Re: debugctl msr

[Markus Metzger <markus.t.metzger@googlemail.com>]

On Wed, 2008-11-19 at 18:13 +0100, stephane eranian wrote:

> Speaking of locking, I also ran into another issue with ds_lock.
> Perfmon sessions each have a spinlock for access serialization, but to
> prevent from PMU and timers interrupts, interrupts are masked. Thus,
> when perfmon
> calls ds.c, interrupts are masked. That means that we lock/unlock ds_lock
> with interrupts disabled. The lock checker triggered when I ran a simple perfmon
> session and warned of possible lock inversion. Suppose you are coming from the
> ptrace code into ds. You grab ds_lock, but the same process is also running
> a perfmon session with PEBS and a counter overflows, you get into
> the PMU interrupt handler which may call into ds.c and try to grab the ds_lock.
> For that reason, I think you should use a
> spin_lock_irqsave/spin_unlock_irqrestore
> pairs to protect your ds context.

OK. So far, there was no user that called ds_*() with interrupts
disabled.


> I found another issue with ds_release(). You need to skip freeing the
> buffer when it
> is NULL, i.e., was already allocated by caller of ds_request_pebs().

ds_release() is not robust with respect to double release, if that's
what you mean. Is that desirable?

For a single ds_release() call matching a corresponding successful
ds_request() call, the buffer is freed if and only if it had been
allocated by ds.c.

Kfree() itself handles NULL pointers and scripts/checkpatch.pl warns on
a check for NULL around a kfree() call.


> I have attached a diff for the ds.c interface. It disables
> ds_validate_access(), export
> the PEBS functions to modules, fixes ds_release().
> 


> As for handling the interrupt is ds.c, not clear how this could work
> with current perfmon.
> I don't know how this work on the BTS side. On the PMU side, that is not because
> I am using PEBS, that I don't also use other counters as well. Longer
> term, I think, there
> needs to be a lower-level PMU interrupt service where you would
> register a callback
> on PMU interrupts. It would be used by NMI watchdog, perfmon,
> Oprofile, ds.c. 

That's even preferable to having the interrupt code itself in ds.c

The point I was trying to make is that buffer overflows should not be
handled on higher levels (i.e. users of ds.c). That's why I am so
reluctant to expose the interrupt threshold in the ds.c interface.


regards,
markus.