On Fri, Jul 31, 2020 at 07:22:16PM +0200, Markus Wichmann wrote:
> On Fri, Jul 31, 2020 at 08:17:02AM +0800, Zhao Zhengyu wrote:
> > Hello,
> >
> > When chunks are merged, we use "(curr_size + pre_size) ^ pre_size >
> > pre_size" to decide whether to reclaim. I think this may be something
> > related to performance, but I can’t prove it. I want to know the
> > reason.
> >
> > Thank you!
> >
> > Zhengyu
>
> I asked that same question a while ago. For one, this was in the old
> malloc code, which is now usually no longer used. For two, this tries to
> figure out if adding current size to previous size rolls over into a new
> power of two. Usually, curr_size will be small and pre_size will be
> large. Therefore, adding the two will not change much about the high
> bits of pre_size, so due to the XOR operator, those bits will cancel out
> and the result will be smaller than pre_size. However, if the sum does
> roll over, then the sum has one bit set that isn't in pre_size and is
> larger than all the ones in pre_size. So the XOR can't cancel that bit,
> and the result of the XOR is greater than pre_size.
>
> Fundamentally, it is an optimized version of (a_clz(curr_size +
> pre_size) < a_clz(pre_size)).
Yes, it's a heuristic at least approximately equivalent to "crossed
the next power of two size boundary" to limit the frequency of madvise
syscalls when a large free zone keeps getting expanded by adjacent
tiny frees. However it does not work very well in practice, and
doesn't even mitigate the possibility of continuous syscall load when
a repeated malloc/free cycle occurs right at a power-of-two boundary.
mallocng handles this kind of thing much better by grouping same-sized
allocations and returning them as a group when all are freed, only
holding back from doing so if it's observed allocations of this size
"bouncing" (repeatedly creating and destroying the same group).
Rich