On 11.04.2016 03:55, David Gibson wrote:
> On Fri, 8 Apr 2016 13:35:29 +0200
> Thomas Huth <thuth@xxxxxxxxxx> wrote:
>
>> This test can be used to check whether the SPR (special purpose
>> registers) of the PowerPC CPU are migrated right. It first fills
>> the various SPRs with some non-zero value, then reads the values
>> back into a first array, then waits for a key (with the '-w' option)
>> so that it is possible to migrate the VM, and finally reads the
>> values from the SPRs back into another array and then compares it
>> with the initial values.
>> Currently the test only supports the SPRs from the PowerISA v2.07
>> specification (i.e. POWER8 CPUs), but other versions should be
>> pretty easy to add later.
>>
>> Signed-off-by: Thomas Huth <thuth@xxxxxxxxxx>
>
> So, the main concern I have about this is that while writing an
> arbitrary value is ok for quite a few SPRs, there's a significant
> number where that's not safe: either because it could actually cause
> some sort of exception interrupting the test, or because the value will
> get masked or otherwise modifier on write.
If a write to an SPR caused any exception or hang of the guest, I did
not include it in the list of SPRs to be tested here. Sorry, I should
have mentioned that somewhere in the commit message or comment of the
sprs.c file below.
Anyway, working out a detailed list of values that should be written
into each of the SPRs might get quite difficult/cumbersome since some of
these SPRs are not that well documented in the PowerISA. So I'd like to
keep the generic approach here instead. Anyway, if a write causes
trouble, it's not a big issue to take that SPR out of the list again.
And since the kvm-unit-tests are also a very isolated test (i.e. we're
not running a Linux kernel here which might expect certain values in
certain SPRs), it's IMHO also ok to mess up the state in the SPRs since
the VM will get destroyed after the test anyway.
About your second concern, that the values will get masked or modified:
Yes, you're right, of course! But that's not a problem here since I read
back the values from the SPRs before the migration - I do not use the
original value that I wrote into the SPRs for comparison!
> [snip]
>> +/* Common SPRs for all PowerPC CPUs */
>> +static void set_sprs_common(uint64_t val)
>> +{
>> + mtspr(1, val); /* XER */
>> + mtspr(9, val); /* CTR */
>> + mtspr(273, val); /* SPRG1 */
>> + mtspr(274, val); /* SPRG2 */
>> + mtspr(275, val); /* SPRG3 */
>> +}
>> +
>> +/* SPRs from PowerISA 2.07 Book III-S */
>> +static void set_sprs_book3s_207(uint64_t val)
>> +{
>> + mtspr(3, val); /* DSCR */
>> + mtspr(13, val); /* AMR */
>> + mtspr(17, val); /* DSCR */
>> + mtspr(18, val); /* DSISR */
>> + mtspr(19, val); /* DAR */
>> + mtspr(29, val); /* AMR */
>
> AMR seems to be listed twice..
That's because AMR is available via both SPR numbers - one time for
userspace mode (which can be disabled by the kernel if desired), and one
time for privileged mode.
> At a glance SPRs above where writing an arbitrary value might not be
> safe include AMR, IAMR, UAMOR, and MMCR*.
Well, you should not think of this test as a nice, behaving guest kernel
that only puts valid values into safe SPRs ... rather think of it as a
stress test for the host with a bad, misbehaving guest kernel.
So think what's the worst thing that could happen? Host crash? Right,
this test already helped to find one of those bugs:
http://www.spinics.net/lists/kvm/msg128989.html
... but then we've of course got to fix the host kernel (or QEMU), not
this test.
The other bad thing that could of course happen is a guest crash - but
then, as mentioned above, we can simply adjust the list of SPRs that we
test as soon as we see such a crash. With the current list, I do not get
any guest crashes - at least not with kvm-hv. ... but kvm-pr and tcg are
two other candidates that likely need some fixing for this test.
Thomas
Attachment:
signature.asc
Description: OpenPGP digital signature
