When QEMU switches to the XIVE interrupt mode, it creates all the
guest interrupts at the level of the KVM device. These interrupts are
backed by real HW interrupts from the IPI interrupt pool of the XIVE
controller.
Currently, this is done from the QEMU main thread, which results in
allocating all interrupts from the chip on which QEMU is running. IPIs
are not distributed across the system and the load is not well
balanced across the interrupt controllers.
Change the vCPU IPI allocation to run from the vCPU context. The
associated XIVE IPI interrupt will be allocated on the chip on which
the vCPU is running and improve distribution of the IPIs in the system.
When the vCPUs are pinned, this will make the IPI local to the chip of
the vCPU. It will reduce rerouting between interrupt controllers and
gives better performance.
Device interrupts are still treated the same. To improve placement, we
would need some information on the chip owning the virtual source or
the HW source in case of a passthrough device but this reuires
changes in PAPR.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20200820134547.2355743-5-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The vCPU IPIs are now allocated in kvmppc_xive_cpu_connect() when the
vCPU connects to the KVM device and not when all the sources are reset
in kvmppc_xive_source_reset()
This requires extra care for hotplug vCPUs and VM restore.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20200820134547.2355743-4-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This is doing an extra loop but should be equivalent.
It also differentiate the reset of the sources from the restore of the
sources configuration. This will help in allocating the vCPU IPIs
independently.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20200820134547.2355743-3-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
We will use to check if a vCPU IPI has been created.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20200820134547.2355743-2-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Now that kvmppc_xive_cpu_get_state() returns negative on error, use that
and get rid of the temporary Error object and error_propagate().
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707852916.1489912.8376334685349668124.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Now that all these functions return a negative errno on failure, check
that and get rid of the local_err boilerplate.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707851537.1489912.1030839306195472651.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Now that all these functions return a negative errno on failure, check
that because it is preferred to local_err. And most of all, propagate it
because vmstate expects negative errnos.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707850148.1489912.18355118622296682631.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Now that kvmppc_xive_get_queues() returns a negative errno on failure, check
with that because it is preferred to local_err. And most of all, propagate
it because vmstate expects negative errnos.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707849455.1489912.6034461176847728064.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Since kvm_device_access() returns a negative errno on failure, convert
kvmppc_xive_set_source_config() to use it for error checking. This allows
to get rid of the local_err boilerplate.
Propagate the return value so that callers may use it as well to check
failures.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707848764.1489912.17078842252160674523.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Since kvmppc_xive_get_queue_config() has a return value, convert
kvmppc_xive_get_queues() to use it for error checking. This allows
to get rid of the local_err boiler plate.
Propagate the return value so that callers may use it as well to check
failures.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707848069.1489912.14879208798696134531.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Since kvm_device_access() returns a negative errno on failure, convert
kvmppc_xive_get_queue_config() and kvmppc_xive_set_queue_config() to
use it for error checking. This allows to get rid of the local_err
boilerplate.
Propagate the return value so that callers may use it as well to check
failures.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707847357.1489912.2032291280645236480.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
kvm_set_one_reg() returns a negative errno on failure, use that instead
of errno. Also propagate it to callers so they can use it to check
for failures and hopefully get rid of their local_err boilerplate.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707846665.1489912.14267225652103441921.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Callers currently check failures of kvmppc_xive_mmap() through the
@errp argument, which isn't a recommanded practice. It is preferred
to use a return value when possible.
Since NULL isn't an invalid address in theory, it seems better to
return MAP_FAILED and to teach callers to handle it.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707845972.1489912.719896767746375765.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Since kvmppc_xive_source_reset_one() has a return value, convert
kvmppc_xive_source_reset() to use it for error checking. This
allows to get rid of the local_err boiler plate.
Propagate the return value so that callers may use it as well to check
failures.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707845245.1489912.9151822670764690034.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Use error_setg_errno() instead of error_setg(strerror()). While here,
use -ret instead of errno since kvm_vcpu_enable_cap() returns a negative
errno on failure.
Use ERRP_GUARD() to ensure that errp can be passed to error_append_hint(),
and get rid of the local_err boilerplate.
Propagate the return value so that callers may use it as well to check
failures.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159707844549.1489912.4862921680328017645.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
All callers guard these functions with an xive_in_kernel() helper. Make
it clear that they are only to be called when the KVM XIVE device exists.
