Le 27/02/2026 à 10:43 PM, Andy Shevchenko a écrit :
On Fri, Feb 27, 2026 at 06:42:12PM +0800, David Gow wrote:
Le 27/02/2026 à 5:40 PM, Andy Shevchenko a écrit :
I have stumbled over the kunit framework issues that make the respective test
cases useless.
Now to the details.
Consider having today's Linux Next.
Sorry to hear that KUnit is causing trouble. It looks like this is due to
those patches crashing the kernel before KUnit gets to run: by using the
--raw_output=full argument to kunit.py run, the corresponding logs are
shown.
Scenario 1 (good):
I run
./tools/testing/kunit/kunit.py config
./tools/testing/kunit/kunit.py run printf
Everything works as expected:
[10:19:36] Testing complete. Ran 28 tests: passed: 28
[10:19:36] Elapsed time: 15.929s total, 0.001s configuring, 15.761s
building, 0.114s running
This works fine for me, too. :-)
Scenario 2 (BAD):
I applied the following change:
--- a/lib/vsprintf.c
+++ b/lib/vsprintf.c
@@ -18,6 +18,7 @@
*/
#include <linux/stdarg.h>
+#include <linux/bitops.h>
#include <linux/build_bug.h>
#include <linux/clk.h>
#include <linux/clk-provider.h>
@@ -2904,12 +2889,17 @@ int vsnprintf(char *buf, size_t size, const char
*fmt_str, va_list args)
case FORMAT_STATE_NUM: {
unsigned long long num;
+ u8 shift = fmt.size * 8 - 1;
if (fmt.size > sizeof(int))
num = va_arg(args, long long);
else
- num = convert_num_spec(va_arg(args, int),
fmt.size, spec);
- str = number(str, end, num, spec);
+ num = va_arg(args, int);
+ num = sign_extend64(num, shift);
+ if (spec.flags & SIGN)
+ str = number(str, end, num, spec);
+ else
+ str = number(str, end, -(long long)num, spec);
continue;
}
Tests went into cosmos (I waited a few minutes and has to interrupt that):
^CERROR:root:Build interruption occurred. Cleaning console.
^CERROR:root:Build interruption occurred. Cleaning console.
^CERROR:root:Build interruption occurred. Cleaning console.
Command '['.kunit/linux', 'kunit.filter_glob=printf', 'kunit.enable=1',
'mem=1G', 'console=tty', 'kunit_shutdown=halt']' timed out after 300 seconds
[10:29:52] [ERROR] Test: <missing>: Could not find any KTAP output. Did any
KUnit tests run?
[10:29:52] ============================================================
[10:29:52] Testing complete. Ran 0 tests: errors: 1
[10:29:52] Elapsed time: 305.676s total, 0.001s configuring, 5.669s
building, 300.006s running
NOTE!
Independently on how long I waited the Elapsed time is about 5 minutes
(Seems 300 seconds limit as stated in the output).
Interesting: this crashed immediately on my machine. During building, I see
a (harmless) warning:
../lib/vsprintf.c:2827:27: warning: ‘convert_num_spec’ defined but not used
[-Wunused-function]
You need to enable binary printf() in the configuration, or comment out that
function.
I have no such warning as I dropped the function (haven't used it in the above
change
for the sake of simplicity.
Yeah, I assumed it was just an in-progress patch or something, but it
did confirm that at least some build output was making it through kunit.py.
By running KUnit with the --raw_output=full option, I can see a segfault
(though, as you can see, the numbers throughout the stacktrace a wrong):
<18446744073709551610>Pid: 1, comm: swapper/0 Not tainted
7.0.0-rc1-gff0627514551-dirty
<18446744073709551610>RIP: ffffffffffffffcd:0xffffffff9fac7320
<18446744073709551610>RSP: ffffffff5f7fc098 EFLAGS: fffffffffffefdf9
<18446744073709551610>RAX: fffffffffffc0000 RBX: ffffffff9ff1ad4c RCX:
ffffffff9f7b6440
<18446744073709551610>RDX: 0000000000000000 RSI: ffffffff9f7b6388 RDI:
ffffffffc6cfc8cc
<18446744073709551610>RBP: 0000000000000000 R08: ffffffffffffffff R09:
ffffffffffffffd0
<18446744073709551610>R10: fffffffffffffff8 R11: fffffffffffffdba R12:
ffffffff9fac7320
<18446744073709551610>R13: ffffffff9ff1b0d0 R14: 0000000000000000 R15:
ffffffff9f963fe8
<0>Kernel panic - not syncing: Segfault with no mm
<18446744073709551612>CPU: 0 UID: 0 PID: 1 Comm: swapper/0 Not tainted
7.0.0-rc1-gff0627514551-dirty #35 VOLUNTARY
<18446744073709551612>Stack:
<18446744073709551612> ffffffff9fbe2fd0 00000000 ffffffff9fffdaec
ffffffff5f7fc080
<18446744073709551612> ffffffff5f7fc080 ffffffff9ff95050 ffffffff9fab1be0
ffffffff9ff95050
<18446744073709551612> ffffffff9fbe2fd0 00000000 00000000 00000000
<18446744073709551612>Call Trace:
<18446744073709551612> [<ffffffff9fbe2fd0>] ?
