SPL regression failure linux:shrinker

[building39]

When spl regression tests are run (Ubuntu 12.04, kernel 3.2.0-22-generic) this failure occurs:

compress/uncompress: Passed level 9, compressed 131072 bytes to 1836 bytes
linux:shrink_dcache ^[[0;32mPass^[[0m
shrink_dcache: Shrink dcache memory, remain 0 -> 0
linux:shrink_icache ^[[0;32mPass^[[0m
shrink_icache: Shrink icache memory, remain 0 -> 0
linux:shrinker ^[[0;31mFail^[[0m Numerical argument out of domain
shrinker: Failed cache was not shrunk to 0, size now 1024.

[behlendorf]

Interesting, thanks for the bug report we'll look in to it.

[rayvt]

On a new install of Ubuntu precise:
uname -a
Linux igor 3.2.0-24-generic-pae #37-Ubuntu SMP Wed Apr 25 10:47:59 UTC 2012 i686 athlon i386 GNU/Linux

I had this failure, plus the two in issue #80 (generic:divdi3 and atomic:64-bit fail)

After applying the 2 fixes, all tests passed.

[behlendorf]

Could you be more specific, exactly which 2 patches resolved your issue.

[rayvt]

Sorry.
I added "#undef abs64" into include/linux/math64_compat.h (as per openzfs/zfs#240)
and
./configure --enable-atomic-spinlocks

All spl tests passed on Debian 6.0.4 with kernel 2.6.32-5-686 #1 SMP, with the original unpatched spl-0.6.0-rc8

[ghost]

I have come across the same issue on a new install of Ubuntu
uname -a
Linux control 3.2.0-24-generic #39-Ubuntu SMP Mon May 21 16:52:17 UTC 2012 x86_64 x86_64 x86_64 GNU/Linux

However, if I run
sudo splat -t linux:shrinker -v
many times in a row, it will work once in a while (roughly 1/5th of the time)...

[pdf]

Same here:
Linux zfs-ubuntu-test 3.2.0-25-virtual #40-Ubuntu SMP Wed May 23 22:20:17 UTC 2012 x86_64 x86_64 x86_64 GNU/Linux
gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3

My success rate is more like 1/2500 linux:shrinker runs, but that's what you get with timing issues.

Compiling the SPL modules with --enable-atomic-spinlocks has no impact. Why is openzfs/zfs#240 still closed? The resolution seems to be that the user hand-patched some stuff on his system...

Do we have any idea what the cause is, or is there any way I can provide more information?

[behlendorf]

Sorry this one slipped through the cracks. It looks like a flaw in the test which needs to be fixed. The assumption was that once drop slabs returned all entires on our slab must have been dropped. That doesn't seem to be strictly true so we'll need to run down exactly why.

[terry-permeance]

I get the same error with spl-0.6.0-rc12, even after recompiling with the extra arg to ./configure and the patch applied.