When an interrupt has been handled, the OS notifies the interrupt
controller with a EOI sequence. On a POWER9 system using the XIVE
interrupt controller, this can be done with a load or a store
operation on the ESB interrupt management page of the interrupt. The
StoreEOI operation has less latency and improves interrupt handling
performance but it was deactivated during the POWER9 DD2.0 timeframe
because of ordering issues. We use the LoadEOI today but we plan to
reactivate StoreEOI in future architectures.
There is usually no need to enforce ordering between ESB load and
store operations as they should lead to the same result. E.g. a store
trigger and a load EOI can be executed in any order. Assuming the
interrupt state is PQ=10, a store trigger followed by a load EOI will
return a Q bit. In the reverse order, it will create a new interrupt
trigger from HW. In both cases, the handler processing interrupts is
notified.
In some cases, the XIVE_ESB_SET_PQ_10 load operation is used to
disable temporarily the interrupt source (mask/unmask). When the
source is reenabled, the OS can detect if interrupts were received
while the source was disabled and reinject them. This process needs
special care when StoreEOI is activated. The ESB load and store
operations should be correctly ordered because a XIVE_ESB_STORE_EOI
operation could leave the source enabled if it has not completed
before the loads.
For those cases, we enforce Load-after-Store ordering with a special
load operation offset. To avoid performance impact, this ordering is
only enforced when really needed, that is when interrupt sources are
temporarily disabled with the XIVE_ESB_SET_PQ_10 load. It should not
be needed for other loads.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200220081506.31209-1-clg@kaod.org
Use kvm_vcpu_gfn_to_hva() when retrieving the host page size so that the
correct set of memslots is used when handling x86 page faults in SMM.
Fixes: 54bf36aac5 ("KVM: x86: use vcpu-specific functions to read/write/translate GFNs")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We need to check the host page size is big enough to accomodate the
EQ. Let's do this before taking a reference on the EQ page to avoid
a potential leak if the check fails.
Cc: stable@vger.kernel.org # v5.2
Fixes: 13ce3297c5 ("KVM: PPC: Book3S HV: XIVE: Add controls for the EQ configuration")
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The EQ page is allocated by the guest and then passed to the hypervisor
with the H_INT_SET_QUEUE_CONFIG hcall. A reference is taken on the page
before handing it over to the HW. This reference is dropped either when
the guest issues the H_INT_RESET hcall or when the KVM device is released.
But, the guest can legitimately call H_INT_SET_QUEUE_CONFIG several times,
either to reset the EQ (vCPU hot unplug) or to set a new EQ (guest reboot).
In both cases the existing EQ page reference is leaked because we simply
overwrite it in the XIVE queue structure without calling put_page().
This is especially visible when the guest memory is backed with huge pages:
start a VM up to the guest userspace, either reboot it or unplug a vCPU,
quit QEMU. The leak is observed by comparing the value of HugePages_Free in
/proc/meminfo before and after the VM is run.
Ideally we'd want the XIVE code to handle the EQ page de-allocation at the
platform level. This isn't the case right now because the various XIVE
drivers have different allocation needs. It could maybe worth introducing
hooks for this purpose instead of exposing XIVE internals to the drivers,
but this is certainly a huge work to be done later.
In the meantime, for easier backport, fix both vCPU unplug and guest reboot
leaks by introducing a wrapper around xive_native_configure_queue() that
does the necessary cleanup.
Reported-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org # v5.2
Fixes: 13ce3297c5 ("KVM: PPC: Book3S HV: XIVE: Add controls for the EQ configuration")
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Tested-by: Lijun Pan <ljp@linux.ibm.com>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Add a new attribute to both XIVE and XICS-on-XIVE KVM devices so that
userspace can tell how many interrupt servers it needs. If a VM needs
less than the current default of KVM_MAX_VCPUS (2048), we can allocate
less VPs in OPAL. Combined with a core stride (VSMT) that matches the
number of guest threads per core, this may substantially increases the
number of VMs that can run concurrently with an in-kernel XIVE device.
Since the legacy XIVE KVM device is exposed to userspace through the
XICS KVM API, a new attribute group is added to it for this purpose.
While here, fix the syntax of the existing KVM_DEV_XICS_GRP_SOURCES
in the XICS documentation.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The XIVE VP is an internal structure which allow the XIVE interrupt
controller to maintain the interrupt context state of vCPUs non
dispatched on HW threads.
