On Fri, 22 Aug 2025 at 10:25, CJ Chen <cjc...@igel.co.jp> wrote: > > Add documentation to clarify that if `.valid.unaligned = true` but > `.impl.unaligned = false`, QEMU’s memory core will automatically split > unaligned guest accesses into multiple aligned accesses. This helps > devices avoid implementing their own unaligned logic, but can be > problematic for devices with side-effect-heavy registers. Also note > that setting `.valid.unaligned = false` together with > `.impl.unaligned = true` is invalid, as it contradicts itself and > will trigger an assertion. > > Signed-off-by: CJ Chen <cjc...@igel.co.jp> > Acked-by: Tomoyuki Hirose <hrstmyk8...@gmail.com> > Suggested-by: Peter Maydell <peter.mayd...@linaro.org> > --- > docs/devel/memory.rst | 18 ++++++++++++++++++ > 1 file changed, 18 insertions(+) > > diff --git a/docs/devel/memory.rst b/docs/devel/memory.rst > index 57fb2aec76..71d7de7ae5 100644 > --- a/docs/devel/memory.rst > +++ b/docs/devel/memory.rst > @@ -365,6 +365,24 @@ callbacks are called: > - .impl.unaligned specifies that the *implementation* supports unaligned > accesses; if false, unaligned accesses will be emulated by two aligned > accesses. > +- Additionally, if .valid.unaligned = true but .impl.unaligned = false, the > + memory core will emulate each unaligned guest access by splitting it into > + multiple aligned sub-accesses.
Can we say in more specific detail what the core handling is, please? I should be able to read this documentation and know exactly what extra accesses I will get to my device. (Then I can make the decision about whether I will be OK with those or if I will need to do the unaligned handling myself.) Documenting the behaviour we intend to provide will also make it easier to review the implementation and the tests in the later patches. > This ensures that devices which only handle > + aligned requests do not need to implement unaligned logic themselves. For > + example, see xhci_cap_ops in hw/usb/hcd-xhci.c: it sets .valid.unaligned > + = true so guests can do unaligned reads on the xHCI Capability Registers, > + while keeping .impl.unaligned = false to rely on the core splitting logic. > + However, if a device’s registers have side effects on read or write, this > + extra splitting can introduce undesired behavior. Specifically, for devices > + whose registers trigger state changes on each read/write, splitting an > access > + can lead to reading or writing bytes beyond the originally requested > subrange > + thereby triggering repeated or otherwise unintended register side effects. > + In such cases, one should set .valid.unaligned = false to reject unaligned > + accesses entirely. .valid.unaligned has to match what the real hardware requires. So we should instead say something like: If your device must support unaligned accesses and these extra accesses would cause unintended side-effects, then you must set .impl.unaligned and handle the unaligned access cases yourself. > +- Conversely, if .valid.unaligned = false but .impl.unaligned = true, > + that setting is considered invalid; it claims unaligned access is allowed > + by the implementation yet disallowed for the device. QEMU enforces this > with > + an assertion to prevent contradictory usage. > > API Reference > ------------- > -- > 2.25.1 > thanks -- PMM
