Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Thu, Nov 3, 2022 at 11:23 PM Mauro Carvalho Chehab wrote: > > Hi, > > I'm facing a couple of issues when testing KUnit with the i915 driver. > > The DRM subsystem and the i915 driver has, for a long time, his own > way to do unit tests, which seems to be added before KUnit. > > I'm now checking if it is worth start using KUnit at i915. So, I wrote > a RFC with some patches adding support for the tests we have to be > reported using Kernel TAP and KUnit. Thanks very much for looking into this, and sorry for the delayed response (I've been out sick). I think Daniel has answered most of your questions (thanks, Daniel), and I agree with pretty much everything he's said. In short, I think that it'd be great to have the i915 tests use KUnit where appropriate, and even where KUnit isn't the ideal tool, using KTAP as a result format would be great. I definitely think that there's a whole bunch of areas of i915 for which KUnit makes sense: the more hardware independent unit tests (things like swizzling/tiling, maybe some command-buffer creation / validation, "utility" functions generally) are an obvious option. If KUnit isn't working for those sorts of tests, that's clearly a deficiency in KUnit that we'll want to rectify (though it might take some time to do so). The more hardware-specific stuff probably isn't as good a fit for KUnit, but if using KUnit is the easiest way to do test management/assertion macros/KTAP output/etc., then it may be worth using whatever parts of it make sense. I'd prefer it if any tests which depend strongly on specific hardware were marked as such, and maybe lived under a different Kconfig option (which might not be auto-enabled by KUNIT_ALL_TESTS). Though as long as the tests are skipped if the hardware isn't present (which seems to be the case from running them under qemu), it's not a real problem to have them. It's not something we plan to "officially support", though, so if the requirements of hardware-specific tests and more traditional unit tests conflict, KUnit will lean towards supporting the non-hardware-specific ones. > > There are basically 3 groups of tests there: > > - mock tests - check i915 hardware-independent logic; > - live tests - run some hardware-specific tests; > - perf tests - check perf support - also hardware-dependent. > > As they depend on i915 driver, they run only on x86, with PCI > stack enabled, but the mock tests run nicely via qemu. > > The live and perf tests require a real hardware. As we run them > together with our CI, which, among other things, test module > unload/reload and test loading i915 driver with different > modprobe parameters, the KUnit tests should be able to run as > a module. > > While testing KUnit, I noticed a couple of issues: > > 1. kunit.py parser is currently broken when used with modules > > the parser expects "TAP version xx" output, but this won't > happen when loading the kunit test driver. > > Are there any plans or patches fixing this issue? > Yeah: this is on our to-do list to fix, hopefully pretty soon. > 2. current->mm is not initialized > > Some tests do mmap(). They need the mm user context to be initialized, > but this is not happening right now. > > Are there a way to properly initialize it for KUnit? > This is something we've hit before and don't have a good solution for (as you've found out). I'm not an expert on the mm subsystem, so while it's something we want to support, I don't think anyone quite knows how yet. As a totally wild, untested guess, you may have some luck setting current->mm = current->active_mm, or current->mm = _mm? It's definitely true that, even when loaded from modules, current->mm won't be set as KUnit tests run in their own kthreads. Maybe setting mm = active_mm would let us carry that context with us from the module loader to the test... In any case, you're not the only person to hit this issue, so it's definitely something we'd like to work out. > 3. there's no test filters for modules > > In order to be able to do proper CI automation, it is needed to > be able to control what tests will run or not. That's specially > interesting at development time where some tests may not apply > or not run properly on new hardware. > > Are there any plans to add support for it at kunit_test_suites() > when the driver is built as module? Ideally, the best would be to > export a per-module filter_glob parameter on such cases. > Again, this is on the to-do list. It may be implemented as a global property which affects future module loads (and might be able to be changed via, e.g., debugfs), rather than a per-module parameter, but we haven't designed it yet. Alas, module support has always seen a little less love than the built-in UML/qemu-based mode, so it does tend to lag behind a little bit with these sort of features, and tends to be tested less well. Hopefully we can bring it up to scratch soon. > 4. there are actually 3 levels of tests on i915: > - Level 1: mock, live, perf
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Mon, 7 Nov 2022 15:16:17 -0800
Daniel Latypov wrote:
> On Mon, Nov 7, 2022 at 10:38 AM Michał Winiarski
> wrote:
> >
> > On Thu, Nov 03, 2022 at 04:23:02PM +0100, Mauro Carvalho Chehab wrote:
> > > Hi,
> > >
> > > I'm facing a couple of issues when testing KUnit with the i915 driver.
> > >
> > > The DRM subsystem and the i915 driver has, for a long time, his own
> > > way to do unit tests, which seems to be added before KUnit.
> > >
> > > I'm now checking if it is worth start using KUnit at i915. So, I wrote
> > > a RFC with some patches adding support for the tests we have to be
> > > reported using Kernel TAP and KUnit.