When a guest is started, the XIVE KVM device allocates a block of
XIVE VPs in OPAL, enough to accommodate the highest possible vCPU
id KVM_MAX_VCPU_ID (16384) packed down to KVM_MAX_VCPUS (2048).
With a guest's core stride of 8 and a threading mode of 1 (QEMU's
default), a VM must run at least 256 vCPUs to actually need such a
range of VPs.
A POWER9 system has a limited XIVE VP space : 512k and KVM is
currently wasting this HW resource with large VP allocations,
especially since a typical VM likely runs with a lot less vCPUs.
Make the size of the VP block configurable. Add an nr_servers
field to the XIVE structure and a function to set it for this
purpose.
Split VP allocation out of the device create function. Since the
VP block isn't used before the first vCPU connects to the XIVE KVM
device, allocation is now performed by kvmppc_xive_connect_vcpu().
This gives the opportunity to set nr_servers in between:
kvmppc_xive_create() / kvmppc_xive_native_create()
.
.
kvmppc_xive_set_nr_servers()
.
.
kvmppc_xive_connect_vcpu() / kvmppc_xive_native_connect_vcpu()
The connect_vcpu() functions check that the vCPU id is below nr_servers
and if it is the first vCPU they allocate the VP block. This is protected
against a concurrent update of nr_servers by kvmppc_xive_set_nr_servers()
with the xive->lock mutex.
Also, the block is allocated once for the device lifetime: nr_servers
should stay constant otherwise connect_vcpu() could generate a boggus
VP id and likely crash OPAL. It is thus forbidden to update nr_servers
once the block is allocated.
If the VP allocation fail, return ENOSPC which seems more appropriate to
report the depletion of system wide HW resource than ENOMEM or ENXIO.
A VM using a stride of 8 and 1 thread per core with 32 vCPUs would hence
only need 256 VPs instead of 2048. If the stride is set to match the number
of threads per core, this goes further down to 32.
This will be exposed to userspace by a subsequent patch.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Reduce code duplication by consolidating the checking of vCPU ids and VP
ids to a common helper used by both legacy and native XIVE KVM devices.
And explain the magic with a comment.
Signed-off-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Print out the VP id of each connected vCPU, this allow to see:
- the VP block base in which OPAL encodes information that may be
useful when debugging
- the packed vCPU id which may differ from the raw vCPU id if the
latter is >= KVM_MAX_VCPUS (2048)
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
If we cannot allocate the XIVE VPs in OPAL, the creation of a XIVE or
XICS-on-XIVE device is aborted as expected, but we leave kvm->arch.xive
set forever since the release method isn't called in this case. Any
subsequent tentative to create a XIVE or XICS-on-XIVE for this VM will
thus always fail (DoS). This is a problem for QEMU since it destroys
and re-creates these devices when the VM is reset: the VM would be
restricted to using the much slower emulated XIVE or XICS forever.
As an alternative to adding rollback, do not assign kvm->arch.xive before
making sure the XIVE VPs are allocated in OPAL.
Cc: stable@vger.kernel.org # v5.2
Fixes: 5422e95103 ("KVM: PPC: Book3S HV: XIVE: Replace the 'destroy' method by a 'release' method")
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
There are some POWER9 machines where the OPAL firmware does not support
the OPAL_XIVE_GET_QUEUE_STATE and OPAL_XIVE_SET_QUEUE_STATE calls.
The impact of this is that a guest using XIVE natively will not be able
to be migrated successfully. On the source side, the get_attr operation
on the KVM native device for the KVM_DEV_XIVE_GRP_EQ_CONFIG attribute
will fail; on the destination side, the set_attr operation for the same
attribute will fail.
This adds tests for the existence of the OPAL get/set queue state
functions, and if they are not supported, the XIVE-native KVM device
is not created and the KVM_CAP_PPC_IRQ_XIVE capability returns false.
Userspace can then either provide a software emulation of XIVE, or
else tell the guest that it does not have a XIVE controller available
to it.
Cc: stable@vger.kernel.org # v5.2+
Fixes: 3fab2d1058 ("KVM: PPC: Book3S HV: XIVE: Activate XIVE exploitation mode")
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Testing has revealed the existence of a race condition where a XIVE
interrupt being shut down can be in one of the XIVE interrupt queues
(of which there are up to 8 per CPU, one for each priority) at the
point where free_irq() is called. If this happens, can return an
interrupt number which has been shut down. This can lead to various
symptoms:
- irq_to_desc(irq) can be NULL. In this case, no end-of-interrupt
function gets called, resulting in the CPU's elevated interrupt
priority (numerically lowered CPPR) never gets reset. That then
means that the CPU stops processing interrupts, causing device
timeouts and other errors in various device drivers.
