On Tue, Feb 25, 2020 at 09:24:34AM +0000, Stefan Hajnoczi wrote: > On Mon, Feb 24, 2020 at 12:10:25PM -0500, Michael S. Tsirkin wrote: > > On Sat, Feb 22, 2020 at 08:19:16PM +0000, Stefan Hajnoczi wrote: > > > The KVM_IOREGIONFD_POSTED_WRITES flag > > > skips waiting for an acknowledgement on write accesses. This is > > > suitable for accesses that do not require synchronous emulation, such as > > > doorbell register writes. > > > > I would avoid hacks like this until we understand this better. > > Specificlly one needs to be very careful since memory ordering semantics > > can differ between a write into an uncacheable range and host writes into > > a data structure. Reads from one region are also assumed to be ordered with > > writes to another region, and drivers are known to make assumptions > > like this. > > > > Memory ordering being what it is, this isn't a field I'd be comfortable > > device writes know what they are doing. > > Unlike PCI Posted Writes the idea is not to let the write operations sit > in a cache. They will be sent immediately just like ioeventfd is > signalled immediately before re-entering the guest. But ioeventfd sits in the cache: the internal counter. The fact it's signalled does not force a barrier on the signalling thread. It looks like the same happens here: value is saved with the file descriptor, other accesses of the same device can bypass the write. > The purpose of this feature is to let the device emulation program > handle these writes asynchronously (without holding up the vCPU for a > response from the device emulation program) but the order of > reads/writes remains unchanged. > > Stefan I don't see how this can be implemented without guest changes though. For example, how do you make sure two writes to such regions are ordered just like they are on PCI? -- MST
