Thanks for replying. ;)


On 3/2/20 6:36 PM, Nicolai Stange wrote:

Hi,

JeffleXu [off-list ref] writes:

quoted

According to the model, there will be scenario where old function and new

function can co-exist, though for a single thread, it sees either all new

functions or all old functions.

That's correct.

quoted

I can't understand why Vojtech said that 'old func processing new
data' was impossible.

Just to make it explicit: Vojtech was talking about data layout
changes.

Fine

That is, consider you have something like e.g. this in the unmodified
kernel sources:

struct my_driver_work
{
	struct work_struct work;
         struct list_head works_list;
         void *some_payload;
};

In general, you can't change that struct definition from a live
patch. So simply extending it like this

struct my_driver_work
{
	struct work_struct work;
         struct list_head works_list;
         unsigned long flags /* New */
         void *some_payload;
};

won't work.

quoted

Assuming a scenario where a process calls func-A to submit a

work request (inserted into a global list), and then a kthread is
responsible

for calling func-B to process all work requests in the list. What if
this process

has finished the transition (sees new func-A) while kthread still sees
the old func-B?

Going with the example from above, the patched func-A would submit
instances of the new struct my_driver_work whereas the unpatched func-B
would expect the ->some_payload pointer at where ->flags is stored now,
which is bad, obviously.

In this specific example, it could perhaps be possible to make the
patched func-A associate a shadow variable corresponding to the new
->flags member with newly created struct my_driver_work instances (of
original, unmodified layout). Any unpatched func-B would obviously
ignore it and consider the shadow only when it becomes patched. It very
much depends on the specific situation whether or not this is
acceptable. If not, the ->post_patch() can sometimes be used to achieve
some notion of a "global consistency" state (in this context,
c.f. Documentation/livepatch/system-state.rst).

Well, I'm familiar with shadow variable, but didn't consider callbacks 
earlier. Since the version of my

kernel is not new enough, the "system state API" has not been merged in 
my kernel. I will read it later.

Note however, that the patched func-B must always be able to handle
the situation where a struct my_driver_work instance does not have such
a ->flags shadow attached to it, either because the instance had been
created  when the live patch has not been applied at all or because it
has been submitted from a not yet transitioned func-A.

There's another subtlety: the deallocation code for struct
my_driver_work needs to get livepatched as well to make it free the
->flags shadow. Consider the case where func-A has been transitioned,
but the deallocation code hasn't yet. Without any extra measures in
func-A, it could happen that a stale ->flags shadow from a deallocated
struct my_driver_work gets (wrongly) reassociated with a fresh struct
my_driver_work instance allocated at the same address as the old one
(because shadow variables are keyed on addresses of the data they're
associated with). Sometimes that's acceptable, sometimes it's not. In
the latter case you probably had to check for this situation and work
around it in the allocation code, i.e. the live-patched func-A.

I've never thought about this. It's a valuable suggestion.

Finally, let me remark that from my experience, most CVEs (>95%) can be
fixed via live patching without having to resort to either of shadow
variables, callbacks or the state API. For the rest, things usually tend
to become really non-trivial, hackish and subtle.

Thanks for your experience.

quoted

In this case, old func-B has to process new data. If there's some lock
semantic

changes in func-A and func-B, then old func-B has no way identifying
the shadow

variable labeled by new func-A.

I don't understand what you mean by "variable labeled by new func-A"?
Anyway, it's correct that an unpatched func-B would not consider any
shadow variables instantiated by patched func-A. And it's also correct
that changing locking semantics is difficult, if not impossible.

Just means shadow variable allocated by new patched function. I know 
shadow variable can serve

as a flag to enable the new functions, but in this case the post_patch() 
callback is obviously more

appropriate to serve as this role.


So as far as I understand, for all kinds of (data/locking) semantic 
changes, it's the responsibility of the

patch writer to detect the semantic changes, and usually it can only be 
analyzed case by case. Right?