- The irq descriptor or related data structures can be in the process
of being freed as the interrupt code is using them. This typically
leads to crashes due to bad pointer dereferences.
This race is basically what commit 62e0468650 ("genirq: Add optional
hardware synchronization for shutdown", 2019-06-28) is intended to
fix, given a get_irqchip_state() method for the interrupt controller
being used. It works by polling the interrupt controller when an
interrupt is being freed until the controller says it is not pending.
With XIVE, the PQ bits of the interrupt source indicate the state of
the interrupt source, and in particular the P bit goes from 0 to 1 at
the point where the hardware writes an entry into the interrupt queue
that this interrupt is directed towards. Normally, the code will then
process the interrupt and do an end-of-interrupt (EOI) operation which
will reset PQ to 00 (assuming another interrupt hasn't been generated
in the meantime). However, there are situations where the code resets
P even though a queue entry exists (for example, by setting PQ to 01,
which disables the interrupt source), and also situations where the
code leaves P at 1 after removing the queue entry (for example, this
is done for escalation interrupts so they cannot fire again until
they are explicitly re-enabled).
The code already has a 'saved_p' flag for the interrupt source which
indicates that a queue entry exists, although it isn't maintained
consistently. This patch adds a 'stale_p' flag to indicate that
P has been left at 1 after processing a queue entry, and adds code
to set and clear saved_p and stale_p as necessary to maintain a
consistent indication of whether a queue entry may or may not exist.
With this, we can implement xive_get_irqchip_state() by looking at
stale_p, saved_p and the ESB PQ bits for the interrupt.
There is some additional code to handle escalation interrupts
properly; because they are enabled and disabled in KVM assembly code,
which does not have access to the xive_irq_data struct for the
escalation interrupt. Hence, stale_p may be incorrect when the
escalation interrupt is freed in kvmppc_xive_{,native_}cleanup_vcpu().
Fortunately, we can fix it up by looking at vcpu->arch.xive_esc_on,
with some careful attention to barriers in order to ensure the correct
result if xive_esc_irq() races with kvmppc_xive_cleanup_vcpu().
Finally, this adds code to make noise on the console (pr_crit and
WARN_ON(1)) if we find an interrupt queue entry for an interrupt
which does not have a descriptor. While this won't catch the race
reliably, if it does get triggered it will be an indication that
the race is occurring and needs to be debugged.
Fixes: 243e25112d ("powerpc/xive: Native exploitation of the XIVE interrupt controller")
Cc: stable@vger.kernel.org # v4.12+
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20190813100648.GE9567@blackberry
At present, when running a guest on POWER9 using HV KVM but not using
an in-kernel interrupt controller (XICS or XIVE), for example if QEMU
is run with the kernel_irqchip=off option, the guest entry code goes
ahead and tries to load the guest context into the XIVE hardware, even
though no context has been set up.
To fix this, we check that the "CAM word" is non-zero before pushing
it to the hardware. The CAM word is initialized to a non-zero value
in kvmppc_xive_connect_vcpu() and kvmppc_xive_native_connect_vcpu(),
and is now cleared in kvmppc_xive_{,native_}cleanup_vcpu.
Fixes: 5af5099385 ("KVM: PPC: Book3S HV: Native usage of the XIVE interrupt controller")
Cc: stable@vger.kernel.org # v4.12+
Reported-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20190813100100.GC9567@blackberry
When a vCPU is brought done, the XIVE VP (Virtual Processor) is first
disabled and then the event notification queues are freed. When freeing
the queues, we check for possible escalation interrupts and free them
also.
But when a XIVE VP is disabled, the underlying XIVE ENDs also are
disabled in OPAL. When an END (Event Notification Descriptor) is
disabled, its ESB pages (ESn and ESe) are disabled and loads return all
1s. Which means that any access on the ESB page of the escalation
interrupt will return invalid values.
When an interrupt is freed, the shutdown handler computes a 'saved_p'
field from the value returned by a load in xive_do_source_set_mask().
This value is incorrect for escalation interrupts for the reason
described above.
