Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sun, 30 Oct 2022 14:19:22 +0100 (CET)
From:   Christoph Lameter <cl@gentwo.de>
To:     Wen Yao <haiwenyao@uniontech.com>
cc:     paul.walmsley@sifive.com, palmer@dabbelt.com,
        aou@eecs.berkeley.edu, dennis@kernel.org, tj@kernel.org,
        linux-riscv@lists.infradead.org, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org
Subject: Re: [PATCH 1/2] riscv: percpu:Add riscv percpu operations
In-Reply-To: <20221026104015.565468-2-haiwenyao@uniontech.com>
Message-ID: <f1b82a9-a88b-50f5-8b49-69e52d7bdaa9@gentwo.de>
References: <20221026104015.565468-1-haiwenyao@uniontech.com> <20221026104015.565468-2-haiwenyao@uniontech.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Precedence: bulk

On Wed, 26 Oct 2022, Wen Yao wrote:

> This patch use riscv AMO(Atomic Memory Operation) instructions to
> optimise some this_cpu_and this_cpu_or this_cpu_add operations.
> It reuse cmpxchg_local() to impletment this_cpu_cmpxchg macros.
> It reuse xchg_relaxed() to impletment this_cpu_xchg macros.

Are you sure that these changes gives you any benefit vs disabling preempt
or irq offs? I dont know too much about atomics on riscv but it looks like
you are using full atomics. The performance penalty for the use of those
is usually drastic. Often irq/preempt off is better.

Could you run some of the synthetic tests to establish the benefit? F.e.
run the synthetic tests for the slub allcator with and without these
patches.


> +			__asm__ __volatile__(                                  \
> +				"amo" #asm_op ".w"                             \

amo = atomic operation?