Note that the check on xive is dropped in kvmppc_xive_disconnect(). It
really cannot be NULL since it comes from set_active_intc() which only
passes pointers to allocated objects.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <159679994169.876294.11026653581505077112.stgit@bahia.lan>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Depending on whether XIVE is emultated or backed with a KVM XIVE device,
the ESB MMIOs of a XIVE source point to an I/O memory region or a mapped
memory region.
This is currently handled by checking kvm_irqchip_in_kernel() returns
false in xive_source_realize(). This is a bit awkward as we usually
need to do extra things when we're using the in-kernel backend, not
less. But most important, we can do better: turn the existing "xive.esb"
memory region into a plain container, introduce an "xive.esb-emulated"
I/O subregion and rename the existing "xive.esb" subregion in the KVM
code to "xive.esb-kvm". Since "xive.esb-kvm" is added with overlap
and a higher priority, it prevails over "xive.esb-emulated" (ie.
a guest using KVM XIVE will interact with "xive.esb-kvm" instead of
the default "xive.esb-emulated" region.
While here, consolidate the computation of the MMIO region size in
a common helper.
Suggested-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159679992680.876294.7520540158586170894.stgit@bahia.lan>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Since this function begins with:
/* The KVM XIVE device is not in use */
if (!xive || xive->fd == -1) {
return;
}
we obviously don't need to check xive->fd again.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159673297296.766512.14780055521619233656.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
If the creation of the KVM XIVE device fails for some reasons, the
negative errno ends up in xive->fd, but the rest of the code assumes
that xive->fd either contains an open fd, ie. positive value, or -1.
This doesn't cause any misbehavior except kvmppc_xive_disconnect()
that will try to close(xive->fd) during rollback and likely be
rewarded with an EBADF.
Only set xive->fd with a open fd.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <159673296585.766512.15404407281299745442.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
And use it instead of reaching out to the machine. This allows to get rid
of a call to qdev_get_machine() and to reduce the scope of another one so
that it is only used within the argument list of error_append_hint(). This
is an acceptable tradeoff compared to all it would require to know about
the maximum number of CPUs here without calling qdev_get_machine().
Signed-off-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20200106145645.4539-10-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The XIVE KVM devices now has an attribute to configure the number of
interrupt servers. This allows to greatly optimize the usage of the VP
space in the XIVE HW, and thus to start a lot more VMs.
Only set this attribute if available in order to support older POWER9
KVM.
The XIVE KVM device now reports the exhaustion of VPs upon the
connection of the first VCPU. Check that in order to have a chance
to provide a hint to the user.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <157478679392.67101.7843580591407950866.stgit@bahia.tlslab.ibm.com>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The XIVE and XICS-on-XIVE KVM devices on POWER9 hosts can greatly reduce
their consumption of some scarce HW resources, namely Virtual Presenter
identifiers, if they know the maximum number of vCPUs that may run in the
VM.
Prepare ground for this by passing the value down to xics_kvm_connect()
and kvmppc_xive_connect(). This is purely mechanical, no functional
change.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <157478678301.67101.2717368060417156338.stgit@bahia.tlslab.ibm.com>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
When using the XIVE KVM device, the trigger page is directly accessible
in QEMU. Unlike with XICS, no need to ask KVM to fire the interrupt. A
simple store on the trigger page does the job.
Just call xive_esb_trigger().
This may improve performance of emulated devices that go through
qemu_set_irq(), eg. virtio devices created with ioeventfd=off or
configured by the guest to use LSI interrupts, which aren't really
recommended setups.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <157408992731.494439.3405812941731584740.stgit@bahia.lan>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Both XICS and XIVE have routines to connect and disconnect KVM with
similar but not identical signatures. This adjusts them to match
exactly, which will be useful for further cleanups later.
While we're there, we add an explicit return value to the connect path
to streamline error reporting in the callers. We remove error
reporting the disconnect path. In the XICS case this wasn't used at
all. In the XIVE case the only error case was if the KVM device was
set up, but KVM didn't have the capability to do so which is pretty
obviously impossible.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
spapr_xive_irq_claim() returns a bool to indicate if it succeeded.
