclusion rules as
    we had and makes sure that actual IO (including rereading partition
    tables) is done only when we are ready to do it.
    
    It actually sounds a lot nastier than it is.  do_open() is a one sick
    puppy right now, but we have everything in one place and _out_ of drivers
    (and 20-odd equally sick puppies are gone from them, along with about
    the same number of races).
    
    Now we are almost ready to clean it up for good - all that remains to
    do before that is to get the rest of drivers (cciss, DAC960, i2o and
    a couple of ancients - xd and acsi) using per-disk gendisks.  Then
    most of that crap will disappear.
    
    BTW, the only generic ioctl remaining in the drivers is HDIO_GETGEO -
    a lot of foo_ioctl() starts with if (cmd != HDIO_GETGEO) return -EINVAL; ;-)
    951f4be9
super.c 15.3 KB