kernel_init+0x0/0xfffffffffffffe20
<18446744073709551612> [<ffffffff9fffdaec>] ?
kernel_init_freeable+0xfffffffffffffe8b/0xfffffffffffffc82
<18446744073709551612> [<ffffffff9ff95050>] ?
uml_curr_cpu+0x0/0xfffffffffffffff0
<18446744073709551612> [<ffffffff9ff95050>] ?
uml_curr_cpu+0x0/0xfffffffffffffff0
<18446744073709551612> [<ffffffff9fbe2fd0>] ?
kernel_init+0x0/0xfffffffffffffe20
<18446744073709551612> [<ffffffff9fbe2faa>] ?
kernel_init+0xffffffffffffffda/0xfffffffffffffe20
<18446744073709551612> [<ffffffff9ff95050>] ?
uml_curr_cpu+0x0/0xfffffffffffffff0
<18446744073709551612> [<ffffffff9ffa5c27>] ?
new_thread_handler+0xffffffffffffff87/0xffffffffffffff60
(Trying the same thing with --arch x86_64 suggested that some stack
corruption is occurring.)
Got it!
Scenario 3 (BAD):
Now I took again a clean tree and applied this change:
--- a/lib/vsprintf.c
+++ b/lib/vsprintf.c
@@ -18,6 +18,7 @@
*/
#include <linux/stdarg.h>
+#include <linux/bitops.h>
#include <linux/build_bug.h>
#include <linux/clk.h>
#include <linux/clk-provider.h>
@@ -2904,11 +2889,17 @@ int vsnprintf(char *buf, size_t size, const char
*fmt_str, va_list args)
case FORMAT_STATE_NUM: {
unsigned long long num;
+ u8 shift = fmt.size * 8 - 1;
if (fmt.size > sizeof(int))
num = va_arg(args, long long);
+ else {
+ num = va_arg(args, int);
+ if ((spec.flags & SIGN))
+ num = sign_extend64(num, shift);
else
- num = convert_num_spec(va_arg(args, int),
fmt.size, spec);
+ num &= ~(BIT_ULL(shift) - 1);
+ }
str = number(str, end, num, spec);
continue;
}
and run tests again.
[10:39:16] [ERROR] Test: <missing>: Could not find any KTAP output. Did any
KUnit tests run?
[10:39:16] ============================================================
[10:39:16] Testing complete. Ran 0 tests: errors: 1
[10:39:16] Elapsed time: 5.762s total, 0.001s configuring, 5.694s building,
0.067s running
it runs fast and completely useless. (There is no build error)
This one also kernel panics, and when run with --raw_output=full, we can see
that it's due to all of the character devices' sysfs entries being
duplicates, because the minor/major are being formatted as '/dev/char/0:0':
<0>sysfs: cannot create duplicate filename '/dev/char/0:0'
(...)
<0>Kernel panic - not syncing: Couldn't register pty driver
<0>CPU: 0 UID: 0 PID: 1 Comm: swapper/0 Tainted: G W
7.0.0-rc1-gff0627514551-dirty #34 VOLUNTARY
Thanks for trying it and explaining me what's going on. At the bottom line I
missed
--raw_output=full which should be enough for me.
...
Yeah, --raw_output=full (or its synonym --raw_output=all) are very
useful for debugging this sort of thing.
Basically, --raw_output or --raw_output=kunit will give only the kernel
logs from after the tests have started running (i.e., the KTAP header
line has been printed), and --raw_output=all will pass through all
kernel output, which is required if the KUnit tests never get run.
Please, fix this as it is a serious issue and really makes kunit useless.
There's not much KUnit can do if the kernel panics before any tests can be
run -- and unfortunately, vsprintf() seems able to cause lots of trouble
early in the boot process.
One idea is to support building tests as independent userspace executables,
which wouldn't depend on all of those parts of the kernel which break (and
would be easier to debug). I discussed this a bit at Plumbers a couple of
years ago[1], but haven't had a chance to work on it since. Even then, it'd
require a little bit of test-specific work to get an isolated version of the
kernel vsprintf to build and be testable.
In the short term, maybe we can improve the interface of kunit.py in cases
where the kernel crashes. At the moment, we simply report that no tests had
run (as you've noticed), but maybe we should check more actively for panics
and/or make a more explicit difference between "no tests were run" and "the
KUnit framework never exectuted". At the very least, we should suggest that
--raw_output=full is a good way to debug this issue if the user wasn't
expecting it in the error message. (I'll send a patch out to do this now.)
Most annoying part is hanging the console, and then after Ctrl+C pressed,
+300 seconds (unneeded!) timeout occurs.
Yes, that's been annoying for a while. I've sent a patch out [1] trying
to fix it -- at least we won't go out of our way to eat SIGINT any more
-- but I'm sure there are still some cases where UML or QEMU could hang
more nastily.
I hope that helps (at least a little bit), and thanks for sticking with
KUnit despite these issues!
[1]: https://lpc.events/event/18/contributions/1790/
Cheers,
-- David
[1]: https://lore.kernel.org/all/[email protected]/
