qemu-kvm failure on linux 4.14.96

[donald]

qemu-kvm keeps failing on me. Error message "Inconsistency detected by ld.so: dl-version.c: 205: _dl_check_map_versions: Assertion `needed != NULL' failed!" on host and kernel panic on guest

[    2.264720] Freeing unused kernel memory: 2012K
[    2.266174] Freeing unused kernel memory: 592K
Inconsistency detected by ld.so: dl-version.c: 205: _dl_check_map_versions: Assertion `needed != NULL' failed!
[    2.295178] Kernel panic - not syncing: Attempted to kill init! exitcode=0x00007f00
[    2.295178] 
[    2.296107] CPU: 2 PID: 1 Comm: bash Not tainted 4.14.96.mx64.240 #1
[    2.296744] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-prebuilt.qemu-project.org 04/01/2014
[    2.297909] Call Trace:

After downgrading the host to 4.14.87, the emulation runs.

[donald]

Test tldr : /scratch/local/vmtest/test.sh

Will get to a kernel panic (bad) or a login prompt (ok) , login not possible, though.

Image is read-only and no network, so this test doesn't mess up anything. Can be aborted with ^C

• 4.14.87-236 : ok
• 4.14.94-239: bad
• 4.14.94-239 without '-enable-kvm' (qemu only) : ok
• 4.14.96-240: bad

[pmenzel]

I tried to test this in the VM dose, but the automounter and so on also failed with 4.14.87-236. Do I have to configure dose in some kind of way to get to the login prompt?

[pmenzel]

• Linux 4.14.90-241: good (on dose)

[donald]

tldr : /scratch/local/vmtest/test.sh is a copy of dose. When started without network, of course some mariux-specific things fail.

[donald]

https://lore.kernel.org/stable/7cedf2fd-ef08-44c8-8db5-a34187d035b8@gmail.com/T/#t looks promising.

[pmenzel]

Nice, so 4.14.97 should be tried. Unfortunately /src/mariux/patches/0001-Revert-genirq-affinity-assign-vectors-to-all-possibl.patch does not apply cleanly anymore. (Let’s move to 4.19.x. ;-)).

[donald]

4.14.97: good.

Probably 9485d5d2318b ("KVM: x86: Fix a 4.14 backport regression related to userspace/guest FPU") really fixed it. I'll check 4.19.19 now.

[donald]

4.19.19: good.

[donald]

We don't need a 'Revert "genirq/affinity: assign vectors to all possible CPUs' patch in 4.19, do we?

[thomas]

The patch was specific for 4.14 kernels after .56 and shouldn't be needed for the successors (although it was mentioned that the underlying problem wasn't fixed in the aacraid driver). So no clear Yes/No from my side -- but an optimistic 'it should work' without the patch.

[thomas]

As for the aacraid-driver (tests on deadbird):

• Drives are detected
• Plug/unplug of enclosures works

Mirror is used as 'long-run' test (phaedra:C8016 mirror to deadbird:C8014)
Right now C8014 is handled by the HBA1000 (aacraid), before it was attached to a HBA1100 (smartpqi).

[donald]

Okay, the qmeu-kvm issue is fixed. Systems wanting to use kvm should avoid kernel 4.14 < 4.14.97. I close the issue.

[thomas]

BTW: Microsemi has released new (Jan 2019) off-tree drivers for both adapters, so in case of trouble we have some more options :)