On Wed, 2018-04-04 at 17:54 +0900, Damien Le Moal wrote:
> Since SCSI scanning occurs asynchronously, since sd_revalidate_disk()
> is called from sd_probe_async() and since sd_revalidate_disk() calls
> sd_zbc_read_zones() it can happen that sd_zbc_read_zones() is called
> concurrently with operations referencing a drive zone bitmaps and number
^^^^^^^^^^^^^^^^^^^^
Should "a" be changed into "the"?
> [Damien] Updated commit message and changed nr_zones/bitmap swap order.
Updating the number of zones after having updated the bitmap pointers is not
sufficient to avoid trouble if the number of zones as reported by the drive
changes while I/O is in progress. With the current implementation if the
number of zones changes the seq_zones_bitmap is cleared. Can this cause
trouble for the mq-deadline scheduler? Additionally, CPUs other than x86 can
reorder store operations. Even worse, a CPU could cache the zone bitmap
pointers which means that at least RCU protection + kfree_rcu() is needed to
avoid trouble. I think we either should handle this case properly or issue a
kernel warning.
Thanks,
Bart.