On 28/02/2019 13:39, Halil Pasic wrote:
On Thu, 28 Feb 2019 10:42:23 +0100
Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
On 27.02.2019 19:00, Tony Krowiak wrote:
On 2/27/19 3:09 AM, Pierre Morel wrote:
On 26/02/2019 16:47, Tony Krowiak wrote:
On 2/26/19 6:47 AM, Pierre Morel wrote:
On 25/02/2019 19:36, Tony Krowiak wrote:
On 2/22/19 10:29 AM, Pierre Morel wrote:
We prepare the interception of the PQAP/AQIC instruction for
the case the AQIC facility is enabled in the guest.
We add a callback inside the KVM arch structure for s390 for
a VFIO driver to handle a specific response to the PQAP
instruction with the AQIC command.
We inject the correct exceptions from inside KVM for the case the
callback is not initialized, which happens when the vfio_ap driver
is not loaded.
If the callback has been setup we call it.
If not we setup an answer considering that no queue is available
for the guest when no callback has been setup.
We do consider the responsability of the driver to always initialize
the PQAP callback if it defines queues by initializing the CRYCB for
a guest.
Signed-off-by: Pierre Morel <pmorel@xxxxxxxxxxxxx>
...snip...
@@ -592,6 +593,55 @@ static int handle_io_inst(struct kvm_vcpu *vcpu)
}
}
+/*
+ * handle_pqap: Handling pqap interception
+ * @vcpu: the vcpu having issue the pqap instruction
+ *
+ * We now support PQAP/AQIC instructions and we need to correctly
+ * answer the guest even if no dedicated driver's hook is available.
+ *
+ * The intercepting code calls a dedicated callback for this instruction
+ * if a driver did register one in the CRYPTO satellite of the
+ * SIE block.
+ *
+ * For PQAP/AQIC instructions only, verify privilege and specifications.
+ *
+ * If no callback available, the queues are not available, return this to
+ * the caller.
+ * Else return the value returned by the callback.
+ */
+static int handle_pqap(struct kvm_vcpu *vcpu)
+{
+ uint8_t fc;
+ struct ap_queue_status status = {};
+
+ /* Verify that the AP instruction are available */
+ if (!ap_instructions_available())
+ return -EOPNOTSUPP;
How can the guest even execute an AP instruction if the AP instructions
are not available? If the AP instructions are not available on the host,
they will not be available on the guest (i.e., CPU model feature
S390_FEAT_AP will not be set). I suppose it doesn't hurt to check this
here given QEMU may not be the only client.
+ /* Verify that the guest is allowed to use AP instructions */
+ if (!(vcpu->arch.sie_block->eca & ECA_APIE))
+ return -EOPNOTSUPP;
+ /* Verify that the function code is AQIC */
+ fc = vcpu->run->s.regs.gprs[0] >> 24;
+ if (fc != 0x03)
+ return -EOPNOTSUPP;
You must have missed my suggestion to move this to the
vcpu->kvm->arch.crypto.pqap_hook(vcpu) in the following responses:
Please consider what happen if the vfio_ap module is not loaded.
I have considered it and even verified my expectations empirically. If
the vfio_ap module is not loaded, you will not be able to create an mdev device.
OK, now please consider that another userland tool, not QEMU uses KVM.
What does that have to do with loading the vfio_ap module? Without the
vfio_ap module, there will be no AP devices for the guest. What are you
suggesting here?
If you don't have an mdev device, you will not be able to
start a guest with a vfio-ap device. If you start a guest without a
vfio-ap device, but enable AP instructions for the guest, there will be
no AP devices attached to the guest. Without any AP devices attached,
the PQAP(AQIC) instructions will not ever get executed.
This is not right. The instruction will be executed, eventually, after decoding.
Please explain why the PQAP(AQIC) instruction will be executed on a
guest without any devices? Point me to the code in the AP bus where
PQAP(AQIC) is executed without a queue?
The host must be prepared to handle malicous and broken guests. So if
a guest does PQAP, we must handle that gracefully (e.g. by injecting an
exception)
Nod.
Even if for some
unknown reason the PQAP(AQIC) instruction is executed - for some unknown
reason, it will fail with response code 0x01, AP-queue number not valid.
