Re: [PATCH 3/4] cputlb: introduce tlb_flush_other_cpu for reset use

[Richard Henderson]

On 2/26/25 06:29, Alex Bennée wrote:
> I guess we want something like:
>
>
> /* tlb_reset() - reset the TLB when the CPU is not running
>   * cs: the cpu
>   *
>   * Only to be used when the CPU is definitely not running
>   */
>
> void tlb_reset(CPUState *cs) {
>       g_assert(cs->cpu_stopped);
>
>      for (i = 0; i < NB_MMU_MODES; i++) {
>          tlb_mmu_flush_locked(&cpu->neg.tlb.d[i], &cpu->neg.tlb.f[i]);
>      }
> }
>
> ?

I like the name, and the separate assert.
I'm not convinced skipping the tlb resize and (especially) accounting is a good 
idea.

I suspect that the tlb_flush_by_mmuidx_async_work should be split vs its 
assert_cpu_is_self, and you just should use that.  I'll note that tcg_cpu_reset_hold and 
tlb_flush_by_mmuidx_async_work both call tcg_flush_jmp_cache, so we've got a double-flush 
in there.

If you don't want to use tlb_flush_by_mmuidx_async_work, I think you need at 
minimum

- take the lock
- tlb_window_reset()
- honor and update cpu->neg.tlb.c.dirty


r~