> > >
> > > There are basically 3 groups of tests there:
> > >
> > > - mock tests - check i915 hardware-independent logic;
> > > - live tests - run some hardware-specific tests;
> > > - perf tests - check perf support - also hardware-dependent.
> > >
> > > As they depend on i915 driver, they run only on x86, with PCI
> > > stack enabled, but the mock tests run nicely via qemu.
> > >
> > > The live and perf tests require a real hardware. As we run them
> > > together with our CI, which, among other things, test module
> > > unload/reload and test loading i915 driver with different
> > > modprobe parameters, the KUnit tests should be able to run as
> > > a module.
> >
> > Note that KUnit tests that are doing more of a functional/integration
> > testing (on "live" hardware) rather than unit testing (where hardware
> > interactions are mocked) are not very common.
> > Do we have other KUnit tests like this merged?
>
> I don't think we have other tests like this.
>
> > Some of the "live tests" are not even that, being more of a pure
> > hardware tests (e.g. live_workarounds, which is checking whether values
> > in MMIO regs stick over various HW state transitions).
> >
> > I'm wondering, is KUnit the right tool for this job?
>
> The main focus of KUnit is for hw-independent tests.
> So in theory: no.
>
> But I can imagine it could be easier to write the validation via
> KUNIT_EXPECT_EQ and friends as opposed to writing your own kernel
> module w/ its own set of macros, etc.
Right now, i915 has its own way of doing that, for both hw-independent
and live tests. The current patches are keeping them, because it helps
comparing both approaches, and won't disrupt CI jobs.
However, if we're willing to merge KUnit support, the best would be
to use the same macros/logic for both, as having two different unit
test frameworks used on i915 doesn't make much sense.
So yeah, using KUNIT_EXPECT_EQ and friends for live unit and hw tests
makes sense to me.
The main difference between i915 selftest and KUnit seems to be at
the way the tests are started:
- i915 selftest uses module parameters to run the driver on test mode.
When such parameters are enabled, the probe() function will run the
tests.
- from what I understood, KUnit uses a .kunit_test_suites code
section, which does an early initialization of the tests, calling
the test suite a lot earlier than probe().
Due to such differences, running KUnit against a real hardware requires
to load the real driver first, letting it probe the hardware. Once the
driver is loaded and have the hardware properly initialized, load a
separate KUnit module that would call the test functions from the driver.
That's the approach took on this series.
It sounds possible to merge both approaches, by adding some helper
functions similar to what kunit_test_suites() do, but, instead of
using a separate segment, run the tests at probe() time.
That would mean a cleaner solution with the usage of all KUnit
macros and, in thesis, it will be equivalent to what i915 selftest
already does.
Another possibility would be to make sure that kunit_test_suites()
will fully initiate the user context, allowing mmap() to work.
> So my first thought is: "if it works, then you can try using it."
> (Might want to take steps like make sure they don't get enabled by
> CONFIG_KUNIT_ALL_TESTS=y).
Yeah, it makes sense to have hw-dependent tests not enabled with
KUNIT_ALL_TESTS. Another approach would be to just skip them when
hw is required[1]. Skipping the tests is needed anyway, as the
hardware probing may fail.
[1] Btw, despite the comments:
/**
* kunit_skip() - Marks @test_or_suite as skipped
*
* @test_or_suite: The test context object.
* @fmt: A printk() style format string.
*
* Skips the test. @fmt is given output as the test status
* comment, typically the reason the test was skipped.
*
* Test execution is halted after kunit_skip() is called.
*/
kunit_skip() doesn't work inside .suite_init, so this would cause
compilation errors:
static int i915_pci_init_suite(struct kunit_suite *suite)
{
if (!hw)
kunit_skip(suite);
}
> Talking with David, he seems to have echoed my thoughts.
