kill_ioctx() used to have an explicit RCU delay between removing the
reference from ->ioctx_table and percpu_ref_kill() dropping the refcount.
At some point that delay had been removed, on the theory that
percpu_ref_kill() itself contained an RCU delay.  Unfortunately, that was
the wrong kind of RCU delay and it didn't care about rcu_read_lock() used
by lookup_ioctx().  As the result, we could get ctx freed right under
lookup_ioctx().  Tejun has fixed that in a6d7cff4 ("fs/aio: Add explicit
RCU grace period when freeing kioctx"); however, that fix is not enough.

Suppose io_destroy() from one thread races with e.g. io_setup() from another;
CPU1 removes the reference from current->mm->ioctx_table[...] just as CPU2
has picked it (under rcu_read_lock()).  Then CPU1 proceeds to drop the
refcount, getting it to 0 and triggering a call of free_ioctx_users(),
which proceeds to drop the secondary refcount and once that reaches zero
calls free_ioctx_reqs().  That does
        INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
        queue_rcu_work(system_wq, &ctx->free_rwork);
and schedules freeing the whole thing after RCU delay.

In the meanwhile CPU2 has gotten around to percpu_ref_get(), bumping the
refcount from 0 to 1 and returned the reference to io_setup().

Tejun's fix (that queue_rcu_work() in there) guarantees that ctx won't get
freed until after percpu_ref_get().  Sure, we'd increment the counter before
ctx can be freed.  Now we are out of rcu_read_lock() and there's nothing to
stop freeing of the whole thing.  Unfortunately, CPU2 assumes that since it
has grabbed the reference, ctx is *NOT* going away until it gets around to
dropping that reference.

The fix is obvious - use percpu_ref_tryget_live() and treat failure as miss.
It's not costlier than what we currently do in normal case, it's safe to
call since freeing *is* delayed and it closes the race window - either
lookup_ioctx() comes before percpu_ref_kill() (in which case ctx->users
won't reach 0 until the caller of lookup_ioctx() drops it) or lookup_ioctx()
fails, ctx->users is unaffected and caller of lookup_ioctx() doesn't see
the object in question at all.

Cc: stable@kernel.org
Fixes: a6d7cff4

 "fs/aio: Add explicit RCU grace period when freeing kioctx"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>