From: 'Marcelo Ricardo Leitner'
> Sent: 21 August 2020 21:47
...
> >
> > 2) Optimise the genradix lookup for the case where there
> > is a single page - it can be completely inlined.
>
> And/or our struct sizes.
>
> __idx_to_offset() will basically do:
> if (!is_power_of_2(obj_size)) {
> return (idx / objs_per_page) * PAGE_SIZE +
> (idx % objs_per_page) * obj_size;
> } else {
> return idx * obj_size;
>
> if we can get it to his the else block, it saves some CPU cycles already (at
> the expense of memory).
I'm penning some changes to the genradix code:
That expression can be written:
idx * obj_size + (idx / objs_per_page) * (PAGE_SIZE % obj_size)
which is simpler and doesn't need the is_power_of_2() test.
You can then do the early test:
if (idx * obj_size <= PAGE_SIZE - obj_size) && root && !llevel)
return root + idx * obj_size;
in the inline function.
A slight increase in code size, but a reduction in code path.
(Better still if you increment 'level'.)
Which catches most SCTP uses and the only other place I've found it used.
There is actually a nasty bug in the genradix code.
If the cmpxchg 'fails' when adding extra root levels then the page
can get reused later - but child[0] has been set.
This could add a loop to the tree!
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
