Re: [RFC] Design + prototype: Multiple kprobes at an address - take 2

["Frank Ch. Eigler" <fche at redhat dot com>]

Hi -


> [...]
> +static DEFINE_SPINLOCK(multprobe_lock);

It would be better for this to be used as a read/write- rather than
mutex- type spinlock, since the hot path (handler execution) is a
read-only operation that should before long allow concurrent
execution.


> [...]
> /* common kprobes fault handler that gets control when the registered probe
>  * gets fired. This routines is wrapper over the inserted multiple handlers
>  * at a given address and calls individual handlers.
>  */

As we discussed, this iteration policy will likely need changes.
Maybe the concept of a fault handler needs to be forked into
protection *for* the pre/post-handlers, which should naturally be
per-kprobes-client, and for the single-stepping phase, which is not.

I actually still haven't seen a useful thing that custom a fault
handler could do for the second case.  Is there some known kprobing
scenario where single-stepping faults are dealt with usefully *by* the
custom handler (and not the generic unconditional kprobes code)?

Also, is there no risk of deadlock by having this fault handler
routine also take multiprobe_lock, considering that the lock is also
held around the entire pre/post handling iterations?


> [...]
> int register_multiprobe(struct mult_handler *multh)
> [...]
> 	spin_lock_irqsave(&multprobe_lock, flags);
> 
> 	spin_lock_irqsave(&kprobe_lock, flags1);
> 	temp = get_kprobe(multh->kp.addr);
> 	spin_unlock_irqrestore(&kprobe_lock, flags1);
> [...]

Can you explain what potential concurrency problems are prevented by
explicitly holding kprobe_lock here (and at the unregister call)?


- FChE

Attachment: pgp00000.pgp
Description: PGP signature