+open list:KERNEL SELFTEST FRAMEWORK In case anyone in kselftest has any thoughts.
On Thu, Oct 10, 2019 at 7:05 PM Theodore Ts'o theodore.tso@gmail.com wrote:
I've been experimenting with the ext4 kunit test case, and something that would be really helpful is if the default is to store the object files for the ARCM=um kernel and its .config file in the top-level directory .kunit. That is, that the default for --build_dir should be .kunit.
Why does this important? Because the kernel developer will want to be running unit tests as well as building kernels that can be run under whatever architecture they are normally developing for (for example, an x86 kernel that can be run using kvm; or a arm64 kernel that gets run on an Android device by using the "fastboot" command). So that means we don't want to be overwriting the object files and .config files for building the kernel for x86 when building the kunit kernel using the um arch. For example, for ext4, my ideal workflow might go something like this:
That's a good point.
<hack hack hack> % ./tools/testing/kunit/kunit.py run <watch to see that unit tests succeed, and since most of the object files have already been built for the kunit kernel in be stored in the .kunit directory, this will be fast, since only the modified files will need to be recompiled> % kbuild <this is a script that builds an x86 kernel in /build/ext4-64 that is designed to be run under either kvm or in a GCE VM; since the kunit object files are stored in /build/ext4-kunit, the pre-existing files when building for x86_64 haven't been disturbed, so this build is fast as well> % kvm-xfstests smoke <this will run xfstests using the kernel plucked from /build/ext-64, using kvm>
The point is when I'm developing an ext4 feature, or reviewing and merging ext4 commits, I need to be able to maintain separate build trees and separate config files for ARCH=um as well as ARCH=x86_64, and if the ARCH=um are stored in the kernel sources, then building with O=... doesn't work:
tytso@lambda {/usr/projects/linux/kunit} (kunit) 1084% make O=/build/test-dir make[1]: Entering directory '/build/test-dir'
*** The source tree is not clean, please run 'make mrproper' *** in /usr/projects/linux/kunit
Should we maybe drop `--build_dir` in favor of `O`?
One of the other reasons why it would be good to use --build_dir by default is that way, building with a separate O= build directory is regularly tested. Right now, "kunit.py --build_dir=" seems to be broken.
Good point.
% ./tools/testing/kunit/kunit.py run --build_dir=/build/ext4-kunit Generating .config ... [22:04:12] Building KUnit Kernel ... /usr/projects/linux/kunit/arch/x86/um/user-offsets.c:20:10: fatal error: asm/syscalls_64.h: No such file or directory 20 | #include <asm/syscalls_64.h> | ^~~~~~~~~~~~~~~~~~~ compilation terminated.
(This appears to be an ARCH=um bug, not a kunit bug, though.)
Yeah, I encountered this before. Some file is not getting properly cleaned up by `make mrproper`. It works if you do `git clean -fdx` (I know that's not a real solution for most people). Nevertheless, it sounds like we need to sit down and actually solve this problem since it is affecting users now.
I think you make a compelling argument. Anyone else have any thoughts on this?
On 10/11/19 4:19 AM, Brendan Higgins wrote:
+open list:KERNEL SELFTEST FRAMEWORK In case anyone in kselftest has any thoughts.
On Thu, Oct 10, 2019 at 7:05 PM Theodore Ts'o theodore.tso@gmail.com wrote:
I've been experimenting with the ext4 kunit test case, and something that would be really helpful is if the default is to store the object files for the ARCM=um kernel and its .config file in the top-level directory .kunit. That is, that the default for --build_dir should be .kunit.
Why does this important? Because the kernel developer will want to be running unit tests as well as building kernels that can be run under whatever architecture they are normally developing for (for example, an x86 kernel that can be run using kvm; or a arm64 kernel that gets run on an Android device by using the "fastboot" command). So that means we don't want to be overwriting the object files and .config files for building the kernel for x86 when building the kunit kernel using the um arch. For example, for ext4, my ideal workflow might go something like this:
That's a good point.
