On 20 November 2017 at 14:27, Richard Henderson <r...@twiddle.net> wrote:
> On 11/20/2017 02:27 PM, Peter Maydell wrote:
>> - return (void *)((uintptr_t)addr + tlbe->addend);
>> + hostaddr = (void *)((uintptr_t)addr + tlbe->addend);
>> +
>> + memory_notdirty_write_prepare(ndi, ENV_GET_CPU(env), addr,
>> + qemu_ram_addr_from_host_nofail(hostaddr),
>> + 1 << s_bits);
>
> These calls should be conditional on TLB_NOTDIRTY being set.
> We certainly don't need to do anything like taking a lock otherwise.
We only take the lock if the dirty flag is set for the
physical memory, which it won't be if TLB_NOTDIRTY wasn't set...
> Perhaps an extra bool in NDI, like
>
> ndi->active = false;
> if (unlikely(tlb_addr & TLB_NOTDIRTY)) {
> ndi->active = true;
> memory_notdirty_write_prepare(ndi, ...);
> }
>
> and
>
> #define ATOMIC_MMU_CLEANUP \
> do { \
> if (unlikely(ndi->active)) { \
> memory_notdirty_write_complete(ndi); \
> } \
> } while (0)
...but yes, we could skip the whole thing if the TLB entry
isn't marked notdirty.
(Having NDI be "mostly opaque apart from this one thing" is a
little ugly, though.)
thanks
-- PMM
