Wednesday, August 28, 2013

Recent drm/i915 Testsuite Improvements

So recently I've again wreaked decent havoc in our intel-gpu-tools kernel testsuite. And then shockingly noticed that I've never done a big pompous blog post to announce what we've started about one-a-half years ago. So besides just describing the new infrastructure for writing testcases (apparently every decent hacker must reinvent a testframework at least once ...) I'll also recap a bit what's been going on in the past.

What Has Happened Thus Far ...

drm/i915 kernel testing was in a very sorry state two years ago - we've had pretty much zero regression testing of corner cases. Of course we've implicitly tested the kernel by running userspace drivers, e.g. testing mesa with piglit. But that only exercise the best case code and not at all the fun stuff hidden behind memory and other resource exhausting handling, racing multiple threads (or the cpu against the gpu) or trying to provoke coherency issues.

But there was a really basic set of manually run testcases in intel-gpu-tools. So we've started with them and wrapped them up in a very hackish testrunner using autotools and went ahead integrating this into our nightly regression testing. Like I've said in my fosdem presentation this year I was shocked how much fallout and regressions even the shittiest testsuite catches ...

Since then an awful lot of stuff has happened:

  • We've switching to piglit as our testrunner framework. The big reason for that was to automatically enumerate subtests and even more important to be able to easily filter them. Often we have tests that check combinations of different things. One example is gem_concurrent_blt which races gpu access against cpu access in different ways, using different cpu access methods and different caching policies, leading to a nice combinatorial explosion of subtests. And some tests also take an extreme amount of time to run, like the coherency checkers, especially the ones that exercise the swap paths ...

    The reason for using piglit was that no other established testrunner seems to sanely support skipping testcases out of the box, and we have tons of platform specific tests. And obviously tests for new features shouldn't fall over on older kernels. The other reason for picking piglit was that we could cobble together an integration with the existing tests, without really changing them, in a few lines of python. All other testsuites I've looked at were way more invasive to convert over to .

  • We've added tons more testcases. Starting with just twenty-odd test tools we've ended up with currently over 300 testcases. And very often when we've started to test an existing feature bugs just kept on showing up. So nowadays new features don't go in without decent test coverage.

  • We've added a lot of debugfs interfaces to the kernel to allow us to exercise tricky corner conditions. One of the first examples is the ring stop interface to simulate gpu hangs without actually hanging the gpu - the latter has the ugly tendency to take down the entire system, hard. With simulated gpu hangs we can exercise the hangcheck code, error capturing and dumping and the hw and soft reset logic without too much risk. This way we can exercise gpu hang vs. pageflip races and other nasty stuff that up to then has all been nicely broken. Other examples are interfaces to provoke gpu completion number wrap-arounds, disabling the prefetching to exercise slowpaths or interfaces to our shrinker code to more accurately exercise some of the memory trashing scenarios without actually thrashing the machine too badly.

  • We've also added tons of infrastructure for writing testcases in lib/drmtest.c. Detailing a bit what's all in there is the aim of this blog post. The most recent improvement here is that tests with subcases can now properly enumerate all subtests even when the drm kernel driver isn't available.

Running i-g-t Kernel Tests

At the moment there are no provisions for installing the tests as binaries and shipping them in distros - no one yet demanded this. So first you need to grab intel-gpu-tools and piglit sources:

$ git clone git://anongit.freedesktop.org/piglit
$ git clone git://anongit.freedesktop.org/xorg/app/intel-gpu-tools

Then build them:

$ cd intel-gpu-tools
$ ./autogen.sh && make
$ cd ../piglit
$ cmake . && make

We also need to tell piglit where the kernel testsuite is with a symlink:

$ cd bin
$ ln ../../intel-gpu-tools igt -s

Finally we can run testcases. Note that currently we assume that testcases are run as root, and there's no other drm client running. The rte testcase (short for runtime enviroment) will fail if that's not the case. Yeah, one of the projects is to improve the piglit testrunner to notice this earlier. Anyway, best get used to always run piglit for the kernel as root:

$ sudo ./piglit-run.py tests/igt result

Just enumerating the testcases doesn't need root though. But it will run the binaries of tests with subtests since igt subtests are enumerated at runtime - that helps greatly with constructing combinatorial testcases. So to list all tests just run

$ ./piglit-run.py -d tests/igt result

For further options and tools to display the test results, list the test commands and other useful stuff please consult the piglit documentation.

Helper Functions

The biggest part of the helper library is a pile of small wrappers around drm ioctls. libdrm is often at a way to high abstraction level, especially if you try to be nasty and provoke the kernel into doing stupid things. There's also the usual duplication of headers that we replicate (in slightly different versions) in all our userspace driver repositories ...

But the far more interesting functionality is in the other code in lib/drmtest.c. One of the most often used helpers is the rude interruptor, which forks a second process to constantly interrupt the first with a signal. That's useful for three reasons: X makes heavy use of signals, so we need to be able to exercise these paths, too. Then the kernel code itself relies on the ioctl restarting to bail out threads when the gpu is hung - this way we can avoid ugly issues when threads hog locks waiting for the gpu, while the reset code needs those logs to complete the requests (or restart them if we ever implement them) after the hardware resets. And finally it's just a great way to exercise error paths, since most of the tricky error handling code for e.g. out of memory conditions also needs to deal with interruptions due to a pending signal.

