linux-kernel - Re: [PATCH] rhashtable: add likely() to __rht

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <CADxym3ZA7FsdeA3zz34V7mHHjBC358UoJjrpV6wieZ1LF2aFxA@mail.gmail.com>
Date: Wed, 24 Sep 2025 21:30:55 +0800
From: Menglong Dong <menglong8.dong@...il.com>
To: NeilBrown <neil@...wn.name>
Cc: Herbert Xu <herbert@...dor.apana.org.au>, tgraf@...g.ch, 
	linux-crypto@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] rhashtable: add likely() to __rht_ptr()

On Tue, Sep 23, 2025 at 7:31 PM NeilBrown <neilb@...mail.net> wrote:
>
> On Tue, 23 Sep 2025, Menglong Dong wrote:
> > On Tue, Sep 23, 2025 at 2:36 PM Herbert Xu <herbert@...dor.apana.org.au> wrote:
> > >
> > > Menglong Dong <menglong8.dong@...il.com> wrote:
> > > > In the fast path, the value of "p" in __rht_ptr() should be valid.
> > > > Therefore, wrap it with a "likely". The performance increasing is tiny,
> > > > but it's still worth to do it.
> > > >
> > > > Signed-off-by: Menglong Dong <dongml2@...natelecom.cn>
> > > > ---
> > > > include/linux/rhashtable.h | 5 +++--
> > > > 1 file changed, 3 insertions(+), 2 deletions(-)
> > >
> > > It's not obvious that rht_ptr would be non-NULL.  It depends on the
> > > work load.  For example, if you're doing a lookup where most keys
> > > are non-existent then it would most likely be NULL.
> >
> > Yeah, I see. In my case, the usage of the rhashtable will be:
> > add -> lookup, and rht_ptr is alway non-NULL. You are right,
> > it can be NULL in other situations, and it's not a good idea to
> > use likely() here ;)
>
> Have you measured a performance increase?  How tiny is it?
>
> It might conceivably make sense to have a rhashtable_lookup_likely() and
> rhashtable_lookup_unlikely(), but concrete evidence of the benefit would
> be needed.

I made a more accurate bench testing:  call the rhashtable_lookup()
100000000 times.

Without the likely(), it cost  123697645ns. And with the likely(), only
84507668ns.

I add the likely() not only to the __rht_ptr(), but also rht_for_each_rcu_from()
and rhashtable_lookup().

Below is the part code of the testing:

    for (i = 0; i < num_elems; i++) {
        objs[i] = kmalloc(sizeof(**objs), GFP_KERNEL);
        KUNIT_ASSERT_NOT_ERR_OR_NULL(test, objs[i]);
        objs[i]->key = i;
        INIT_RHT_NULLS_HEAD(objs[i]->node.next);
        ret = rhashtable_insert_fast(&ht, &objs[i]->node, bench_params);
        KUNIT_ASSERT_EQ(test, ret, 0);
    }

    /* for CPU warm up */
    for (i = 0; i < 1000000000; i++) {
        u32 key = 0;
        struct bench_obj *found;

        found = rhashtable_lookup(&ht, &key, bench_params);
        KUNIT_ASSERT_NOT_ERR_OR_NULL(test, found);
        KUNIT_ASSERT_EQ(test, found->key, key);
    }

    rcu_read_lock();
    t0 = ktime_get();
    for (i = 0; i < 100000000; i++) {
        u32 key = 0;
        struct bench_obj *found;

        found = rhashtable_lookup(&ht, &key, bench_params);
        if (unlikely(!found)) {
            pr_info("error!\n");
            break;
        }
    }
    t1 = ktime_get();
    rcu_read_unlock();

>
> Thanks,
> NeilBrown