Re: [PATCH 1/8] perf/hw_breakpoint: Optimize list of per-task breakpoints

[Marco Elver]

On Thu, Jun 09, 2022 at 02:30PM +0200, Dmitry Vyukov wrote:
[...]
> > +       rcu_read_lock();
> 
> Why do we need rcu_read_lock() here?
> The patch does not change anything with respect to locking, so all
> accesses to the container should still be protected by nr_bp_mutex.
> Similarly for the rcu variant of for_each below.
[...]
> > +       head = rhltable_lookup(&task_bps_ht, &bp->hw.target, task_bps_ht_params);
> > +       if (!head)
> > +               goto out;
> > +
> > +       rhl_for_each_entry_rcu(iter, pos, head, hw.bp_list) {

It's part of rhashtable's interface requirements:

	/**
	 * rhltable_lookup - search hash list table
	 * @hlt:	hash table
	 * @key:	the pointer to the key
	 * @params:	hash table parameters
	 *
	 * Computes the hash value for the key and traverses the bucket chain looking
	 * for a entry with an identical key.  All matching entries are returned
	 * in a list.
	 *
	 * This must only be called under the RCU read lock.
	 *
	 * Returns the list of entries that match the given key.
	 */

Beyond that, even though there might not appear to be any concurrent
rhashtable modifications, it'll be allowed in patch 6/8. Furthermore,
rhashtable actually does concurrent background compactions since I
selected 'automatic_shrinking = true' (so we don't leak tons of memory
after starting and killing those 1000s of tasks) -- there's this
call_rcu() in lib/rhashtable.c that looks like that's when it's used.
This work is done in a deferred work by rht_deferred_worker().