Re: [PATCH 22/24] accel/tcg: Use interval tree for user-only page tracking

[Richard Henderson]

On 10/28/22 01:59, Alex Bennée wrote:
> > I'm unwilling to put an expensive test like a function call
> > (have_mmap_lock) before an inexpensive test like pointer != NULL.
>
> Is it really that more expensive?

Well, yes.  I mean, the function call isn't really slow, but it isn't single-cycle like a 
comparison against 0.

> Sure, I guess I'm just trying to avoid having so many returns out of
> the code at various levels of nesting. The page_get_target_data code is
> harder to follow. What about:
>
> int page_get_flags(target_ulong address)
> {
>      PageFlagsNode *p = pageflags_find(address, address);
>
>      /*
>       * See util/interval-tree.c re lockless lookups: no false positives but
>       * there are false negatives.  If we had the lock and found
>       * nothing we are done, otherwise retry with the mmap lock acquired.
>       */
>      if (p) {
>          return p->flags;
>      } else if (have_mmap_lock()) {
>          return 0;
>      }
>
>      mmap_lock();
>      p = pageflags_find(address, address);
>      mmap_unlock();
>
>      return p ? p->flags : 0;
> }

Ok, I can use this.  In for v3.

As for page_get_target_data, see v2, in which there has been clean up.


r~