<hack hack hack> % ./tools/testing/kunit/kunit.py run <watch to see that unit tests succeed, and since most of the object files have already been built for the kunit kernel in be stored in the .kunit directory, this will be fast, since only the modified files will need to be recompiled> % kbuild <this is a script that builds an x86 kernel in /build/ext4-64 that is designed to be run under either kvm or in a GCE VM; since the kunit object files are stored in /build/ext4-kunit, the pre-existing files when building for x86_64 haven't been disturbed, so this build is fast as well> % kvm-xfstests smoke <this will run xfstests using the kernel plucked from /build/ext-64, using kvm>
The point is when I'm developing an ext4 feature, or reviewing and merging ext4 commits, I need to be able to maintain separate build trees and separate config files for ARCH=um as well as ARCH=x86_64, and if the ARCH=um are stored in the kernel sources, then building with O=... doesn't work:
tytso@lambda {/usr/projects/linux/kunit} (kunit) 1084% make O=/build/test-dir make[1]: Entering directory '/build/test-dir'
*** The source tree is not clean, please run 'make mrproper' *** in /usr/projects/linux/kunit
Should we maybe drop `--build_dir` in favor of `O`?
Yes, preferably be consistent with the rest of the kernel makefiles.
One of the other reasons why it would be good to use --build_dir by default is that way, building with a separate O= build directory is regularly tested. Right now, "kunit.py --build_dir=" seems to be broken.
Good point.
% ./tools/testing/kunit/kunit.py run --build_dir=/build/ext4-kunit Generating .config ... [22:04:12] Building KUnit Kernel ... /usr/projects/linux/kunit/arch/x86/um/user-offsets.c:20:10: fatal error: asm/syscalls_64.h: No such file or directory 20 | #include <asm/syscalls_64.h> | ^~~~~~~~~~~~~~~~~~~ compilation terminated.
(This appears to be an ARCH=um bug, not a kunit bug, though.)
Yeah, I encountered this before. Some file is not getting properly cleaned up by `make mrproper`. It works if you do `git clean -fdx` (I know that's not a real solution for most people). Nevertheless, it sounds like we need to sit down and actually solve this problem since it is affecting users now.
I think you make a compelling argument. Anyone else have any thoughts on this?
On Fri, Oct 11, 2019 at 7:56 AM Randy Dunlap rdunlap@infradead.org wrote:
On 10/11/19 4:19 AM, Brendan Higgins wrote:
+open list:KERNEL SELFTEST FRAMEWORK In case anyone in kselftest has any thoughts.
On Thu, Oct 10, 2019 at 7:05 PM Theodore Ts'o theodore.tso@gmail.com wrote:
I've been experimenting with the ext4 kunit test case, and something that would be really helpful is if the default is to store the object files for the ARCM=um kernel and its .config file in the top-level directory .kunit. That is, that the default for --build_dir should be .kunit.
Why does this important? Because the kernel developer will want to be running unit tests as well as building kernels that can be run under whatever architecture they are normally developing for (for example, an x86 kernel that can be run using kvm; or a arm64 kernel that gets run on an Android device by using the "fastboot" command). So that means we don't want to be overwriting the object files and .config files for building the kernel for x86 when building the kunit kernel using the um arch. For example, for ext4, my ideal workflow might go something like this:
That's a good point.
<hack hack hack> % ./tools/testing/kunit/kunit.py run <watch to see that unit tests succeed, and since most of the object files have already been built for the kunit kernel in be stored in the .kunit directory, this will be fast, since only the modified files will need to be recompiled> % kbuild <this is a script that builds an x86 kernel in /build/ext4-64 that is designed to be run under either kvm or in a GCE VM; since the kunit object files are stored in /build/ext4-kunit, the pre-existing files when building for x86_64 haven't been disturbed, so this build is fast as well> % kvm-xfstests smoke <this will run xfstests using the kernel plucked from /build/ext-64, using kvm>
The point is when I'm developing an ext4 feature, or reviewing and merging ext4 commits, I need to be able to maintain separate build trees and separate config files for ARCH=um as well as ARCH=x86_64, and if the ARCH=um are stored in the kernel sources, then building with O=... doesn't work:
tytso@lambda {/usr/projects/linux/kunit} (kunit) 1084% make O=/build/test-dir make[1]: Entering directory '/build/test-dir'
*** The source tree is not clean, please run 'make mrproper' *** in /usr/projects/linux/kunit
Should we maybe drop `--build_dir` in favor of `O`?
Yes, preferably be consistent with the rest of the kernel makefiles.
Alright, probably a good idea to make this change fairly soon then before we have to worry about backwards compatibility and such.