This has no impact on Linux/KVM today because we don't make use of it
but we will introduce in future changes a xive_get_irqchip_state()
handler. This handler will use the 'saved_p' field to return the state
of an interrupt and 'saved_p' being incorrect, softlockup will occur.
Fix the vCPU cleanup sequence by first freeing the escalation interrupts
if any, then disable the XIVE VP and last free the queues.
Fixes: 90c73795af ("KVM: PPC: Book3S HV: Add a new KVM device for the XIVE native exploitation mode")
Fixes: 5af5099385 ("KVM: PPC: Book3S HV: Native usage of the XIVE interrupt controller")
Cc: stable@vger.kernel.org # v4.12+
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20190806172538.5087-1-clg@kaod.org
The XIVE device structure is now allocated in kvmppc_xive_get_device()
and kfree'd in kvmppc_core_destroy_vm(). In case of an OPAL error when
allocating the XIVE VPs, the kfree() call in kvmppc_xive_*create()
will result in a double free and corrupt the host memory.
Fixes: 5422e95103 ("KVM: PPC: Book3S HV: XIVE: Replace the 'destroy' method by a 'release' method")
Cc: stable@vger.kernel.org # v5.2+
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/6ea6998b-a890-2511-01d1-747d7621eb19@kaod.org
Under XIVE, the ESB pages of an interrupt are used for interrupt
management (EOI) and triggering. They are made available to guests
through a mapping of the XIVE KVM device.
When a device is passed-through, the passthru_irq helpers,
kvmppc_xive_set_mapped() and kvmppc_xive_clr_mapped(), clear the ESB
pages of the guest IRQ number being mapped and let the VM fault
handler repopulate with the correct page.
The ESB pages are mapped at offset 4 (KVM_XIVE_ESB_PAGE_OFFSET) in the
KVM device mapping. Unfortunately, this offset was not taken into
account when clearing the pages. This lead to issues with the
passthrough devices for which the interrupts were not functional under
some guest configuration (tg3 and single CPU) or in any configuration
(e1000e adapter).
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
According to Documentation/virtual/kvm/locking.txt, the srcu read lock
should be taken when accessing the memslots of the VM. The XIVE KVM
device needs to do so when configuring the page of the OS event queue
of vCPU for a given priority and when marking the same page dirty
before migration.
This avoids warnings such as :
[ 208.224882] =============================
[ 208.224884] WARNING: suspicious RCU usage
[ 208.224889] 5.2.0-rc2-xive+ #47 Not tainted
[ 208.224890] -----------------------------
[ 208.224894] ../include/linux/kvm_host.h:633 suspicious rcu_dereference_check() usage!
[ 208.224896]
other info that might help us debug this:
[ 208.224898]
rcu_scheduler_active = 2, debug_locks = 1
[ 208.224901] no locks held by qemu-system-ppc/3923.
[ 208.224902]
stack backtrace:
[ 208.224907] CPU: 64 PID: 3923 Comm: qemu-system-ppc Kdump: loaded Not tainted 5.2.0-rc2-xive+ #47
[ 208.224909] Call Trace:
[ 208.224918] [c000200cdd98fa30] [c000000000be1934] dump_stack+0xe8/0x164 (unreliable)
[ 208.224924] [c000200cdd98fa80] [c0000000001aec80] lockdep_rcu_suspicious+0x110/0x180
[ 208.224935] [c000200cdd98fb00] [c0080000075933a0] gfn_to_memslot+0x1c8/0x200 [kvm]
[ 208.224943] [c000200cdd98fb40] [c008000007599600] gfn_to_pfn+0x28/0x60 [kvm]
[ 208.224951] [c000200cdd98fb70] [c008000007599658] gfn_to_page+0x20/0x40 [kvm]
[ 208.224959] [c000200cdd98fb90] [c0080000075b495c] kvmppc_xive_native_set_attr+0x8b4/0x1480 [kvm]
[ 208.224967] [c000200cdd98fca0] [c00800000759261c] kvm_device_ioctl_attr+0x64/0xb0 [kvm]
[ 208.224974] [c000200cdd98fcf0] [c008000007592730] kvm_device_ioctl+0xc8/0x110 [kvm]
[ 208.224979] [c000200cdd98fd10] [c000000000433a24] do_vfs_ioctl+0xd4/0xcd0
[ 208.224981] [c000200cdd98fdb0] [c000000000434724] ksys_ioctl+0x104/0x120
[ 208.224984] [c000200cdd98fe00] [c000000000434768] sys_ioctl+0x28/0x80
[ 208.224988] [c000200cdd98fe20] [c00000000000b888] system_call+0x5c/0x70
legoater@boss01:~$
Fixes: 13ce3297c5 ("KVM: PPC: Book3S HV: XIVE: Add controls for the EQ configuration")
Fixes: e6714bd167 ("KVM: PPC: Book3S HV: XIVE: Add a control to dirty the XIVE EQ pages")
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The XICS-on-XIVE KVM device needs to allocate XIVE event queues when a
priority is used by the OS. This is referred as EQ provisioning and it
is done under the hood when :
1. a CPU is hot-plugged in the VM
2. the "set-xive" is called at VM startup
3. sources are restored at VM restore
The kvm->lock mutex is used to protect the different XIVE structures
being modified but in some contexts, kvm->lock is taken under the
vcpu->mutex which is not permitted by the KVM locking rules.
