When a MSI interrupt is bound to a guest using
xc_domain_update_msi_irq (XEN_DOMCTL_bind_pt_irq) the interrupt is
left masked by default.
This causes problems with guests that first configure interrupts and
clean the per-entry MSIX table mask bit and afterwards enable MSIX
globally. In such scenario the Xen internal msixtbl handlers would not
detect the unmasking of MSIX entries because vectors are not yet
registered since MSIX is not enabled, and vectors would be left
masked.
Introduce a new flag in the gflags field to signal Xen whether a MSI
interrupt should be unmasked after being bound.
This also requires to track the mask register for MSI interrupts, so
QEMU can also notify to Xen whether the MSI interrupt should be bound
masked or unmasked
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reported-by: Andreas Kinzler <hfp@posteo.de>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Check the return status of the xen_host_pci_get_* functions we call in
xen_pt_msix_init(), and fail device init if the reads failed rather than
ploughing ahead. (Spotted by Coverity: CID 777338.)
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
qemu/osdep.h checks whether MAP_ANONYMOUS is defined, but this check
is bogus without a previous inclusion of sys/mman.h. Include it in
sysemu/os-posix.h and remove it from everywhere else.
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On Xen MSIs can be remapped into pirqs, which are a type of event
channels. It's mostly for the benefit of PCI passthrough devices, to
avoid the overhead of interacting with the emulated lapic.
However remapping interrupts and MSIs is also supported for emulated
devices, such as the e1000 and virtio-net.
When an interrupt or an MSI is remapped into a pirq, masking and
unmasking is done by masking and unmasking the event channel. The
masking bit on the PCI config space or MSI-X table should be ignored,
but it isn't at the moment.
As a consequence emulated devices which use MSI or MSI-X, such as
virtio-net, don't work properly (the guest doesn't receive any
notifications). The mechanism was working properly when xen_apic was
introduced, but I haven't narrowed down which commit in particular is
causing the regression.
Fix the issue by ignoring the masking bit for MSI and MSI-X which have
been remapped into pirqs.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Clean up includes so that osdep.h is included first and headers
which it implies are not included manually.
This commit was created with scripts/clean-includes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 1453832250-766-14-git-send-email-peter.maydell@linaro.org
The way the generic infrastructure works the intention of not allowing
unaligned accesses can't be achieved by simply setting .unaligned to
false. The benefit is that we can now replace the conditionals in
{get,set}_entry_value() by assert()-s.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
The remaining log message in pci_msix_write() is wrong, as there guest
behavior may only appear to be wrong: For one, the old logic didn't
take the mask-all bit into account. And then this shouldn't depend on
host device state (i.e. the host may have masked the entry without the
guest having done so). Plus these writes shouldn't be dropped even when
an entry gets unmasked. Instead, if they can't be made take effect
right away, they should take effect on the next unmasking or enabling
operation - the specification explicitly describes such caching
behavior.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
msix->mmio is added to XenPCIPassthroughState's object as property.
object_finalize_child_property is called for XenPCIPassthroughState's
object, which calls object_property_del_all, which is going to try to
delete msix->mmio. object_finalize_child_property() will access
msix->mmio's obj. But the whole msix struct has already been freed
by xen_pt_msix_delete. This will cause segment fault when msix->mmio
has been overwritten.
This patch is to fix the issue.
Signed-off-by: Lan Tianyu <tianyu.lan@intel.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
We seem to only use these functions when de-activating the
MSI - so just log errors.
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
As we do not use it outside our code.
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Functions setting errno commonly return just -1, which is of no
particular use in the log file.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Limit error messages resulting from bad guest behavior to avoid allowing
the guest to cause the control domain's disk to fill.
The first message in pci_msix_write() can simply be deleted, as this
is indeed bad guest behavior, but such out of bounds writes don't
really need to be logged.
The second one is more problematic, as there guest behavior may only
appear to be wrong: For one, the old logic didn't take the mask-all bit
into account. And then this shouldn't depend on host device state (i.e.
the host may have masked the entry without the guest having done so).
Plus these writes shouldn't be dropped even when an entry is unmasked.
Instead, if they can't be made take effect right away, they should take
effect on the next unmasking or enabling operation - the specification
explicitly describes such caching behavior. Until we can validly drop
the message (implementing such caching/latching behavior), issue the
message just once per MSI-X table entry.
Note that the log message in pci_msix_read() similar to the one being
removed here is not an issue: "addr" being of unsigned type, and the
maximum size of the MSI-X table being 32k, entry_nr simply can't be
negative and hence the conditonal guarding issuing of the message will
never be true.
This is XSA-130.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
The function is empty after the previous patch, so remove it.
Reviewed-by: Peter Crosthwaite <peter.crosthwaite@xilinx.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pirqs are not freed when driver unload, then new pirqs are allocated when
driver reload. This could exhaust pirqs if do it in a loop.
This patch fixes the bug by freeing pirqs when ENABLE bit is cleared in
msi/msix control reg.
There is also other way of fixing it such as reuse pirqs between driver reload,
but this way is better.
Xen-devel: http://marc.info/?l=xen-devel&m=136800120304275&w=2
Signed-off-by: Zhenzhong Duan <zhenzhong.duan@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