No, before accessing the AP-queue the instruction will be decoded and depending on the installed micro-code it will fail with
- OPERATION EXCEPTION if the micro-code is not installed
- PRIVILEDGE OPERATION if the instruction is issued from userland (programm state)
- SPECIFICATION exception if the instruction do not respect the usage specification
then it will be interpreted by the microcode and access the queue and only then it will fail with RC 0x01, AP queue not valid.
In the case of KVM, we intercept the instruction because it is issued by the guest and we set the AQIC facility on to force interception.
KVM do for us all the decode steps I mention here above, if there is or not a pqap hook to be call to simulate the QP queue access.
That done, the AP queue virtualisation can be called, this is done by calling the hook.
Okay, let's go back to the genesis of this discussion; namely, my
suggestion about moving the fc == 0x03 check into the hook code. If
the vfio_ap module is not loaded, there will be no hook code. In that
case, the check for the hook will fail and ultimately response code
0x01 will be set in the status word (which may not be the right thing
to do?). You have not stated a single good reason for keeping this
check, but I'm done with this silly argument. It certainly doesn't
hurt anything.
The instruction handler must handle the basic checks for the
instruction itself as outlined above.
Nod.
Do we want to allow QEMU to fully emulate everything (the ECA_APIE case being off)?
The we should pass along everything to QEMU, but this is already done with the
ECA_APIE check, correct?
Nod.
Do we agree that when we are beyond the ECA_APIE check, that we do not emulate
in QEMU and we have enabled the AP instructions interpretion?
At least the intention is to not emulate. ECA_APIE is an effective
control though...
If yes then this has some implication:
1. ECA is on and we should only get PQAP interception for specific FC (namely 3).
Not necessarily true. TAPQ can be intercepted as well (APFT depends
IC.3). But for now we don't care about that.
2. What we certainly should check is the facility bit of the guest (65) and reject fc==3
right away with a specification exception. I do not want the hook to mess with
the kvm cpu model. @Pierre would be good to actually check test_kvm_facility(vcpu->kvm, 65))
As far as I can tell he already does test_kvm_facility(vcpu->kvm, 65). I
agree we need a spec exception if guest does not have facility 65, but
does have ap instructions.
3. What shall we do when fc == 0x3? We can certainly do the check here OR in the
hook. As long as we have only fc==3 this does not matter.
I guess Tony's point is that we may have fc == 0 that is TAPQ in the
APFT flavor. IMHO we don't need to care about that at the moment.
Correct?
IMHO mostly.
I also doing the facility checks in kvm is easier, and I think this is
something we can change later if needed without any major trouble.
There are a couple of things I would do differently than Pierre does:
1) Do the PGM_PRIVILEGED_OP before the fc == 3 check.
Idea was not to modify existing behavior for fc != 3
Also Christian already proposed to handle all FC codes. So in this idea,
this must be done as you say.
2) Do the test_kvm_facility(vcpu->kvm, 65) check in the context of fc ==
3. I.e. decide if this hook is about pqap or just about pqap aqic and
make the code convey that decision to its reader.
3) I would most probably test if the queue is available by looking at the
masks in CRYCB here. If not AP_RESPONSE_Q_NOT_AVAIL is what we need.
This I do not agree with, it is typically the responsibility of the part
in charge of the virtualization to do this, also the vfio_driver.
4) If we have APIE and queues authorized by the CRYCB (i.e. we have a
vfio_ap module loaded an an mdev associated with the kvm) the callback
not set (!(vcpu->kvm->arch.crypto.pqap_hook)) is a BUG!
I do not agree with this either, the maintainers ;) will not allow this.
In that case
lying that the queue is not available does not seem right. BTW this is
something Pierre changed since the last version quietly (I can't recall
a mention in the change log or somebody asking for this). If we want to
be very pedantic about this bug scenario our best bet is probably
response code 6.
RC 06 means "Invalid address of AP-queue notification byte"
So you must have think about another code or I do not understand at all
what you mean.
Regards,
Pierre
--
Pierre Morel
Linux/KVM/QEMU in Böblingen - Germany