Introduce a new mutex 'lock' for the KVM devices for them to
synchronize accesses to the XIVE device structures.
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
When a vCPU is connected to the KVM device, it is done using its vCPU
identifier in the guest. Fix the enforced limit on the vCPU identifier
by taking into account the SMT mode.
Reported-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Tested-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
When a CPU is hot-unplugged, the EQ is deconfigured using a zero size
and a zero address. In this case, there is no need to check the flag
and queue size validity. Move the checks after the queue reset code
section to fix CPU hot-unplug.
Reported-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Tested-by: Satheesh Rajendran <sathnaga@linux.vnet.ibm.com>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Improve the release of the XIVE KVM device by clearing the file
address_space, which is used to unmap the interrupt ESB pages when a
device is passed-through.
Suggested-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Currently, kvmppc_xive_release() and kvmppc_xive_native_release() clear
kvm->arch.mmu_ready and call kick_all_cpus_sync() as a way of ensuring
that no vcpus are executing in the guest. However, future patches will
change the mutex associated with kvm->arch.mmu_ready to a new mutex that
nests inside the vcpu mutexes, making it difficult to continue to use
this method.
In fact, taking the vcpu mutex for a vcpu excludes execution of that
vcpu, and we already take the vcpu mutex around the call to
kvmppc_xive_[native_]cleanup_vcpu(). Once the cleanup function is
done and we release the vcpu mutex, the vcpu can execute once again,
but because we have cleared vcpu->arch.xive_vcpu, vcpu->arch.irq_type,
vcpu->arch.xive_esc_vaddr and vcpu->arch.xive_esc_raddr, that vcpu will
not be going into XIVE code any more. Thus, once we have cleaned up
all of the vcpus, we are safe to clean up the rest of the XIVE state,
and we don't need to use kvm->arch.mmu_ready to hold off vcpu execution.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
There is a spelling mistake in a pr_err message, fix it.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Now that we have the possibility of a XIVE or XICS-on-XIVE device being
released while the VM is still running, we need to be careful about
races and potential use-after-free bugs. Although the kvmppc_xive
struct is not freed, but kept around for re-use, the kvmppc_xive_vcpu
structs are freed, and they are used extensively in both the XIVE native
and XICS-on-XIVE code.
There are various ways in which XIVE code gets invoked:
- VCPU entry and exit, which do push and pull operations on the XIVE hardware
- one_reg get and set functions (vcpu->mutex is held)
- XICS hypercalls (but only inside guest execution, not from
kvmppc_pseries_do_hcall)
- device creation calls (kvm->lock is held)
- device callbacks - get/set attribute, mmap, pagefault, release/destroy
- set_mapped/clr_mapped calls (kvm->lock is held)
- connect_vcpu calls
- debugfs file read callbacks
Inside a device release function, we know that userspace cannot have an
open file descriptor referring to the device, nor can it have any mmapped
regions from the device. Therefore the device callbacks are excluded,
as are the connect_vcpu calls (since they need a fd for the device).
Further, since the caller holds the kvm->lock mutex, no other device
creation calls or set/clr_mapped calls can be executing concurrently.
To exclude VCPU execution and XICS hypercalls, we temporarily set
kvm->arch.mmu_ready to 0. This forces any VCPU task that is trying to
enter the guest to take the kvm->lock mutex, which is held by the caller
of the release function. Then, sending an IPI to all other CPUs forces
any VCPU currently executing in the guest to exit.