> David also
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Fri, Nov 4, 2022 at 12:20 PM Daniel Latypov wrote: > rm...@google.com was working on them and should hopefully be able to > send them out real soon. > You should get CC'd on those. > > I think the follow-up work is just crafting an example parser input > file and iterating until > $ ./tools/testing/kunit/kunit.py parse < /tmp/example_input > produces our desired results. I briefly played around with this on top of https://lore.kernel.org/linux-kselftest/20221104194705.3245738-1-rm...@google.com/ But instead, I tried to make it so kunit.py parse could be just given a directory as input so we don't have to do `find | xargs cat`. I can't get module tests to work locally, so I'll replicate the setup with some regular files. Say we have /tmp/results with two files: file1 contains # Subtest: suite1 KTAP version 1 1..1 ok 1 test ok 1 suite1 and file2 contains # Subtest: suite2 KTAP version 1 1..1 ok 1 test ok 1 suite2 I think this should match what module output in debugfs would look like. A key thing to note: the top level test counter resets between modules, iirc. The patch below makes this almost work $ ./tools/testing/kunit/kunit.py parse /tmp/results/ suite2 (1 subtest) [PASSED] test = [PASSED] suite2 == suite1 (1 subtest) [PASSED] test [ERROR] Test: suite1: Expected test number 2 but found 1 = [PASSED] suite1 == Testing complete. Ran 2 tests: passed: 2, errors: 1 But there's a few issues a. the error about multiple top level tests with "test number 1" b. how do we make this handle kernel output with prefixes (timestamps) c. and what happens if files each have different length prefixes? diff --git a/tools/testing/kunit/kunit.py b/tools/testing/kunit/kunit.py index 4d4663fb578b..c5b2eb416c2d 100755 --- a/tools/testing/kunit/kunit.py +++ b/tools/testing/kunit/kunit.py @@ -189,6 +189,14 @@ def _map_to_overall_status(test_status: kunit_parser.TestStatus) -> KunitStatus: return KunitStatus.SUCCESS return KunitStatus.TEST_FAILURE +def _kernel_output_from_dir(dir: str) -> Iterable[str]: + yield 'KTAP version 1' + for root, dirs, files in os.walk(dir): + for sub in files: + with open(os.path.join(root, sub), errors='backslashreplace') as f: + for line in f: + yield line + def parse_tests(request: KunitParseRequest, metadata: kunit_json.Metadata, input_data: Iterable[str]) -> Tuple[KunitResult, kunit_parser.Test]: parse_start = time.time() @@ -496,6 +504,8 @@ def main(argv): if cli_args.file is None: sys.stdin.reconfigure(errors='backslashreplace') # pytype: disable=attribute-error kunit_output = sys.stdin + elif os.path.isdir(cli_args.file): + kunit_output = _kernel_output_from_dir(cli_args.file) else: with open(cli_args.file, 'r', errors='backslashreplace') as f: kunit_output = f.read().splitlines()
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Mon, Nov 7, 2022 at 10:38 AM Michał Winiarski wrote: > > On Thu, Nov 03, 2022 at 04:23:02PM +0100, Mauro Carvalho Chehab wrote: > > Hi, > > > > I'm facing a couple of issues when testing KUnit with the i915 driver. > > > > The DRM subsystem and the i915 driver has, for a long time, his own > > way to do unit tests, which seems to be added before KUnit. > > > > I'm now checking if it is worth start using KUnit at i915. So, I wrote > > a RFC with some patches adding support for the tests we have to be > > reported using Kernel TAP and KUnit. > > > > There are basically 3 groups of tests there: > > > > - mock tests - check i915 hardware-independent logic; > > - live tests - run some hardware-specific tests; > > - perf tests - check perf support - also hardware-dependent. > > > > As they depend on i915 driver, they run only on x86, with PCI > > stack enabled, but the mock tests run nicely via qemu. > > > > The live and perf tests require a real hardware. As we run them > > together with our CI, which, among other things, test module > > unload/reload and test loading i915 driver with different > > modprobe parameters, the KUnit tests should be able to run as > > a module. > > Note that KUnit tests that are doing more of a functional/integration > testing (on "live" hardware) rather than unit testing (where hardware > interactions are mocked) are not very common. > Do we have other KUnit tests like this merged? I don't think we have other tests like this. > Some of the "live tests" are not even that, being more of a pure > hardware tests (e.g. live_workarounds, which is checking whether values > in MMIO regs stick over various HW state transitions). > > I'm wondering, is KUnit the right tool for this job? The main focus of KUnit is for hw-independent tests. So in theory: no. But I can imagine it could be easier to write the validation via KUNIT_EXPECT_EQ and friends as opposed to writing your own kernel module w/ its own set of macros, etc. So my first thought is: "if it works, then you can try using it." (Might want to take steps like make sure they don't get enabled by CONFIG_KUNIT_ALL_TESTS=y). Talking with David, he seems to have echoed my thoughts. David also suggested that maybe the test could use a fake of the hw by default, but have an option to run against real hw when available. I think that sounds like a good chunk of work, so I don't know if you need to worry about that. Daniel