But most of the callers and one callee use int return values and/or an
Error * with more information instead. In any case, ints are a more
common idiom for success/failure states than bools (one never knows
what sense they'll be in).
So instead change to an int return value to indicate presence of error
+ an Error * to describe the details through that call chain.
It also didn't actually check if the irq was already claimed, which is
one of the primary purposes of the claim path, so do that.
spapr_xive_irq_free() also returned a bool... which no callers checked
and was always true, so just drop it.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
A typical pseries VM with 16 vCPUs, one disk, one network adapater
uses less than 100 interrupts but the whole IRQ number space of the
QEMU machine is allocated at reset time and it is 8K wide. This is
wasting a considerable amount of interrupt numbers in the global IRQ
space which has 1M interrupts per socket on a POWER9.
To optimise the HW resources, only request at the KVM level interrupts
which have been claimed by the guest. This will help to increase the
maximum number of VMs per system and also help supporting nested guests
using the XIVE interrupt mode.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190911133937.2716-3-clg@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <156942766014.1274533.10792048853177121231.stgit@bahia.lan>
[dwg: Folded in fix up from Greg Kurz]
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
sysemu/sysemu.h is a rather unfocused dumping ground for stuff related
to the system-emulator. Evidence:
* It's included widely: in my "build everything" tree, changing
sysemu/sysemu.h still triggers a recompile of some 1100 out of 6600
objects (not counting tests and objects that don't depend on
qemu/osdep.h, down from 5400 due to the previous two commits).
* It pulls in more than a dozen additional headers.
Split stuff related to run state management into its own header
sysemu/runstate.h.
Touching sysemu/sysemu.h now recompiles some 850 objects. qemu/uuid.h
also drops from 1100 to 850, and qapi/qapi-types-run-state.h from 4400
to 4200. Touching new sysemu/runstate.h recompiles some 500 objects.
Since I'm touching MAINTAINERS to add sysemu/runstate.h anyway, also
add qemu/main-loop.h.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20190812052359.30071-30-armbru@redhat.com>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
[Unbreak OS-X build]
The migration sequence of a guest using the XIVE exploitation mode
relies on the fact that the states of all devices are restored before
the machine is. This is not true for hot-plug devices such as CPUs
which state come after the machine. This breaks migration because the
thread interrupt context registers are not correctly set.
Fix migration of hotplugged CPUs by restoring their context in the
'post_load' handler of the XiveTCTX model.
Fixes: 277dd3d771 ("spapr/xive: add migration support for KVM")
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190813064853.29310-1-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Make kvmppc_xive_disconnect() able to undo the changes of a partial
execution of kvmppc_xive_connect() and use it to perform rollback.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <156198735673.293938.7313195993600841641.stgit@bahia>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Today, the interrupt device is fully initialized at reset when the CAS
negotiation process has completed. Depending on the KVM capabilities,
the SpaprXive memory regions (ESB, TIMA) are initialized with a host
MMIO backend or a QEMU emulated backend. This results in a complex
initialization sequence partially done at realize and later at reset,
and some memory region leaks.
To simplify this sequence and to remove of the late initialization of
the emulated device which is required to be done only once, we
introduce new memory regions specific for KVM. These regions are
mapped as overlaps on top of the emulated device to make use of the
host MMIOs. Also provide proper cleanups of these regions when the
XIVE KVM device is destroyed to fix the leaks.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190614165920.12670-2-clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Today, when a reset occurs on a pseries machine using the 'dual'
interrupt mode, the KVM devices are released and recreated depending
on the interrupt mode selected by CAS. If XIVE is selected, the SysBus
memory regions of the SpaprXive model are initialized by the KVM
backend initialization routine each time a reset occurs. This leads to
a crash after a couple of resets because the machine reaches the
QDEV_MAX_MMIO limit of SysBusDevice :
qemu-system-ppc64: hw/core/sysbus.c:193: sysbus_init_mmio: Assertion `dev->num_mmio < QDEV_MAX_MMIO' failed.
To fix, initialize the SysBus memory regions in spapr_xive_realize()
called only once and remove the same inits from the QEMU and KVM
backend initialization routines which are called at each reset.