Finally, we take the vcpu->mutex for each VCPU around the process of
cleaning up and freeing its XIVE data structures, in order to exclude
any one_reg get/set calls.
To exclude the debugfs read callbacks, we just need to ensure that
debugfs_remove is called before freeing any data structures. Once it
returns we know that no CPU can be executing the callbacks (for our
kvmppc_xive instance).
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
When a P9 sPAPR VM boots, the CAS negotiation process determines which
interrupt mode to use (XICS legacy or XIVE native) and invokes a
machine reset to activate the chosen mode.
We introduce 'release' methods for the XICS-on-XIVE and the XIVE
native KVM devices which are called when the file descriptor of the
device is closed after the TIMA and ESB pages have been unmapped.
They perform the necessary cleanups : clear the vCPU interrupt
presenters that could be attached and then destroy the device. The
'release' methods replace the 'destroy' methods as 'destroy' is not
called anymore once 'release' is. Compatibility with older QEMU is
nevertheless maintained.
This is not considered as a safe operation as the vCPUs are still
running and could be referencing the KVM device through their
presenters. To protect the system from any breakage, the kvmppc_xive
objects representing both KVM devices are now stored in an array under
the VM. Allocation is performed on first usage and memory is freed
only when the VM exits.
[paulus@ozlabs.org - Moved freeing of xive structures to book3s.c,
put it under #ifdef CONFIG_KVM_XICS.]
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The KVM XICS-over-XIVE device and the proposed KVM XIVE native device
implement an IRQ space for the guest using the generic IPI interrupts
of the XIVE IC controller. These interrupts are allocated at the OPAL
level and "mapped" into the guest IRQ number space in the range 0-0x1FFF.
Interrupt management is performed in the XIVE way: using loads and
stores on the addresses of the XIVE IPI interrupt ESB pages.
Both KVM devices share the same internal structure caching information
on the interrupts, among which the xive_irq_data struct containing the
addresses of the IPI ESB pages and an extra one in case of pass-through.
The later contains the addresses of the ESB pages of the underlying HW
controller interrupts, PHB4 in all cases for now.
A guest, when running in the XICS legacy interrupt mode, lets the KVM
XICS-over-XIVE device "handle" interrupt management, that is to
perform the loads and stores on the addresses of the ESB pages of the
guest interrupts. However, when running in XIVE native exploitation
mode, the KVM XIVE native device exposes the interrupt ESB pages to
the guest and lets the guest perform directly the loads and stores.
The VMA exposing the ESB pages make use of a custom VM fault handler
which role is to populate the VMA with appropriate pages. When a fault
occurs, the guest IRQ number is deduced from the offset, and the ESB
pages of associated XIVE IPI interrupt are inserted in the VMA (using
the internal structure caching information on the interrupts).
Supporting device passthrough in the guest running in XIVE native
exploitation mode adds some extra refinements because the ESB pages
of a different HW controller (PHB4) need to be exposed to the guest
along with the initial IPI ESB pages of the XIVE IC controller. But
the overall mechanic is the same.
When the device HW irqs are mapped into or unmapped from the guest
IRQ number space, the passthru_irq helpers, kvmppc_xive_set_mapped()
and kvmppc_xive_clr_mapped(), are called to record or clear the
passthrough interrupt information and to perform the switch.
The approach taken by this patch is to clear the ESB pages of the
guest IRQ number being mapped and let the VM fault handler repopulate.
The handler will insert the ESB page corresponding to the HW interrupt
of the device being passed-through or the initial IPI ESB page if the
device is being removed.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Each source is associated with an Event State Buffer (ESB) with a
even/odd pair of pages which provides commands to manage the source:
to trigger, to EOI, to turn off the source for instance.
The custom VM fault handler will deduce the guest IRQ number from the
offset of the fault, and the ESB page of the associated XIVE interrupt
will be inserted into the VMA using the internal structure caching
information on the interrupts.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Each thread has an associated Thread Interrupt Management context
composed of a set of registers. These registers let the thread handle
priority management and interrupt acknowledgment. The most important
are :
- Interrupt Pending Buffer (IPB)
- Current Processor Priority (CPPR)
- Notification Source Register (NSR)
They are exposed to software in four different pages each proposing a
view with a different privilege. The first page is for the physical
thread context and the second for the hypervisor. Only the third
(operating system) and the fourth (user level) are exposed the guest.
