Re: Plugin Memory Callback Debugging

[Richard Henderson <richard.henderson@linaro.org>]

On 11/21/22 13:51, Alex Bennée wrote:
> 
> Aaron Lindsay <aaron@os.amperecomputing.com> writes:
> 
>> On Nov 15 22:36, Alex Bennée wrote:
>>> Aaron Lindsay <aaron@os.amperecomputing.com> writes:
>>>> I believe the code *should* always reset `cpu->plugin_mem_cbs` to NULL at the
>>>> end of an instruction/TB's execution, so its not exactly clear to me how this
>>>> is occurring. However, I suspect it may be relevant that we are calling
>>>> `free_dyn_cb_arr()` because my plugin called `qemu_plugin_reset()`.
>>>
>>> Hmm I'm going to have to remind myself about how this bit works.
>>
>> When is it expected that cpu->plugin_mem_cbs is reset to NULL if it is
>> set for an instruction? Is it guaranteed it is reset by the end of the
>> tb?
> 
> It should be by the end of the instruction. See
> inject_mem_disable_helper() which inserts TCG code to disable the
> helpers. We also have plugin_gen_disable_mem_helpers() which should
> catch every exit out of a block (exit_tb, goto_tb, goto_ptr). That is
> why qemu_plugin_disable_mem_helpers() is only really concerned about
> when we longjmp out of the loop.
> 
>> If I were to put an assertion in cpu_tb_exec() just after the call
>> to tcg_qemu_tb_exec(), should cpu->plugin_mem_cbs always be NULL
>> there?
> 
> Yes I think so.

Indeed.

As an aside, this field should be moved out of CPUState to DisasContextBase, because it's 
only used in code generation.

> 
>> In my debugging, I *think* I'm seeing a tb set cpu->plugin_mem_cbs
>> for an instruction, and then not reset it to NULL. I'm wondering if its
>> getting optimized away or something, but want to make sure I've got my
>> assumptions correct about how this is intended to be working.
> 
> We are expecting some stuff to dead code away (but hopefully in pairs).
> We don't know ahead of the instruction decode if it will be a memory
> instruction so opportunistically insert our empty helper calls. If no
> memory access is done those ops should be gone. This is all done to
> avoid doing a 2 pass translate.
> 
> Richard,
> 
> Have I got that right? I think thats how I remember the TCG code working.

That's about right.


r~