The raven_io_ops MemoryRegionOps is the only one in the source tree
which sets .valid.unaligned to indicate that it should support
unaligned accesses and which does not also set .impl.unaligned to
indicate that its read and write functions can do the unaligned
handling themselves. This is a problem, because at the moment the
core memory system does not implement the support for handling
unaligned accesses by doing a series of aligned accesses and
combining them (system/memory.c:access_with_adjusted_size() has a
TODO comment noting this).
Fortunately raven_io_read() and raven_io_write() will correctly deal
with the case of being passed an unaligned address, so we can fix the
missing unaligned access support by setting .impl.unaligned in the
MemoryRegionOps struct.
Fixes: 9a1839164c9c8f06 ("raven: Implement non-contiguous I/O region")
Signed-off-by: Peter Maydell <peter.mayd...@linaro.org>
---
Spotted by code inspection: I was looking for devices whose behaviour
might be changed by a patch I'm reviewing that adds that missing
support for unaligned accesses in the core memory system. But even
if we do implement it there, it's more efficient for the raven MR
to correctly mark it as handling unaligned accesses itself.
Tested with 'make check' and 'make check-avocado' only.
---
hw/pci-host/raven.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/hw/pci-host/raven.c b/hw/pci-host/raven.c
index c7a0a2878ab..a7dfddd69ea 100644
--- a/hw/pci-host/raven.c
+++ b/hw/pci-host/raven.c
@@ -200,6 +200,7 @@ static const MemoryRegionOps raven_io_ops = {
.write = raven_io_write,
.endianness = DEVICE_LITTLE_ENDIAN,
.impl.max_access_size = 4,
+ .impl.unaligned = true,
.valid.unaligned = true,
};
--
2.34.1
