Legacy virtio defines the virtqueue base using a 32-bit PFN field, with
a read-only register indicating a fixed page size of 4k.
This can cause problems for DMA allocators that allocate top down from
the DMA mask, which is set to 64 bits. In this case, the addresses are
silently truncated to 44-bit, leading to IOMMU faults, failure to read
from the queue or data corruption.
This patch restricts the coherent DMA mask for legacy PCI virtio devices
to 44 bits, which matches the specification.
Cc: stable@vger.kernel.org
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Benjamin Serebrin <serebrin@google.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This switches to vring_create_virtqueue, simplifying the driver and
adding DMA API support.
This fixes virtio-pci on platforms and busses that have IOMMUs. This
will break the experimental QEMU Q35 IOMMU support until QEMU is
fixed. In exchange, it fixes physical virtio hardware as well as
virtio-pci running under Xen.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Move resource allocation from common code to legacy and modern code.
Only request resources actually used, i.e. bar0 in legacy mode and
the bar(s) specified by capabilities in modern mode.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Most of initialization is device-independent.
Let's move it to common.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Device VQs were getting freed twice: once in every device's removal
functions, and then again in virtio_pci_legacy_remove(). The ones in
devices are called first, so drop the useless second call.
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The reason we defer kfree until release function is because it's a
general rule for kobjects: kfree of the reference counter itself is only
legal in the release function.
Previous patch didn't make this clear, document this in code.
Cc: stable@vger.kernel.org
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
A struct device which has just been unregistered can live on past the
point at which a driver decides to drop it's initial reference to the
kobject gained on allocation.
This implies that when releasing a virtio device, we can't free a struct
virtio_device until the underlying struct device has been released,
which might not happen immediately on device_unregister().
Unfortunately, this is exactly what virtio pci does:
it has an empty release callback, and frees memory immediately
after unregistering the device.
This causes an easy to reproduce crash if CONFIG_DEBUG_KOBJECT_RELEASE
it enabled.
To fix, free the memory only once we know the device is gone in the release
callback.
Cc: stable@vger.kernel.org
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
It turns out we need to add device-specific code
in release callback. Move it to virtio_pci_legacy.c.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
kbuild does not seem to like it when we name source
files same as the module.
Let's rename virtio_pci -> virtio_pci_common,
and get rid of #include-ing c files.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Move everything dealing with legacy devices out to virtio_pci_legacy.c.
Expose common code APIs in virtio_pci.h
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Will Deacon