Re: [PATCH v2] perf ordered_events: Optimise event object reuse

[Matt Fleming]

On Wed, 27 May, at 12:25:33PM, Jiri Olsa wrote:
> On Tue, May 26, 2020 at 02:59:28PM +0100, Matt Fleming wrote:
> > +/*
> > + * Allocate a new event object from the free event cache.
> > + *
> > + * Find the first address range in the cache and carve out enough bytes
> > + * for an ordered_event objects. The object with the lowest address is
> > + * always returned so that subsequent allocations benefit from
> > + * contiguous memory accesses (spatial locality).
> > + */
> > +static struct ordered_event *free_event_get_tree(struct ordered_events *oe)
> > +{
> > +	struct interval_tree_node *it;
> > +	struct ordered_event *new;
> > +	size_t bytes = sizeof(*new);
> > +
> > +	it = interval_tree_iter_first(&oe->cache.rb, 0, ULONG_MAX);
> > +	if (!it)
> > +		return NULL;
> > +
> > +	/* Has the cache memory been exhausted? */
> > +	assert(cache_region_size(it) >= bytes);
> > +
> > +	new = (void *)it->start;
> > +
> > +	if (cache_region_size(it) == bytes) {
> > +		interval_tree_remove(it, &oe->cache.rb);
> > +		free(it);
> > +	}
> > +
> > +	it->start += bytes;
> 
> this does not look right.. should this go to else path in above condition?

Urgh, yes. You're right -- this should be in the else path.