On 12/1/2023 12:31 AM, Casey Schaufler wrote:
On 11/30/2023 1:34 PM, Roberto Sassu wrote:
On 11/30/2023 5:15 PM, Casey Schaufler wrote:
On 11/30/2023 12:30 AM, Petr Tesarik wrote:
Hi all,
On 11/30/2023 1:41 AM, Casey Schaufler wrote:
...
It would be nice if the solution directly addresses the problem.
EVM needs to be after the LSMs that use xattrs, not after all LSMs.
I suggested LSM_ORDER_REALLY_LAST in part to identify the notion as
unattractive.
Excuse me to chime in, but do we really need the ordering in code?
tl;dr - Yes.
FWIW
the linker guarantees that objects appear in the order they are seen
during the link (unless --sort-section overrides that default, but this
option is not used in the kernel). Since *.a archive files are used in
kbuild, I have also verified that their use does not break the
assumption; they are always created from scratch.
In short, to enforce an ordering, you can simply list the corresponding
object files in that order in the Makefile. Of course, add a big fat
warning comment, so people understand the order is not arbitrary.
Not everyone builds custom kernels.
Sorry, I didn't understand your comment.
Most people run a disto supplied kernel. If the LSM ordering were determined
only at compile time you could never run a kernel that omitted an LSM.
Ah, ok. We are talking about the LSMs with order LSM_ORDER_LAST which
are always enabled and the last.
This is the code in security.c to handle them:
/* LSM_ORDER_LAST is always last. */
for (lsm = __start_lsm_info; lsm < __end_lsm_info; lsm++) {
if (lsm->order == LSM_ORDER_LAST)
append_ordered_lsm(lsm, " last");
}
Those LSMs are not affected by lsm= in the kernel command line, or the
order in the kernel configuration (those are the mutable LSMs).
In this case, clearly, what matters is how LSMs are stored in the
.lsm_info.init section. See the DEFINE_LSM() macro:
#define DEFINE_LSM(lsm) \
static struct lsm_info __lsm_##lsm \
__used __section(".lsm_info.init") \
__aligned(sizeof(unsigned long))
With Petr, we started to wonder if somehow the order in which LSMs are
placed in this section is deterministic. I empirically tried to swap the
order in which IMA and EVM are compiled in the Makefile, and that led to
'evm' being placed in the LSM list before 'ima'.
The question is if this behavior is deterministic, or there is a case
where 'evm' is before 'ima', despite they are in the inverse order in
the Makefile.
Petr looked at the kernel linking process, which is relevant for the
order of LSMs in the .lsm_info.init section, and he found that the order
in the section always corresponds to the order in the Makefile.
Thanks
Roberto
Everyone builds the kernel, also Linux distros. What Petr was
suggesting was that it does not matter how you build the kernel, the
linker will place the LSMs in the order they appear in the Makefile.
And for this particular case, we have:
obj-$(CONFIG_IMA) += ima/
obj-$(CONFIG_EVM) += evm/
In the past, I also verified that swapping these two resulted in the
swapped order of LSMs. Petr confirmed that it would always happen.
LSM execution order is not based on compilation order. It is specified
by CONFIG_LSM, and may be modified by the LSM_ORDER value. I don't
understand why the linker is even being brought into the discussion.
Thanks
Roberto
