Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [RFC v2 00/34] SLUB: reduce irq disabled scope and make it RT
 compatible
To:     Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc:     linux-mm@kvack.org, linux-kernel@vger.kernel.org,
        Christoph Lameter <cl@linux.com>,
        David Rientjes <rientjes@google.com>,
        Pekka Enberg <penberg@kernel.org>,
        Joonsoo Kim <iamjoonsoo.kim@lge.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Mel Gorman <mgorman@techsingularity.net>,
        Jesper Dangaard Brouer <brouer@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Jann Horn <jannh@google.com>
References: <20210609113903.1421-1-vbabka@suse.cz>
 <20210702182944.lqa7o2a25to6czju@linutronix.de>
 <35b26e48-a96a-41b0-826e-43e43660c9d6@suse.cz>
 <20210729134939.iulryxjarhjmpugz@linutronix.de>
From:   Vlastimil Babka <vbabka@suse.cz>
Message-ID: <627eddeb-3dc0-056e-ae07-f14c4b1a1b8e@suse.cz>
Date:   Thu, 29 Jul 2021 16:17:26 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.12.0
MIME-Version: 1.0
In-Reply-To: <20210729134939.iulryxjarhjmpugz@linutronix.de>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 7/29/21 3:49 PM, Sebastian Andrzej Siewior wrote:

> now that I'm slowly catching up…

> 

> On 2021-07-02 22:25:05 [+0200], Vlastimil Babka wrote:

>> > - perf_5.10 stat -r 10 hackbench -g200 -s 4096 -l500

>> > Old:

>> > |         464.967,20 msec task-clock                #   27,220 CPUs utilized            ( +-  0,16% )

>> > New:

>> > |         422.865,71 msec task-clock                #    4,782 CPUs utilized            ( +-  0,34% )

>> 

>> The series shouldn't significantly change the memory allocator

>> interaction, though.

>> Seems there's less cycles, but more time elapsed, thus more sleeping -

>> is it locks becoming mutexes on RT?

> 

> yes, most likely since the !RT parts are mostly unchanged.

> 

>> My second guess - list_lock remains spinlock with my series, thus RT

>> mutex, but the current RT tree converts it to raw_spinlock. I'd hope

>> leaving that one as non-raw spinlock would still be much better for RT

>> goals, even if hackbench (which is AFAIK very slab intensive) throughput

>> regresses - hopefully not that much.

> 

> Yes, the list_lock seems to be the case. I picked your

> slub-local-lock-v3r0 and changed the list_lock (+slab_lock()) to use

> raw_spinlock_t and disable interrupts and CPUs utilisation went to

> ~23CPUs (plus a bunch of warnings which probably made it a little slower

> again).


I forgot to point that out in the cover letter, but with v3 this change to
raw_spinlock_t is AFAICS no longer possible (at least with
CONFIG_SLUB_CPU_PARTIAL) because in put_cpu_partial() we now take the local_lock
and it can be called from get_partial_node() which takes the list_lock.


> The difference between a sleeping lock (spinlock_t) and a mutex is

> that we attempt not to preempt a task that acquired a spinlock_t even if

> it is running for some time and the scheduler would preempt it (like it

> would do if the task had a mutex acquired. These are the "lazy preempt"

> bits in the RT patch).

> 

> By making the list_lock a raw_spinlock_t a lot of IRQ-flags dancing

> needs to be done as the page-allocator must be entered with enabled

> interrupts.


Hm but SLUB should never call the page allocator from under list_lock in my series?


> And then there is the possibility that you may need to free

> some memory even if you allocate memory which requires some extra steps

> on RT due to the IRQ-off part. All this vanishes by keeping list_lock a

> spinlock_t.

> The kernel-build test on /dev/shm remained unchanged so that is good.

> Unless there is a real-world use-case, that gets worse, I don't mind

> keeping the spinlock_t here. I haven't seen tglx complaining so far.


Good. IIRC hackbench is very close to being a slab microbenchmark, so
regressions there are expected, but should not translate to notable real world
regressions.


> Sebastian

>