A custom VM fault handler will populate the VMA with the appropriate
pages, which should only be the OS page for now.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The state of the thread interrupt management registers needs to be
collected for migration. These registers are cached under the
'xive_saved_state.w01' field of the VCPU when the VPCU context is
pulled from the HW thread. An OPAL call retrieves the backup of the
IPB register in the underlying XIVE NVT structure and merges it in the
KVM state.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
When migration of a VM is initiated, a first copy of the RAM is
transferred to the destination before the VM is stopped, but there is
no guarantee that the EQ pages in which the event notifications are
queued have not been modified.
To make sure migration will capture a consistent memory state, the
XIVE device should perform a XIVE quiesce sequence to stop the flow of
event notifications and stabilize the EQs. This is the purpose of the
KVM_DEV_XIVE_EQ_SYNC control which will also marks the EQ pages dirty
to force their transfer.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This control will be used by the H_INT_SYNC hcall from QEMU to flush
event notifications on the XIVE IC owning the source.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This control is to be used by the H_INT_RESET hcall from QEMU. Its
purpose is to clear all configuration of the sources and EQs. This is
necessary in case of a kexec (for a kdump kernel for instance) to make
sure that no remaining configuration is left from the previous boot
setup so that the new kernel can start safely from a clean state.
The queue 7 is ignored when the XIVE device is configured to run in
single escalation mode. Prio 7 is used by escalations.
The XIVE VP is kept enabled as the vCPU is still active and connected
to the XIVE device.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
These controls will be used by the H_INT_SET_QUEUE_CONFIG and
H_INT_GET_QUEUE_CONFIG hcalls from QEMU to configure the underlying
Event Queue in the XIVE IC. They will also be used to restore the
configuration of the XIVE EQs and to capture the internal run-time
state of the EQs. Both 'get' and 'set' rely on an OPAL call to access
the EQ toggle bit and EQ index which are updated by the XIVE IC when
event notifications are enqueued in the EQ.
The value of the guest physical address of the event queue is saved in
the XIVE internal xive_q structure for later use. That is when
migration needs to mark the EQ pages dirty to capture a consistent
memory state of the VM.
To be noted that H_INT_SET_QUEUE_CONFIG does not require the extra
OPAL call setting the EQ toggle bit and EQ index to configure the EQ,
but restoring the EQ state will.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This control will be used by the H_INT_SET_SOURCE_CONFIG hcall from
QEMU to configure the target of a source and also to restore the
configuration of a source when migrating the VM.
The XIVE source interrupt structure is extended with the value of the
Effective Interrupt Source Number. The EISN is the interrupt number
pushed in the event queue that the guest OS will use to dispatch
events internally. Caching the EISN value in KVM eases the test when
checking if a reconfiguration is indeed needed.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The XIVE KVM device maintains a list of interrupt sources for the VM
which are allocated in the pool of generic interrupts (IPIs) of the
main XIVE IC controller. These are used for the CPU IPIs as well as
for virtual device interrupts. The IRQ number space is defined by
QEMU.
The XIVE device reuses the source structures of the XICS-on-XIVE
device for the source blocks (2-level tree) and for the source
interrupts. Under XIVE native, the source interrupt caches mostly
configuration information and is less used than under the XICS-on-XIVE
device in which hcalls are still necessary at run-time.
When a source is initialized in KVM, an IPI interrupt source is simply
allocated at the OPAL level and then MASKED. KVM only needs to know
about its type: LSI or MSI.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The user interface exposes a new capability KVM_CAP_PPC_IRQ_XIVE to
let QEMU connect the vCPU presenters to the XIVE KVM device if
required. The capability is not advertised for now as the full support
for the XIVE native exploitation mode is not yet available. When this
is case, the capability will be advertised on PowerNV Hypervisors
only. Nested guests (pseries KVM Hypervisor) are not supported.
Internally, the interface to the new KVM device is protected with a
new interrupt mode: KVMPPC_IRQ_XIVE.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This is the basic framework for the new KVM device supporting the XIVE
native exploitation mode. The user interface exposes a new KVM device
to be created by QEMU, only available when running on a L0 hypervisor.
Support for nested guests is not available yet.
The XIVE device reuses the device structure of the XICS-on-XIVE device
as they have a lot in common. That could possibly change in the future
if the need arise.
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
