lkml.org 
[lkml]   [2008]   [Oct]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] e1000e: write protect ICHx NVM to prevent malicious write/erase


On Wed, 1 Oct 2008, Jesse Brandeburg wrote:
>
> From: Bruce Allan <bruce.w.allan@intel.com>
>
> Set the hardware to ignore all write/erase cycles to the GbE region in
> the ICHx NVM. This feature can be disabled by the WriteProtectNVM module
> parameter (enabled by default) only after a hardware reset, but
> the machine must be power cycled before trying to enable writes.

Thanks, applied.

One thing that I did notice when I looked at the driver is that I don't
see any serialization what-so-ever around a lot of the special accesses.

There's all these different routines that do

ret_val = e1000_acquire_swflag_ich8lan(hw);
if (ret_val)
return retval;
...
e1000_release_swflag_ich8lan(hw);


but as far as I can tell, there is absolutely _nothing_ that prevents
these from being done concurrently by software.

Yeah, yeah, I'm sure most of them end up being single-threaded and only
called over the probe sequence (well, at least I _hope_ so), but it sure
isn't obvious. People call e1000_read_nvm() from various different
contexts, and I'm not seeing what - if anything - protects two concurrent
ethtool queries, for example.

Imagine that you run ethtool concurrently (even on a UP machine with
preemption of just a sleeping op), and tell me that having two
e1000_acquire_swflag_ich8lan/e1000_release_swflag_ich8lan sequences nest
(or overlap) works. I don't think it does.

That E1000_EXTCNF_CTRL_SWFLAG is _not_ a lock against other threads, it's
purely a lock against the hardware itself. And maybe I'm missing some
locking, but I can't see it.

Same goes for the PHY accesses etc afaik. Hmm?

Linus


\
 
 \ /
  Last update: 2008-10-02 02:47    [W:0.052 / U:1.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site