Re: [PATCH v9 3/9] plugins: Add enforcement of QEMU_PLUGIN_CB flags in register R/W callbacks

["Alex Bennée" <alex.bennee@linaro.org>]

Rowan Hart <rowanbhart@gmail.com> writes:

>  However, it should be safe at least for a subset of those callbacks and
>
>  I believe there are very valid use-cases for allowing such usage. For
>  example, we are currently working on a plugin API exposing traps. In
>  those callbacks, users may want to peek at some registers such as
>  "ecause" and "tval" (on RISC-V). We certainly will want to do so for the
>  use-case we are pushing that API for.
>
>  We could add a QEMU_PLUGIN_CB flag parameter to the respective
>  registration functions. But since they are not run from translated
>  blocks but _outside_ that context, I feel they would just clutter the
>  API without introducing any real benefit. That is, if there is no valid
>  safety or correctness concern that I'm not aware of.
>
> This makes sense, I think we could just set QEMU_PLUGIN_CB_RW_REGS for these callbacks if they're always called in a state
> where the stated requirements are met (I think they are). This would avoid breaking compatibility while maintaining the
> functionality. Same as you, I looked around and it seems like the vcpu_(init|idle|exit) locations are definitely okay, I'll check
> into the tb_trans callback, I'm not entirely sure about it.

It should be OK here too.

The CB_RW_REGS flags are only needed for callback from within the
translated code - much like their cousins the TCG helpers. All register
state should rectified if we are no longer running translated code -
indeed currently it is rectified by the time we leave the translated
block and before we start the next one.

There are some edge cases when we handle exceptions but we should have
fixed up any register state before we longjmp back to the start of the
run loop.

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro