Re: [PATCH 05/10] drivers/char: Support compiling out /dev/zero

[Josh Triplett <josh@joshtriplett.org>]

On Wed, Jan 28, 2015 at 05:20:08PM -0600, Tom Zanussi wrote:
> On Wed, 2015-01-28 at 13:51 -0800, josh@joshtriplett.org wrote:
> > On Wed, Jan 28, 2015 at 10:07:51PM +0100, Pavel Machek wrote:
> > > On Fri 2015-01-23 12:37:11, Tom Zanussi wrote:
> > > > Some embedded systems with tightly controlled userspace have no use
> > > > for /dev/zero, and could benefit from the size savings gained by
> > > > omitting it.  Add a new EMBEDDED config option to disable it.
> > > > 
> > > > bloat-o-meter (based on tinyconfig):
> > > > 
> > > > add/remove: 0/3 grow/shrink: 0/1 up/down: 0/-391 (-391)
> > > > function                                     old     new   delta
> > > > chr_dev_init                                 162     147     -15
> > > > mmap_zero                                     16       -     -16
> > > > zero_fops                                    116       -    -116
> > > > zero_bdi                                     244       -    -244
> > > > 
> > > > Signed-off-by: Tom Zanussi <tom.zanussi@linux.intel.com>
> > > 
> > > I'm not sure that 400 bytes are worth additional Kconfig noise. .. and
> > > pretty much everyone needs /dev/zero...
> > 
> > Relatively few, actually, given MMAP_ANONYMOUS.  Memory isn't allocated
> > via an mmap of /dev/zero.  It's useful for systems with shells that want
> > to redirect from it or read from it, but less useful for environments
> > with entirely compiled code.
> > 
> > /dev/null is much more commonly needed, though there are still systems
> > that won't need it (and can just disable read/writes on an fd entirely
> > rather than duping /dev/null to that fd).
> 
> For testing, I was able to boot my dev system (Ubuntu) into a usable
> shell with networking with only /dev/null and /dev/urandom.  A
> restricted userspace could be made/verified to not touch /dev/null. 

And a carefully constructed system with narrower needs (rather than a
general-purpose distribution) could avoid touching *any* files in /dev.

> > That said, I'd be entirely in favor of consolidating many of these
> > "miscellaneous character device" options into a couple of Kconfig
> > options.  It doesn't seem critical to *individually* control each of
> > these files in /dev.
> 
> I can easily create a small set of groupings instead - how about
> something like:  DEVMEM_RANDOM(/dev/random, /dev/urandom, getrandom()),
> DEVMEM_MEM (/dev/mem, /dev/kmem, /dev/port), DEVMEM_RW
> (/dev/null, /dev/zero), and DEVMEM_MISC (/dev/full, /dev/kmsg)?
> 
> That cuts the number in half, from 8->4 (we still have a separate
> DEVPORT and DEVKMEM regardless).

I don't think it makes sense to group full and kmsg.  I'd suggest giving
kmsg its own option (which likely needs to depend on printk, since it
doesn't make sense without a kernel message buffer), and putting null,
zero, and full together.  (Though full seems like the least useful of
the lot.)

> I can also get rid of DEVMEM_RANDOM entirely as mentioned in a previous
> post, and not allow those to be disabled at all, which saves even more
> on Kconfig noise.

Please don't.

> > Personally, I'm hoping that we eventually end up with a disableable
> > CONFIG_CHAR similar to CONFIG_BLOCK.
> 
> If we ditch DEVMEM_RANDOM i.e. make certain devices undisableable then
> it seems CONFIG_CHAR wouldn't be possible either.

An excellent reason not to ditch it, then.

- Josh Triplett