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Thu, Nov 03, 2022 at 04:23:02PM +0100, Mauro Carvalho Chehab wrote: > Hi, > > I'm facing a couple of issues when testing KUnit with the i915 driver. > > The DRM subsystem and the i915 driver has, for a long time, his own > way to do unit tests, which seems to be added before KUnit. > > I'm now checking if it is worth start using KUnit at i915. So, I wrote > a RFC with some patches adding support for the tests we have to be > reported using Kernel TAP and KUnit. > > There are basically 3 groups of tests there: > > - mock tests - check i915 hardware-independent logic; > - live tests - run some hardware-specific tests; > - perf tests - check perf support - also hardware-dependent. > > As they depend on i915 driver, they run only on x86, with PCI > stack enabled, but the mock tests run nicely via qemu. > > The live and perf tests require a real hardware. As we run them > together with our CI, which, among other things, test module > unload/reload and test loading i915 driver with different > modprobe parameters, the KUnit tests should be able to run as > a module. Note that KUnit tests that are doing more of a functional/integration testing (on "live" hardware) rather than unit testing (where hardware interactions are mocked) are not very common. Do we have other KUnit tests like this merged? Some of the "live tests" are not even that, being more of a pure hardware tests (e.g. live_workarounds, which is checking whether values in MMIO regs stick over various HW state transitions). I'm wondering, is KUnit the right tool for this job? -Michał
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Thu, Nov 3, 2022 at 8:23 AM Mauro Carvalho Chehab wrote: > > Hi, > > I'm facing a couple of issues when testing KUnit with the i915 driver. > > The DRM subsystem and the i915 driver has, for a long time, his own > way to do unit tests, which seems to be added before KUnit. > > I'm now checking if it is worth start using KUnit at i915. So, I wrote > a RFC with some patches adding support for the tests we have to be > reported using Kernel TAP and KUnit. > > There are basically 3 groups of tests there: > > - mock tests - check i915 hardware-independent logic; > - live tests - run some hardware-specific tests; > - perf tests - check perf support - also hardware-dependent. > > As they depend on i915 driver, they run only on x86, with PCI > stack enabled, but the mock tests run nicely via qemu. > > The live and perf tests require a real hardware. As we run them > together with our CI, which, among other things, test module > unload/reload and test loading i915 driver with different > modprobe parameters, the KUnit tests should be able to run as > a module. > > While testing KUnit, I noticed a couple of issues: > > 1. kunit.py parser is currently broken when used with modules > > the parser expects "TAP version xx" output, but this won't > happen when loading the kunit test driver. > > Are there any plans or patches fixing this issue? Partially. Note: we need a header to look for so we can strip prefixes (like timestamps). But there is a patch in the works to add a TAP header for each subtest, hopefully in time for 6.2. This is to match the KTAP spec: https://kernel.org/doc/html/latest/dev-tools/ktap.html That should fix it so you can parse one suite's results at a time. I'm pretty sure it won't fix the case where there's multiple suites and/or you're trying to parse all test results at once via $ find /sys/kernel/debug/kunit/ -type f | xargs cat | ./tools/testing/kunit/kunit.py parse I think that in-kernel code change + some more python changes could make the above command work, but no one has actively started looking at that just yet. Hopefully we can pick this up and also get it done for 6.2 (unless I'm underestimating how complicated this is). > > 2. current->mm is not initialized > > Some tests do mmap(). They need the mm user context to be initialized, > but this is not happening right now. > > Are there a way to properly initialize it for KUnit? Right, this is a consequence of how early built-in KUnit tests are run after boot. I think for now, the answer is to make the test module-only. I know David had some ideas here, but I can't speak to them. > > 3. there's no test filters for modules > > In order to be able to do proper CI automation, it is needed to > be able to control what tests will run or not. That's specially > interesting at development time where some tests may not apply > or not run properly on new hardware. > > Are there any plans to add support for it at kunit_test_suites() > when the driver is built as module? Ideally, the best would be to > export a per-module filter_glob parameter on such cases. I think this is a good idea and is doable. (I think I said as much on the other thread). The thinking before was that people would make group tests together in modules. But if you want to share a single module for many tests, this becomes more useful. This has some potential merge conflicts w/ other pending work. I was also prototyping the ability to tell KUnit "run tests #2 - #5", so that also touches the filtering code very heavily. (The goal there is to have kunit.py able to shard up tests and boot multiple kernels concurrently.) > > 4. there are actually 3 levels of tests on i915: > - Level 1: mock, live, perf > - Level 2: test group (mmap, fences, ...) > - Level 3: unit tests > > Currently, KUnit seems to have just two levels (test suite and tests). > Are there a way to add test groups there? Parameterized tests are the closest we have to a third-level of tests. But other than that, the answer is no. I'd need to get more familiar with the existing tests, but I'm pretty sure parameters won't work for you. And I don't know if this will get done. Note: the kunit_parser.py code should be able to handle arbitrary levels of tests in the output. This restriction is purely in the in-kernel code. I had brought up the idea of more layers of tests before. It would also be useful for a) sharing expensive setup between multiple tests b) allowing more granular scope for cleanups (kunit_kmalloc and others) c) more flexibility in dynamically generating subtests than parameterized testing There's some precedent in other unit testing frameworks, for example: https://pkg.go.dev/testing#T.Run Daniel
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Fri, Nov 4, 2022 at 12:50 AM Mauro Carvalho Chehab
wrote:
>
> On Thu, 3 Nov 2022 15:43:26 -0700
> Daniel Latypov wrote:
>
> > On Thu, Nov 3, 2022 at 8:23 AM Mauro Carvalho Chehab
> > wrote:
> > >
> > > Hi,
> > >
> > > I'm facing a couple of issues when testing KUnit with the i915 driver.
