On Monday January 14, viro@ZenIV.linux.org.uk wrote:
On Mon, Jan 14, 2008 at 02:21:45PM +1100, Neil Brown wrote:
quoted
Maybe it isn't there any more....
Once upon a time, when I
echo remove > /sys/block/mdX/md/dev-YYY/state
Egads. And just what will protect you from parallel callers
of state_store()? buffer->mutex does *not* do that - it only
gives you exclusion on given struct file. Run the command
above from several shells and you've got independent open
from each redirect => different struct file *and* different
buffer for each => no exclusion whatsoever.
well in -mm, rdev_attr_store gets a lock on
rdev->mddev->reconfig_mutex.
It doesn't test is rdev->mddev is NULL though, so if the write happens
after unbind_rdev_from_array, we lose.
A test for NULL would be easy enough. And I think that the mddev
won't actually disappear until the rdevs are all gone (you subsequent
comment about kobject_del ordering seems to confirm that) so a simple test
for NULL should be sufficient.
And _that_ is present right in the mainline tree - it's unrelated
to -mm kobject changes.
BTW, yes, you do have a deadlock there - kobject_del() will try to evict
children, which will include waiting for currently running ->store()
to finish, which will include the caller since .../state *is* a child of
that sucker.
The real problem is the lack of any kind of exclusion considerations in
md.c itself, AFAICS. Fun with ordering is secondary (BTW, yes, it is
a problem - will sysfs ->store() to attribute between export_rdev() and
kobject_del() work correctly?)
Probably not. The possibility that rdev->mddev could be NULL would
break a lot of these. Maybe I should delay setting rdev->mddev to
NULL until after the kobject_del. Then audit them all.
Thanks. I'll see what I can some up with.
NeilBrown