Re: [libnftnl PATCH] src: Fix nftnl_assert() on data_len

[Phil Sutter <phil@nwl.cc>]

Hi Pablo,

On Tue, Feb 18, 2020 at 02:42:27PM +0100, Pablo Neira Ayuso wrote:
> On Sat, Feb 15, 2020 at 11:58:55PM +0100, Phil Sutter wrote:
> > On Sat, Feb 15, 2020 at 02:17:13PM +0100, Pablo Neira Ayuso wrote:
> > > On Sat, Feb 15, 2020 at 01:43:11AM +0100, Phil Sutter wrote:
> > > > On Fri, Feb 14, 2020 at 06:42:00PM +0100, Pablo Neira Ayuso wrote:
> > > > > On Fri, Feb 14, 2020 at 06:34:50PM +0100, Phil Sutter wrote:
> > > > > > On Fri, Feb 14, 2020 at 06:32:47PM +0100, Pablo Neira Ayuso wrote:
> > > > > > > On Fri, Feb 14, 2020 at 06:24:17PM +0100, Phil Sutter wrote:
> > > > > > > > Typical idiom for *_get_u*() getters is to call *_get_data() and make
> > > > > > > > sure data_len matches what each of them is returning. Yet they shouldn't
> > > > > > > > trust *_get_data() to write into passed pointer to data_len since for
> > > > > > > > chains and NFTNL_CHAIN_DEVICES attribute, it does not. Make sure these
> > > > > > > > assert() calls trigger in those cases.
> > > > > > > 
> > > > > > > The intention to catch for unset attributes through the assertion,
> > > > > > > right?
> > > > > > 
> > > > > > No, this is about making sure that no wrong getter is called, e.g.
> > > > > > nftnl_chain_get_u64() with e.g. NFTNL_CHAIN_HOOKNUM attribute which is
> > > > > > only 32bits.
> > > > > 
> > > > > I think it will also catch the case I'm asking. If attribute is unset,
> > > > > then nftnl_chain_get_data() returns NULL and the assertion checks
> > > > > data_len, which has not been properly initialized.
> > > > 
> > > > With nftnl_assert() being (shortened):
> > > > 
> > > > | #define nftnl_assert(val, attr, expr) \
> > > > |  ((!val || expr) ? \
> > > > |  (void)0 : __nftnl_assert_fail(attr, __FILE__, __LINE__))
> > > > 
> > > > Check for 'expr' (which is passed as 'data_len == sizeof(<something>)')
> > > > will only happen if 'val' is not NULL. Callers then return like so:
> > > > 
> > > > | return val ? *val : 0;
> > > > 
> > > > This means that if you pass an unset attribute to the getter, it will
> > > > simply return 0.
> > > 
> > > Thanks for explaining, Phil. If the problem is just
> > > NFTNL_CHAIN_DEVICES and NFTNL_FLOWTABLE_DEVICES, probably this is just
> > > fine? So zero data-length is reversed for arrays and update
> > > nftnl_assert() to skip data_len == 0, ie.
> > > 
> > > > | #define nftnl_assert(val, attr, expr) \
> > > > |  ((!val || data_len == 0 || expr) ? \
> > > > |  (void)0 : __nftnl_assert_fail(attr, __FILE__, __LINE__))
> > 
> > Your proposed patch would allow to call e.g.:
> > 
> > | nftnl_chain_get_u32(c, NFTNL_CHAIN_DEVICES)
> > 
> > This would return (uint32_t)*(&c->dev_array[0]), I highly doubt we
> > should allow this. Unless I miss something, it is certainly a
> > programming error if someone calls any of the nftnl_chain_get_{u,s}*
> > getters on NFTNL_CHAIN_DEVICES attribute. So aborting with error message
> > in nftnl_assert() is not only OK but actually helpful, no?
> 
> Indeed, good point.
> 
> I don't think nftnl_flowtable_set_data() is good for these two device
> array.

Well, right now it serves as a backend for all attribute setters, and
your patch continues in that tradition. So while it may be a bit
"rustic", I'd say it's good enough for its purpose. :)

> I just sent a patch, I forgot to finish the _set_array() and
> _get_array() helpers for the flowtable, the definition in the header
> file prooves this.
> 
> Can we introduce these new interfaces? Then, update nftables to use it.
> Then, at some point, set *data_len = 0 for these array datatypes. Yes,
> it's a bit longer term, but better fix this interface. But setting all
> these data_len to zero when in most cases it is going to be thereafter
> properly set to the datatype length is...
> 
> Would this work for you? I know it is not so short term.

While I think your patch is the right way to providing a sanitized
access to the array attributes, I don't think it's really related to
what my original patch was fixing, which is:

Right now we are preventing users from passing wrong attribute types to
getters by checking the attribute length. This does not work for
NFTNL_CHAIN_DEVICES or NFTNL_FLOWTABLE_DEVICES because they don't set
data_len. Hence the expression in nftnl_asser() call:

| nftnl_assert(val, attr, data_len == sizeof(<something>));

Will lead to comparing with garbage from stack. This may in most cases
fail as expected, but there's no guarantee.

Your patch allows to use "a better" getter/setter for those problematic
attributes, but it doesn't prevent the above from happening.

My first approach was to make nftnl_chain_get_data() and
nftnl_flowtable_get_data() set:

| *data_len = 0;

for the problematic attributes, but the value is not really correct - a
"more correct" value, e.g.:

| *data_len = c->dev_array_len * sizeof(char *);

Could lead to a pass in getter sanitizing by accident although e.g.
nftnl_chain_get_u64() is completely unfit even if c->dev_array_len was
1.

So I decided to go the safe way and initialize data_len variables to zero
instead which has the benefit of catching new attributes added later as
well.

If you don't like the approach of initializing all data_len variables, I
would rather suggest to go with setting '*data_len = 0' in _get_data()
routines as described above. This has the same effect but it's just a
two lines change. What do you think?

Cheers, Phil