> > >
> > > The DRM subsystem and the i915 driver has, for a long time, his own
> > > way to do unit tests, which seems to be added before KUnit.
> > >
> > > I'm now checking if it is worth start using KUnit at i915. So, I wrote
> > > a RFC with some patches adding support for the tests we have to be
> > > reported using Kernel TAP and KUnit.
> > >
> > > There are basically 3 groups of tests there:
> > >
> > > - mock tests - check i915 hardware-independent logic;
> > > - live tests - run some hardware-specific tests;
> > > - perf tests - check perf support - also hardware-dependent.
> > >
> > > As they depend on i915 driver, they run only on x86, with PCI
> > > stack enabled, but the mock tests run nicely via qemu.
> > >
> > > The live and perf tests require a real hardware. As we run them
> > > together with our CI, which, among other things, test module
> > > unload/reload and test loading i915 driver with different
> > > modprobe parameters, the KUnit tests should be able to run as
> > > a module.
> > >
> > > While testing KUnit, I noticed a couple of issues:
> > >
> > > 1. kunit.py parser is currently broken when used with modules
> > >
> > > the parser expects "TAP version xx" output, but this won't
> > > happen when loading the kunit test driver.
> > >
> > > Are there any plans or patches fixing this issue?
> >
> > Partially.
> > Note: we need a header to look for so we can strip prefixes (like
> > timestamps).
> >
> > But there is a patch in the works to add a TAP header for each
> > subtest, hopefully in time for 6.2.
>
> Good to know.
>
> > This is to match the KTAP spec:
> > https://kernel.org/doc/html/latest/dev-tools/ktap.html
>
> I see.
>
> > That should fix it so you can parse one suite's results at a time.
> > I'm pretty sure it won't fix the case where there's multiple suites
> > and/or you're trying to parse all test results at once via
> >
> > $ find /sys/kernel/debug/kunit/ -type f | xargs cat |
> > ./tools/testing/kunit/kunit.py parse
>
> Could you point me to the changeset? perhaps I can write a followup
> patch addressing this case.
rm...@google.com was working on them and should hopefully be able to
send them out real soon.
You should get CC'd on those.
I think the follow-up work is just crafting an example parser input
file and iterating until
$ ./tools/testing/kunit/kunit.py parse < /tmp/example_input
produces our desired results.
>
> > I think that in-kernel code change + some more python changes could
> > make the above command work, but no one has actively started looking
> > at that just yet.
> > Hopefully we can pick this up and also get it done for 6.2 (unless I'm
> > underestimating how complicated this is).
> >
> > >
> > > 2. current->mm is not initialized
> > >
> > > Some tests do mmap(). They need the mm user context to be initialized,
> > > but this is not happening right now.
> > >
> > > Are there a way to properly initialize it for KUnit?
> >
> > Right, this is a consequence of how early built-in KUnit tests are run
> > after boot.
> > I think for now, the answer is to make the test module-only.
> >
> > I know David had some ideas here, but I can't speak to them.
>
> This is happening when test-i915 is built as module as well.
Oh, I didn't expect that at all.
>
> I suspect that the function which initializes it is mm_alloc() inside
> kernel/fork.c:
>
> struct mm_struct *mm_alloc(void)
> {
> struct mm_struct *mm;
>
> mm = allocate_mm();
> if (!mm)
> return NULL;
>
> memset(mm, 0, sizeof(*mm));
> return mm_init(mm, current, current_user_ns());
> }
>
> As modprobing a test won't fork until all tests run, this never runs.
>
> It seems that the normal usage is at fs/exec.c:
>
> fs/exec.c: bprm->mm = mm = mm_alloc();
>
> but other places also call it:
>
> arch/arm/mach-rpc/ecard.c: struct mm_struct * mm = mm_alloc();
> drivers/dma-buf/dma-resv.c: struct mm_struct *mm = mm_alloc();
> include/linux/sched/mm.h:extern struct mm_struct *mm_alloc(void);
> mm/debug_vm_pgtable.c: args->mm = mm_alloc();
>
> Probably the solution would be to call it inside kunit executor code,
> adding support for modules to use it.
I know basically nothing about the mm code.
I think I vaguely recall there being issues with this on UML or
something, but I could be totally wrong.
I'll wait for David to chime in when he can.
>
> > > 3. there's no test filters for modules
> > >
> > > In order to be able to do proper CI automation, it is needed to
> > > be able to control what tests will run or not. That's specially
> > >
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Fri, 4 Nov 2022 08:49:55 +0100
Mauro Carvalho Chehab wrote:
> On Thu, 3 Nov 2022 15:43:26 -0700
> Daniel Latypov wrote:
>
> > On Thu, Nov 3, 2022 at 8:23 AM Mauro Carvalho Chehab
> > wrote:
> > >
> > > Hi,
> > >
> > > I'm facing a couple of issues when testing KUnit with the i915 driver.
