Re: ipc/msg.c "cleanup" breaks fakeroot on Alpha

[Manfred Spraul <manfred@colorfullife.com>]

Paul E. McKenney wrote:

> I also don't understand why the code in sys_msgrcv() doesn't have
> to remap the msqid, similar to the way it is done in sys_semtimedop().
>
>  
What do you mean with remap?

> So, what am I missing here?  How does a msgrcv() racing with an rmid()
> avoid taking a lock on a message queue that just got freed?  (The
> ipc_lock_by_ptr() in "Lockless receive, part 3".)  My concern is the
> following sequence of steps:
>
> o	expunge_all() invokes wake_up_process() and sets r_msg.
>
> o	sys_msgrcv() is awakened, but for whatever reason does
> 	not actually start executing (e.g., lots of other busy
> 	processes at higher priority).
>
> o	expunge_all() returns to freeque(), which runs through the
> 	rest of its processing, finally calling ipc_rcu_putref().
>
> o	ipc_rcu_putref() invokes call_rcu() to free the message
> 	queue after a grace period.
>
> o	ipc_immediate_free() is invoked at the end of a grace
> 	period, freeing the message queue.
>
> o	sys_msgrcv() finally gets a chance to run, and does an
> 	rcu_read_lock() -- but too late!!!
>
>  
Not too late:
sys_msgrcv() checks msr_d.r_msr, notices that the value is -EIDRM and
returns to user space with -EIDRM immediately. This codepath
doesn't touch the message queue pointer, thus it doesn't matter that the
message queue is already freed.
The code only touches the message queue pointer if msr_d.r_msr
is -EAGAIN - and the rcu_read_lock() guarantees there is no rcu grace
period between the test for -EAGAIN and the ipc_lock_by_ptr.
Thus this should be safe.

But back to the oops:
The oops happens in expunge_all, called from sys_msgctl.
Thus it must be an msgctl(IPC_SET).
IPC_SET is special: it calls expunge_all(-EAGAIN): that's necessary
because IPC_SET can change the permissions.
Unfortunately, faked doesn't use IPC_SET at all :-(

Falk - could you strace your "fakeroot ls" test? Are there any IPC_SET 
calls?
Which gcc version do you use? Is it possible that gcc auto-inlined 
something?

--
   Manfred
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/