Re: [PATCH v4 3.0-rc2-tip 4/22] 4: Uprobes: register/unregister probes.

[Oleg Nesterov <oleg@redhat.com>]

Hi Srikar,

I still hope some day I'll find the time to read the whole series ;)
Trying to continue from where I have stopped, and it seems that this
patch has a couple more problems.

On 06/07, Srikar Dronamraju wrote:
>
> A probe is specified by a file:offset.  While registering, a breakpoint
> is inserted for the first consumer, On subsequent probes, the consumer
> gets appended to the existing consumers. While unregistering a
> breakpoint is removed if the consumer happens to be the last consumer.
> All other unregisterations, the consumer is deleted from the list of
> consumers.
>
> Probe specifications are maintained in a rb tree. A probe specification
> is converted into a uprobe before store in a rb tree.  A uprobe can be
> shared by many consumers.

register/unregister logic looks racy...

Supose that uprobe U has a single consumer C and register_uprobe()
is called with the same inode/offset, while another thread does
unregister(U,C).

	- register() calls alloc_uprobe(), finds the entry in rb tree,
	  and increments U->ref. But this doesn't add the new consumer.

	- uregister() does del_consumer(), and removes the single
	  consumer C.

	  then it takes uprobes_mutex, sees uprobe->consumers == NULL
	  and calls delete_uprobe()->rb_erase()

	- register() continues, takes uprobes_mutex, re-inserts the
	  breakpoints, finds the new consumer and succeeds.

	  However, this uprobe is not in rb-tree, it was deleted
	  by unregister.



OTOH. Suppose we add the new uprobe. register()->alloc_uprobe() sets
new_uprobe->ref == 2. If something goes wrong after that, register()
does delete_uprobe() + put_uprobe(), new_uprobe->ref becomes 1 and
we leak this uprobe.

Oleg.