> > >
> > > The DRM subsystem and the i915 driver has, for a long time, his own
> > > way to do unit tests, which seems to be added before KUnit.
> > >
> > > I'm now checking if it is worth start using KUnit at i915. So, I wrote
> > > a RFC with some patches adding support for the tests we have to be
> > > reported using Kernel TAP and KUnit.
> > >
> > > There are basically 3 groups of tests there:
> > >
> > > - mock tests - check i915 hardware-independent logic;
> > > - live tests - run some hardware-specific tests;
> > > - perf tests - check perf support - also hardware-dependent.
> > >
> > > As they depend on i915 driver, they run only on x86, with PCI
> > > stack enabled, but the mock tests run nicely via qemu.
> > >
> > > The live and perf tests require a real hardware. As we run them
> > > together with our CI, which, among other things, test module
> > > unload/reload and test loading i915 driver with different
> > > modprobe parameters, the KUnit tests should be able to run as
> > > a module.
> > >
> > > While testing KUnit, I noticed a couple of issues:
> > >
> > > 1. kunit.py parser is currently broken when used with modules
> > >
> > > the parser expects "TAP version xx" output, but this won't
> > > happen when loading the kunit test driver.
> > >
> > > Are there any plans or patches fixing this issue?
> >
> > Partially.
> > Note: we need a header to look for so we can strip prefixes (like
> > timestamps).
> >
> > But there is a patch in the works to add a TAP header for each
> > subtest, hopefully in time for 6.2.
>
> Good to know.
>
> > This is to match the KTAP spec:
> > https://kernel.org/doc/html/latest/dev-tools/ktap.html
>
> I see.
>
> > That should fix it so you can parse one suite's results at a time.
> > I'm pretty sure it won't fix the case where there's multiple suites
> > and/or you're trying to parse all test results at once via
> >
> > $ find /sys/kernel/debug/kunit/ -type f | xargs cat |
> > ./tools/testing/kunit/kunit.py parse
>
> Could you point me to the changeset? perhaps I can write a followup
> patch addressing this case.
>
> > I think that in-kernel code change + some more python changes could
> > make the above command work, but no one has actively started looking
> > at that just yet.
> > Hopefully we can pick this up and also get it done for 6.2 (unless I'm
> > underestimating how complicated this is).
> >
> > >
> > > 2. current->mm is not initialized
> > >
> > > Some tests do mmap(). They need the mm user context to be initialized,
> > > but this is not happening right now.
> > >
> > > Are there a way to properly initialize it for KUnit?
> >
> > Right, this is a consequence of how early built-in KUnit tests are run
> > after boot.
> > I think for now, the answer is to make the test module-only.
> >
> > I know David had some ideas here, but I can't speak to them.
>
> This is happening when test-i915 is built as module as well.
>
> I suspect that the function which initializes it is mm_alloc() inside
> kernel/fork.c:
>
> struct mm_struct *mm_alloc(void)
> {
> struct mm_struct *mm;
>
> mm = allocate_mm();
> if (!mm)
> return NULL;
>
> memset(mm, 0, sizeof(*mm));
> return mm_init(mm, current, current_user_ns());
> }
>
> As modprobing a test won't fork until all tests run, this never runs.
>
> It seems that the normal usage is at fs/exec.c:
>
> fs/exec.c: bprm->mm = mm = mm_alloc();
>
> but other places also call it:
>
> arch/arm/mach-rpc/ecard.c: struct mm_struct * mm = mm_alloc();
> drivers/dma-buf/dma-resv.c: struct mm_struct *mm = mm_alloc();
> include/linux/sched/mm.h:extern struct mm_struct *mm_alloc(void);
> mm/debug_vm_pgtable.c: args->mm = mm_alloc();
>
> Probably the solution would be to call it inside kunit executor code,
> adding support for modules to use it.
Hmm... it is not that simple... I tried the enclosed patch, but it caused
another issue at the live/mman/mmap test:
[ 152.815543] test_i915: :00:02.0: it is a i915 device.
[ 152.816456] # Subtest: i915 live selftests
[ 152.816463] 1..1
[ 152.816835] kunit_try_run_case: allocating user context
[ 152.816978] CPU: 1 PID: 1139 Comm: kunit_try_catch Tainted: G
N 6.1.0-rc2-drm-110e9bebcbcc+ #20
[ 152.817063] Hardware name: Intel Corporation Tiger Lake Client
Platform/TigerLake Y LPDDR4x T4 Crb, BIOS TGLSFWI1.R00.3243.A01.2006102133
06/10/2020
[ 152.817583] i915: Performing live_mman selftests with
st_random_seed=0x11aaba4d st_timeout=500
[ 152.817735] test_i915: Setting dangerous
Re: [Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
On Thu, 3 Nov 2022 15:43:26 -0700
Daniel Latypov wrote:
> On Thu, Nov 3, 2022 at 8:23 AM Mauro Carvalho Chehab
> wrote:
> >
> > Hi,
> >
> > I'm facing a couple of issues when testing KUnit with the i915 driver.
