On Thu, Jul 28, 2022 at 4:42 AM David Gow [off-list ref] wrote:

On Thu, Jul 21, 2022 at 4:29 AM Luis Chamberlain [off-list ref] wrote:

quoted

On Wed, Jul 20, 2022 at 05:26:02PM +0800, David Gow wrote:

quoted

On Wed, Jul 20, 2022 at 1:58 AM Luis Chamberlain [off-list ref] wrote:

quoted

On Wed, Jul 13, 2022 at 08:24:32AM -0700, Daniel Latypov wrote:

quoted

On Tue, Jul 12, 2022 at 5:52 PM David Gow [off-list ref] wrote:

quoted

The new KUnit module handling has KUnit test suites listed in a
.kunit_test_suites section of each module. This should be loaded when
the module is, but at the moment this only happens if KUnit is built-in.

Also load this when KUnit is enabled as a module: it'll not be usable
unless KUnit is loaded, but such modules are likely to depend on KUnit
anyway, so it's unlikely to ever be loaded needlessly.

This seems reasonable to me.

Question: what happens in this case?
1. insmod <test-module>
2. insmod kunit
3. rmmod <test-module>

I think on 3, we'll call the cleanup code, __kunit_test_suites_exit(),
for <test-module>, I think?
But we never called __kunit_test_suites_init().
My fear is what breaks as a result of this precondition break.

I don't think this should be possible: any module with KUnit tests
will depend on the 'kunit' module (or, at least, kunit symbols), so
shouldn't load without kunit already present.

If modprobe is used, kunit will automatically be loaded. If insmod is
used directly, loading the first module should error out with
something like:
[   82.393629] list_test: loading test module taints kernel.
[   82.409607] list_test: Unknown symbol kunit_binary_ptr_assert_format (err -2)
[   82.409657] list_test: Unknown symbol kunit_do_failed_assertion (err -2)
[   82.409799] list_test: Unknown symbol kunit_binary_assert_format (err -2)
[   82.409820] list_test: Unknown symbol kunit_unary_assert_format (err -2)
insmod: ERROR: could not insert module
/lib/modules/5.19.0-rc1-15284-g9ec67db0c271/kernel/lib/list-test.ko:
Unknown symbol in module

This can be fixed with a request_module() call. And since this is a
generic requirement, you can have the wrappers do it for you.

I'm not convinced that this is worth the trouble, particularly since
KUnit needs to be loaded already before any test-specific code in a
module is run. _Maybe_ we could put it in the code which looks for the
.kunit_test_suites section, but even then it seems like a bit of an
ugly hack.

Personally, I'm not particularly concerned about test modules failing
to load if KUnit isn't already present -- if people want all of a
module's dependencies loaded, that's what modprobe is for.

That being said, if you feel particularly strongly about it, this is
something we can look at. Let's do so in a separate patch though: this
one does fix a regression as-is.

I agree. We need a fix for 3d6e44623841 to go in for 5.20 - we've
gotten several complaints about it. This patch seems to accomplish
that.

I am not an expert on the module stuff by any means, so I am
absolutely open to continuing this discussion in a follow-up patch,
but I think we need this fix now.

If no one objects, I am going to ask Shuah to take this patch.

quoted

quoted

Maybe you could get into some trouble by force-removing modules at
various points, but you're in undefined behaviour generally at that
point, so I don't think there's much point going out-of-our-way to try
to support that.

You can prevent that by refcounting the kunit module / symbols, by each test.

Again, I don't think KUnit is any more special than any other module
here. I don't think we need to do this ourselves, as it shouldn't be
possible to remove kunit without first removing any dependent modules.

Of course, happy to look into this again if anyone can come up with an
actual crash, but I'd rather get this fix in first. At the very least,
this patch shouldn't introduce any _new_ issues.

Sounds good. I will send my Reviewed-by in a separate email, as per usual.

Cheers