On Mon, Oct 05, 2015 at 01:50:47PM +0100, Peter Maydell wrote:
On 5 October 2015 at 13:40, Gabriel L. Somlo <somlo@cmu.edu> wrote:
In addition, Michael's comment earlier in the thread suggests that even my current acpi version isn't sufficiently "orthodox" w.r.t. ACPI, and I should be providing the hardware access routine as an ACPI/AML routine, to avoid race conditions with the rest of ACPI, and for encapsulation. I.e. it's even rude to use the fw_cfg node's ACPI _CRS method (the part where I'd be treating it like a DT stand-in only to query fw_cfg's hardware specifics).
If you want to try to support "firmware might also be reading fw_cfg at the same time as the kernel" this is a (painful) problem regardless of how the kernel figures out whether a fw_cfg device is present. I had assumed that one of the design assumptions of this series was that firmware would only read the fw_cfg before booting the guest kernel and never touch it afterwards. If it might touch it later then letting the guest kernel also mess with fw_cfg seems like a really bad idea.
I don't know of any case where firmware and kernel might race each other to access fw_cfg. The issue AFAICT is whether it's safe (future-proof) to rely on parsing _CRS for the fw_cfg i/o access information, or whether such logic could be rendered obsolete by potential future updates to fw_cfg's _CRS. If I "outsource" the fw_cfg_dump_blob_by_key() functionality entirely to an ACPI method, my kernel driver won't have to worry about keeping up with said future updates. On the down-side, that means the kernel driver will be ACPI or nothing (but I'm OK with that, at my curent level of understanding :) Thanks, --Gabriel