Reported-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190522074016.10521-2-clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The interrupt mode is chosen by the CAS negotiation process and
activated after a reset to take into account the required changes in
the machine. This brings new constraints on how the associated KVM IRQ
device is initialized.
Currently, each model takes care of the initialization of the KVM
device in their realize method but this is not possible anymore as the
initialization needs to be done globaly when the interrupt mode is
known, i.e. when machine is reseted. It also means that we need a way
to delete a KVM device when another mode is chosen.
Also, to support migration, the QEMU objects holding the state to
transfer should always be available but not necessarily activated.
The overall approach of this proposal is to initialize both interrupt
mode at the QEMU level to keep the IRQ number space in sync and to
allow switching from one mode to another. For the KVM side of things,
the whole initialization of the KVM device, sources and presenters, is
grouped in a single routine. The XICS and XIVE sPAPR IRQ reset
handlers are modified accordingly to handle the init and the delete
sequences of the KVM device.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-15-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The activation of the KVM IRQ device depends on the interrupt mode
chosen at CAS time by the machine and some methods used at reset or by
the migration need to be protected.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190513084245.25755-11-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
If a new interrupt mode is chosen by CAS, the machine generates a
reset to reconfigure. At this point, the connection with the previous
KVM device needs to be closed and a new connection needs to opened
with the KVM device operating the chosen interrupt mode.
New routines are introduced to destroy the XICS and the XIVE KVM
devices. They make use of a new KVM device ioctl which destroys the
device and also disconnects the IRQ presenters from the vCPUs.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-10-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
When the VM is stopped, the VM state handler stabilizes the XIVE IC
and marks the EQ pages dirty. These are then transferred to destination
before the transfer of the device vmstates starts.
The SpaprXive interrupt controller model captures the XIVE internal
tables, EAT and ENDT and the XiveTCTX model does the same for the
thread interrupt context registers.
At restart, the SpaprXive 'post_load' method restores all the XIVE
states. It is called by the sPAPR machine 'post_load' method, when all
XIVE states have been transferred and loaded.
Finally, the source states are restored in the VM change state handler
when the machine reaches the running state.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-7-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This handler is in charge of stabilizing the flow of event notifications
in the XIVE controller before migrating a guest. This is a requirement
before transferring the guest EQ pages to a destination.
When the VM is stopped, the handler sets the source PQs to PENDING to
stop the flow of events and to possibly catch a triggered interrupt
occuring while the VM is stopped. Their previous state is saved. The
XIVE controller is then synced through KVM to flush any in-flight
event notification and to stabilize the EQs. At this stage, the EQ
pages are marked dirty to make sure the EQ pages are transferred if a
migration sequence is in progress.
The previous configuration of the sources is restored when the VM
resumes, after a migration or a stop. If an interrupt was queued while
the VM was stopped, the handler simply generates the missing trigger.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-6-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This extends the KVM XIVE device backend with 'synchronize_state'
methods used to retrieve the state from KVM. The HW state of the
sources, the KVM device and the thread interrupt contexts are
collected for the monitor usage and also migration.
These get operations rely on their KVM counterpart in the host kernel
which acts as a proxy for OPAL, the host firmware. The set operations
will be added for migration support later.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Message-Id: <20190513084245.25755-5-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
XIVE hcalls are all redirected to QEMU as none are on a fast path.
When necessary, QEMU invokes KVM through specific ioctls to perform
host operations. QEMU should have done the necessary checks before
calling KVM and, in case of failure, H_HARDWARE is simply returned.
H_INT_ESB is a special case that could have been handled under KVM
but the impact on performance was low when under QEMU. Here are some
figures :
kernel irqchip OFF ON
H_INT_ESB KVM QEMU
rtl8139 (LSI ) 1.19 1.24 1.23 Gbits/sec
virtio 31.80 42.30 -- Gbits/sec
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-4-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This introduces a set of helpers when KVM is in use, which create the
KVM XIVE device, initialize the interrupt sources at a KVM level and
connect the interrupt presenters to the vCPU.
They also handle the initialization of the TIMA and the source ESB
memory regions of the controller. These have a different type under
KVM. They are 'ram device' memory mappings, similarly to VFIO, exposed
to the guest and the associated VMAs on the host are populated
dynamically with the appropriate pages using a fault handler.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190513084245.25755-3-clg@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