> >
> > The DRM subsystem and the i915 driver has, for a long time, his own
> > way to do unit tests, which seems to be added before KUnit.
> >
> > I'm now checking if it is worth start using KUnit at i915. So, I wrote
> > a RFC with some patches adding support for the tests we have to be
> > reported using Kernel TAP and KUnit.
> >
> > There are basically 3 groups of tests there:
> >
> > - mock tests - check i915 hardware-independent logic;
> > - live tests - run some hardware-specific tests;
> > - perf tests - check perf support - also hardware-dependent.
> >
> > As they depend on i915 driver, they run only on x86, with PCI
> > stack enabled, but the mock tests run nicely via qemu.
> >
> > The live and perf tests require a real hardware. As we run them
> > together with our CI, which, among other things, test module
> > unload/reload and test loading i915 driver with different
> > modprobe parameters, the KUnit tests should be able to run as
> > a module.
> >
> > While testing KUnit, I noticed a couple of issues:
> >
> > 1. kunit.py parser is currently broken when used with modules
> >
> > the parser expects "TAP version xx" output, but this won't
> > happen when loading the kunit test driver.
> >
> > Are there any plans or patches fixing this issue?
>
> Partially.
> Note: we need a header to look for so we can strip prefixes (like timestamps).
>
> But there is a patch in the works to add a TAP header for each
> subtest, hopefully in time for 6.2.
Good to know.
> This is to match the KTAP spec:
> https://kernel.org/doc/html/latest/dev-tools/ktap.html
I see.
> That should fix it so you can parse one suite's results at a time.
> I'm pretty sure it won't fix the case where there's multiple suites
> and/or you're trying to parse all test results at once via
>
> $ find /sys/kernel/debug/kunit/ -type f | xargs cat |
> ./tools/testing/kunit/kunit.py parse
Could you point me to the changeset? perhaps I can write a followup
patch addressing this case.
> I think that in-kernel code change + some more python changes could
> make the above command work, but no one has actively started looking
> at that just yet.
> Hopefully we can pick this up and also get it done for 6.2 (unless I'm
> underestimating how complicated this is).
>
> >
> > 2. current->mm is not initialized
> >
> > Some tests do mmap(). They need the mm user context to be initialized,
> > but this is not happening right now.
> >
> > Are there a way to properly initialize it for KUnit?
>
> Right, this is a consequence of how early built-in KUnit tests are run
> after boot.
> I think for now, the answer is to make the test module-only.
>
> I know David had some ideas here, but I can't speak to them.
This is happening when test-i915 is built as module as well.
I suspect that the function which initializes it is mm_alloc() inside
kernel/fork.c:
struct mm_struct *mm_alloc(void)
{
struct mm_struct *mm;
mm = allocate_mm();
if (!mm)
return NULL;
memset(mm, 0, sizeof(*mm));
return mm_init(mm, current, current_user_ns());
}
As modprobing a test won't fork until all tests run, this never runs.
It seems that the normal usage is at fs/exec.c:
fs/exec.c: bprm->mm = mm = mm_alloc();
but other places also call it:
arch/arm/mach-rpc/ecard.c: struct mm_struct * mm = mm_alloc();
drivers/dma-buf/dma-resv.c: struct mm_struct *mm = mm_alloc();
include/linux/sched/mm.h:extern struct mm_struct *mm_alloc(void);
mm/debug_vm_pgtable.c: args->mm = mm_alloc();
Probably the solution would be to call it inside kunit executor code,
adding support for modules to use it.
> > 3. there's no test filters for modules
> >
> > In order to be able to do proper CI automation, it is needed to
> > be able to control what tests will run or not. That's specially
> > interesting at development time where some tests may not apply
> > or not run properly on new hardware.
> >
> > Are there any plans to add support for it at kunit_test_suites()
> > when the driver is built as module? Ideally, the best would be to
> > export a per-module filter_glob parameter on such cases.
>
> I think this is a good idea and is doable. (I think I said as much on
> the other thread).
>
> The thinking before was that people would make group tests together in
> modules.
> But if you want to share a single module for many tests, this becomes
> more useful.
At least for this RFC, I opted to place everything we have already on
a single module.