On Wed, Oct 16, 2019 at 02:04:35PM -0700, Brendan Higgins wrote:
Should we maybe drop `--build_dir` in favor of `O`?
Yes, preferably be consistent with the rest of the kernel makefiles.
Alright, probably a good idea to make this change fairly soon then before we have to worry about backwards compatibility and such.
I'm not sure how this would work; so something like:
.../kunit.py run O=/build_dir
Should other flags we can pass in via the makefile processing, such as V=1, etc., also work? What other things can we pass in via after the "run" command?
And if we're going to go this far, maybe we should make "make kunit" run tools/testing/kunit/kunit.py?
Some minor other nits if you're going to be making changes to kunit.py's CLI parsing:
1) It would be nice if there was a help command so that "kunit.py help" does what kunit.py -h does.
2) The top-level help message should indicate that "kunit.py run" takes various optional arguments and the way to find them is "kunit.py run -h". This was *not* obvious, and the way I figured out there was even --build_dir option was via purusing the source code. (It wasn't in the documentation that I could find.)
3) And maybe then "kunit.py help run" should display the help message for "kunit.py urn". This would make it consistent with other tools that some of us might be familiar with (e.g., gcloud, gsutil, etc.)
Of course, if the front entry for kunit starts being "make kunit" as opposed to ./tools/testing/kunit/kunit.py, then we really need to figure out how to pass in the equivalent of --timeout. (Maybe --raw_output is enabled if we run make kunit V=1?). And of course, all of this would need to be documented.
- Ted
On Thu, Oct 17, 2019 at 5:51 AM Theodore Y. Ts'o tytso@mit.edu wrote:
On Wed, Oct 16, 2019 at 02:04:35PM -0700, Brendan Higgins wrote:
Should we maybe drop `--build_dir` in favor of `O`?
Yes, preferably be consistent with the rest of the kernel makefiles.
Alright, probably a good idea to make this change fairly soon then before we have to worry about backwards compatibility and such.
I'm not sure how this would work; so something like:
.../kunit.py run O=/build_dir
Seems reasonable to me.
Should other flags we can pass in via the makefile processing, such as V=1, etc., also work? What other things can we pass in via after the "run" command?
Hmmm...that's a good point. I don't know about V; probably need to improve how kunit_tool displays build information for that to be useful. I don't think that W is likely to be useful since I think that is semantically a different operation than just running KUnit tests.
Probably don't want to forward ARCH, or CROSS_COMPILE or any of those.
Supporting some of these[1] seems useful.
And if we're going to go this far, maybe we should make "make kunit" run tools/testing/kunit/kunit.py?
That seems reasonable. I was holding off on that initially because I thought it might be a bridge too far in terms of putting KUnit in a highly visible place. However, in hindsight, I think we crossed that bridge a long time ago with putting tests is very visible places. So yeah, now is probably a good time to do that.
Some minor other nits if you're going to be making changes to kunit.py's CLI parsing:
- It would be nice if there was a help command so that "kunit.py help" does what kunit.py -h does.
Seems reasonable.
- The top-level help message should indicate that "kunit.py run" takes various optional arguments and the way to find them is "kunit.py run -h". This was *not* obvious, and the way I figured out there was even --build_dir option was via purusing the source code. (It wasn't in the documentation that I could find.)
Also reasonable.
- And maybe then "kunit.py help run" should display the help message for "kunit.py urn". This would make it consistent with other tools that some of us might be familiar with (e.g., gcloud, gsutil, etc.)
That's reasonable, but also a little harder. At that point, I think we might need to write our own flag library to support all this. So, I will put this as a TODO, but we probably won't get to it for a while.
Of course, if the front entry for kunit starts being "make kunit" as opposed to ./tools/testing/kunit/kunit.py, then we really need to figure out how to pass in the equivalent of --timeout. (Maybe
Yeah, that's true. We should probably also by default set --timeout to a reasonable default instead of infinite.
--raw_output is enabled if we run make kunit V=1?). And of course,
Hmmm...that seems like a good idea.
all of this would need to be documented.
Yeah, we very much need to improve our documentation, especially for the kunit_tool. That is very high on our TODO list.
[1] https://www.gnu.org/software/make/manual/html_node/Options-Summary.html
linux-kselftest-mirror@lists.linaro.org
-
Brendan Higgins -
Randy Dunlap -
Theodore Y. Ts'o