Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 4 Apr 2019 15:57:24 +0000
From:   Christopher Lameter <cl@linux.com>
To:     Vlastimil Babka <vbabka@suse.cz>
cc:     linux-mm@kvack.org, Pekka Enberg <penberg@kernel.org>,
        David Rientjes <rientjes@google.com>,
        Joonsoo Kim <iamjoonsoo.kim@lge.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Mel Gorman <mgorman@techsingularity.net>,
        linux-kernel@vger.kernel.org
Subject: Re: [RFC 0/2] add static key for slub_debug
In-Reply-To: <20190404091531.9815-1-vbabka@suse.cz>
Message-ID: <01000169e911ae41-0abde43e-18e8-442b-b289-e796c461f0b1-000000@email.amazonses.com>
References: <20190404091531.9815-1-vbabka@suse.cz>
User-Agent: Alpine 2.21 (DEB 202 2017-01-01)
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Feedback-ID: 1.us-east-1.fQZZZ0Xtj2+TD7V5apTT/NrT6QKuPgzCT/IC7XYgDKI=:AmazonSES
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, 4 Apr 2019, Vlastimil Babka wrote:

> I looked a bit at SLUB debugging capabilities and first thing I noticed is
> there's no static key guarding the runtime enablement as is common for similar
> debugging functionalities, so here's a RFC to add it. Can be further improved
> if there's interest.

Well the runtime enablement is per slab cache and static keys are global.

Adding static key adds code to the critical paths. Since the flags for a
kmem_cache have to be inspected anyways there may not be that much of a
benefit.

> It's true that in the alloc fast path the debugging check overhead is AFAICS
> amortized by the per-cpu cache, i.e. when the allocation is from there, no
> debugging functionality is performed. IMHO that's kinda a weakness, especially
> for SLAB_STORE_USER, so I might also look at doing something about it, and then
> the static key might be more critical for overhead reduction.

Moving debugging out of the per cpu fastpath allows that fastpath to be
much simpler and faster.

SLAB_STORE_USER is mostly used only for debugging in which case we are
less concerned with performance.

If you want to use SLAB_STORE_USER in the fastpath then we have to do some
major redesign there.

> In the freeing fast path I quickly checked the stats and it seems that in
> do_slab_free(), the "if (likely(page == c->page))" is not as likely as it
> declares, as in the majority of cases, freeing doesn't happen on the object
> that belongs to the page currently cached. So the advantage of a static key in
> slow path __slab_free() should be more useful immediately.

Right. The freeing logic is actuall a weakness in terms of performance for
SLUB due to the need to operate on a per page queue immediately.