Another big chunk of helper logic is the just recently improved support code for subtests. The next section will go into greater detail on those. Then there's also helper code for system suspend/resume, gpu aperture thrashing, prefault disabling. And finally we have a few helpers for modesetting tests, to draw nice framebuffers with some gradients and corner markers and to get the VT handling right - it's rather tricky to stop the fbcon from interferring by blanking the screen in the middle of a test ...

Subtest Infrastructure and Magic Codeblocks

As mentioned already igt kernel tests enumerate subtests at runtime. Generally the subtest infrastructure is pretty simple. At the beginning of the test igt_subtest_init(argc, argv) needs to be called so that the subtest logic can correctly disable subtest - either all off them when only listing subtest names with --list-subtests or all but one when running a specific subtest with --run-subtes <subtest-name>. There is also a special variant of this init function called igt_subtest_init_parse_opts useful when the test program wants to augment its own option parsing with this functionality.

Then it's just a matter of enclosing the code for each subtest in an igt_subtest code block, e.g.

        igt_subtest("basic")
                run_basic_test();

This works like any C codeblock, i.e. if the block contains more than one statement it needs to be enclosed in curly braces. The block so marked up is only run when the named subtest should be run - when only listing subtests the code itself is skipped. The magic igt_subtest block will also print the result of the subtest to stdout. For combinatorial subtests where the subtest name is constructed at runtime there's also igt_subtest_f which takes a printf-like format string and additional paramters instead of a fixed string.

Now this is all nice and dandy but leaves the question how to do common setup code outside of the actual subtests and make sure it doesn't interfere when just listing the subtests or when running on a unsupported system where all subtests should be skipped. And then there's the question what to do when a subtest should be skipped or failed.

For the first issue there's another magic codeblock, igt_fixture. Anything within such a codeblock will be skipped when only listing subtests, but unconditionally run when any subtests should be run. So code to setup the global test fixture should be enclosed in such blocks.

For testcases without subtests the skipping and failing is easy to do - failed tests simply exit with a nonzero exit code, and the special exit code 77 is reserved to denote a skipped test. But with subtests we want to go through the entire remaining program so that the results for all subsequent subtests is still printed. Here comes igt_skip() and igt_exit(<exitcode>) to the rescue. With some longjmp magic they will bail out of the enclosing magic code block (either a subtest or a fixture). If those magic macros bail out of a subtest only that subtest is skipped/failed and all subsequent code is run - subtest of course only if they should be run, but fixtures (e.g. used for cleanup) are unconditionally run.

But if those magics macros are called within a fixture all subsequent fixture and subtest blocks are skipped since presumably a required piece of state couldn't be set up. Also for all subtests that still follow their status will be reported as either skipped or failed. This way when new features are added to a new combinatorial testcase which depends upon kernel support (like a new buffer object coherency mode) this can be added at the end of the test sequence: If the setup in a fixture block then fails all the subtests which need this new kernel support will be automagically skipped, but all the previous tests still run correctly.

For convenience there's also the igt_assert and igt_require macros which work like a normal assert. But instead of calling abort() internally they call igt_fail() and igt_skip(), respectively. And of course the print the file and line where the failure happened and the condition itself for easy identification. Unfortunately macros get expanded before being passed on, so ioctl numbers are really hard to read ...

Finally there's also an igt_success(), but most tests don't need to call this explicitly since the magic subtest code block will call this function at the very end of the code block, presuming that the subtest succeeded when it completed. And finally there's igt_exit() which will exit the test program with the correct exitcode for the piglit testrunner to interpret the result.

One thing to remember is that both fixtures and subtests internally use long jumps and that those can wreak havoc with stack variables. But when the testcases are written properly (specifically subtests must be independent of each another and only depend upon all the preceeding fixtures to have completed successfully) then this shouldn't ever cause trouble. It did cause a bit of trouble while converting testcase on an earlier version of  the fixture block - that used a long jump even when the fixture successfully complete and so could wreak the just set up stack variables.

So all in all for a testcase with subtests like gem_mmap_gtt this results in a fairly streamlined main function, despite all the magic:

int main(int argc, char **argv)
{

        int fd;

        igt_subtest_init(argc, argv);

        igt_fixture
                fd = drm_open_any();

        igt_subtest("access")
                test_access(fd);
        /* More subtests ... */

        igt_fixture
                close(fd);

        igt_exit();
}

The test functions themselves can also shed a lot of the control flow logic by directly using the igt_require/igt_asssert macros. Compared to the previous state before the introduction of the magic subtest and fixture code blocks all this control flow was explicitly done in each testcase and correspondingly buggy. And dropping all that code also nicely clarified what's actually being tested.

Update: Fix the setup commands a bit as spotted by Mika Kuoppala.

No comments:

Post a Comment