[Patch 0/2] powerpc: avoid userspace poking to legacy ioports

Jean Delvare khali at linux-fr.org
Tue Feb 19 07:58:42 EST 2008


On Tue, 19 Feb 2008 07:42:03 +1100, Benjamin Herrenschmidt wrote:
> 
> > Maybe Christian's patch can be improved to not do the check on these?
> > As long as /dev/port exists, it seems reasonable that the kernel should
> > behave, no matter what I/O ports are accessed from user-space.
> 
> nonsense.
> 
>  /dev/mem exists for example, but you are still not supposed to go
> bang all over the place in it.

You should at least be able to read from it without crashing the
machine. Of course writing is a different story.

> > > I hate that sensors_detect.. or for that matter any other userland code
> > > that pokes random ports like that. It should die.
> > 
> > What do you propose as a replacement?
> 
> Dunno, something less scary, like knowing where your sensors are on a
> given machine...

You mean, having a complete database for the, what, 4000 PC
motherboards out there? And maintaining it day after day? _This_ sounds
much scarier to me than the current situation.

> honestly, it's just scary the risk you guys are taking
> by banging random IO ports.

I don't remember anyone reporting problems with this in the past 3 or 4
years, so it doesn't seem to be a big problem in practice.

> At the very least, that shouldn't be done on non-x86.

I am surprised that anyone would actually run sensors-detect on
non-x86... Non-PC hardware usually doesn't have sensors driven by
"hwmon" drivers anyway, or people know what they have and do not need
detection. But I would be totally fine with updating sensors-detect to
skip some of the probes on non-x86 hardware. There are basically
3 /dev/port probes that are done currently:

* Super-I/O chips at 0x2e/0x2f and 0x4e/0x4f.

* Legacy PC hardware monitoring chips at 0x290-0x297.

* IPMI interface at 0x0ca3 and 0x0cab (read-only).

Please tell me which ones should be skipped on PowerPC.

Christian, can you tell me which of these probes caused trouble for you?

> > And how is userland code poking at random ports different from kernel
> > code poking at random ports? We could move sensors-detect inside the
> > kernel (and I have some plan to do that) but I fail to see how this
> > would solve this particular problem.
> 
> It wouldn't, but at least I could NAK it or make it CONFIG_X86 :-)

The same could be done for user-space (or at the /dev/port level.)

-- 
Jean Delvare



More information about the Linuxppc-dev mailing list