Re: [PATCHv3 05/16] pps: access pps device by direct pointer

[Alexander Gordeev]

Ð Thu, 5 Aug 2010 14:31:47 +0200
Rodolfo Giometti <giometti@xxxxxxxxxxxx> ÐÐÑÐÑ:

> On Thu, Aug 05, 2010 at 03:42:31PM +0400, Alexander Gordeev wrote:
> > ?? Thu, 5 Aug 2010 11:32:36 +0200
> > Rodolfo Giometti <giometti@xxxxxxxxxxxx> ??????????:
> > 
> > > On Thu, Aug 05, 2010 at 01:06:42AM +0400, Alexander Gordeev wrote:
> > > > Using device index as a pointer needs some unnecessary work to be done
> > > > every time the pointer is needed (in irq handler for example).
> > > > Using a direct pointer is much more easy (and safe as well).
> > > > 
> > > > Signed-off-by: Alexander Gordeev <lasaine@xxxxxxxxxxxxx>
> > [snip]
> > > 
> > > If you remove these functions you can't be sure anymore that nobodies
> > > may call pps_event() over a non existent device...
> > 
> > [snip]
> > 
> > > By dropping pps_get_source you may be here by a call from (i.e.) a
> > > serial port driver whose doesn't know if your PPS source is gone or
> > > not...
> > > 
> > > I don't understand how your modifications may resolve this problem.
> > 
> > Well, this can happen only if PPS client module calls pps_event before
> > calling pps_register_source() or after pps_unregister_source(). This
> > means that it's broken! If we try to handle/workaround broken clients it
> 
> Suppose we are using pps-ldisc client. How can you assure that
> nobodies may execute pps_tty_close() while you are into the
> pps_event() related to the same serial port?
> 
> You can't disable serial interrupts in order to avoid
> pps_tty_dcd_change calls...

Hmm, yes, I see...
But this is custom problem of only one client. I think it should be
fixed in place instead of trying to fix it in the subsystem.

Are you 100% sure dcd_change can be called before open or after close?
Then I'll try to deal with this.

> > affects performance. So we have to choose what is the priority:
> > security or performance. My guru told me I shouldn't bother too much
> > about broken kernel-space code which my code interacts with. If it's
> > broken it should be fixed. Some assertions enabled by DEBUG define are
> > enough. For me it makes sense but I don't know what should I check here?
> 
> I'm sorry but I disagree with you. Kernel code can't allow userland
> programs to corrupt it!
> 
> We are not discussing about security or performance but about
> reliability.

Sure, now I see the problem (in the pps-ldisc).

-- 
  Alexander

Attachment: signature.asc
Description: PGP signature