Perhaps I could create, instead, 3 separate modules. This way, I would gain
a "third level" and a poor man's way of filtering what test type
will run (mock,
[Intel-gfx] KUnit issues - Was: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL
Hi, I'm facing a couple of issues when testing KUnit with the i915 driver. The DRM subsystem and the i915 driver has, for a long time, his own way to do unit tests, which seems to be added before KUnit. I'm now checking if it is worth start using KUnit at i915. So, I wrote a RFC with some patches adding support for the tests we have to be reported using Kernel TAP and KUnit. There are basically 3 groups of tests there: - mock tests - check i915 hardware-independent logic; - live tests - run some hardware-specific tests; - perf tests - check perf support - also hardware-dependent. As they depend on i915 driver, they run only on x86, with PCI stack enabled, but the mock tests run nicely via qemu. The live and perf tests require a real hardware. As we run them together with our CI, which, among other things, test module unload/reload and test loading i915 driver with different modprobe parameters, the KUnit tests should be able to run as a module. While testing KUnit, I noticed a couple of issues: 1. kunit.py parser is currently broken when used with modules the parser expects "TAP version xx" output, but this won't happen when loading the kunit test driver. Are there any plans or patches fixing this issue? 2. current->mm is not initialized Some tests do mmap(). They need the mm user context to be initialized, but this is not happening right now. Are there a way to properly initialize it for KUnit? 3. there's no test filters for modules In order to be able to do proper CI automation, it is needed to be able to control what tests will run or not. That's specially interesting at development time where some tests may not apply or not run properly on new hardware. Are there any plans to add support for it at kunit_test_suites() when the driver is built as module? Ideally, the best would be to export a per-module filter_glob parameter on such cases. 4. there are actually 3 levels of tests on i915: - Level 1: mock, live, perf - Level 2: test group (mmap, fences, ...) - Level 3: unit tests Currently, KUnit seems to have just two levels (test suite and tests). Are there a way to add test groups there? Regards, Mauro Forwarded message: Date: Thu, 3 Nov 2022 14:51:38 + From: Mauro Carvalho Chehab To: Cc: Thomas Hellström , linux-kselft...@vger.kernel.org, Michał Winiarski , dri-de...@lists.freedesktop.org, intel-gfx@lists.freedesktop.org, Daniel Latypov , linux-ker...@vger.kernel.org, igt-...@lists.freedesktop.org, Matthew Auld , Daniel Vetter , Rodrigo Vivi , sk...@linuxfoundation.org, Isabella Basso , David Airlie , Christian König Subject: [igt-dev] [PATCH RFC v2 8/8] drm/i915: check if current->mm is not NULL The mmap tests require mm in order to work. Failing to do that will cause a crash: [ 316.820722] BUG: kernel NULL pointer dereference, address: 00e8 [ 316.822517] #PF: supervisor write access in kernel mode [ 316.823430] #PF: error_code(0x0002) - not-present page [ 316.824390] PGD 0 P4D 0 [ 316.825357] Oops: 0002 [#1] PREEMPT SMP NOPTI [ 316.826350] CPU: 0 PID: 1517 Comm: kunit_try_catch Tainted: G U N 6.1.0-rc2-drm-266703e6f163+ #14 [ 316.827503] Hardware name: Intel Corporation Tiger Lake Client Platform/TigerLake Y LPDDR4x T4 Crb, BIOS TGLSFWI1.R00.3243.A01.2006102133 06/10/2020 [ 316.828633] RIP: 0010:down_write_killable+0x50/0x110 [ 316.829756] Code: 24 10 45 31 c9 31 c9 41 b8 01 00 00 00 31 d2 31 f6 48 89 ef e8 e1 74 4a ff bf 01 00 00 00 e8 87 d6 46 ff 31 c0 ba 01 00 00 00 48 0f b1 13 0f 94 c0 5a 84 c0 74 62 8b 05 49 12 e4 00 85 c0 74 [ 316.830896] RSP: 0018:c90001eabc58 EFLAGS: 00010246 [ 316.832008] RAX: RBX: 00e8 RCX: [ 316.833141] RDX: 0001 RSI: 81c94fc9 RDI: 81c94fc9 [ 316.834195] RBP: 0158 R08: 0001 R09: [ 316.835231] R10: R11: 8883a13350b8 R12: 0002 [ 316.836259] R13: 0001 R14: 0010 R15: 00e8 [ 316.837237] FS: () GS:8883a380() knlGS: [ 316.838214] CS: 0010 DS: ES: CR0: 80050033 [ 316.839190] CR2: 00e8 CR3: 02812003 CR4: 00770ef0 [ 316.840147] PKRU: 5554 [ 316.841099] Call Trace: [ 316.842047] [ 316.842990] ? vm_mmap_pgoff+0x78/0x150 [ 316.843936] vm_mmap_pgoff+0x78/0x150 [ 316.844884] igt_mmap_offset+0x178/0x1b9 [i915] [ 316.846119] __igt_mmap+0xfe/0x680 [i915] Unfortunately, when KUnit module runs, it doesn't create an user context, causing mmap tests to fail. Signed-off-by: Mauro Carvalho Chehab --- To avoid mailbombing on a large number of people, only mailing lists were C/C on the cover. See [PATCH RFC v2 0/8] at: https://lore.kernel.org/all/cover.1667486144.git.mche...@kernel.org/ drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c | 5 + 1 file changed, 5 insertions(+)
