From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-1.web.codeaurora.org [10.30.226.201]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id 4893928EB for ; Mon, 30 Jan 2023 14:03:03 +0000 (UTC) Received: by smtp.kernel.org (Postfix) with ESMTPSA id A0D12C433D2; Mon, 30 Jan 2023 14:03:02 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=linuxfoundation.org; s=korg; t=1675087383; bh=QZ78tu706ZlnOvpAZIFjADwDCdpKi+VTJmtcQvgZalk=; h=From:To:Cc:Subject:Date:In-Reply-To:References:From; b=azH70JCcc9KdTz/iurXGeAE0vBI7mS0Zr/NoF+fl0biXYoCMUtZw30SFAHN/CtpJV 4R85zZgoUk15csoiSG8XydHcANc6KEZqEOLxJll0Rmu52qTCeP3C9sOAMYMIc+IWjn X0R8a0BInxu1wW25pJYZkRFdOVfX+8HEccEYzN6Q= From: Greg Kroah-Hartman To: stable@vger.kernel.org Cc: Greg Kroah-Hartman , patches@lists.linux.dev, Mateusz Guzik , Peter Zijlstra , Linus Torvalds , Sasha Levin , Tony Luck , Nicholas Piggin , Will Deacon Subject: [PATCH 6.1 189/313] lockref: stop doing cpu_relax in the cmpxchg loop Date: Mon, 30 Jan 2023 14:50:24 +0100 Message-Id: <20230130134345.502806140@linuxfoundation.org> X-Mailer: git-send-email 2.39.1 In-Reply-To: <20230130134336.532886729@linuxfoundation.org> References: <20230130134336.532886729@linuxfoundation.org> User-Agent: quilt/0.67 Precedence: bulk X-Mailing-List: patches@lists.linux.dev List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit From: Mateusz Guzik [ Upstream commit f5fe24ef17b5fbe6db49534163e77499fb10ae8c ] On the x86-64 architecture even a failing cmpxchg grants exclusive access to the cacheline, making it preferable to retry the failed op immediately instead of stalling with the pause instruction. To illustrate the impact, below are benchmark results obtained by running various will-it-scale tests on top of the 6.2-rc3 kernel and Cascade Lake (2 sockets * 24 cores * 2 threads) CPU. All results in ops/s. Note there is some variance in re-runs, but the code is consistently faster when contention is present. open3 ("Same file open/close"): proc stock no-pause 1 805603 814942 (+%1) 2 1054980 1054781 (-0%) 8 1544802 1822858 (+18%) 24 1191064 2199665 (+84%) 48 851582 1469860 (+72%) 96 609481 1427170 (+134%) fstat2 ("Same file fstat"): proc stock no-pause 1 3013872 3047636 (+1%) 2 4284687 4400421 (+2%) 8 3257721 5530156 (+69%) 24 2239819 5466127 (+144%) 48 1701072 5256609 (+209%) 96 1269157 6649326 (+423%) Additionally, a kernel with a private patch to help access() scalability: access2 ("Same file access"): proc stock patched patched +nopause 24 2378041 2005501 5370335 (-15% / +125%) That is, fixing the problems in access itself *reduces* scalability after the cacheline ping-pong only happens in lockref with the pause instruction. Note that fstat and access benchmarks are not currently integrated into will-it-scale, but interested parties can find them in pull requests to said project. Code at hand has a rather tortured history. First modification showed up in commit d472d9d98b46 ("lockref: Relax in cmpxchg loop"), written with Itanium in mind. Later it got patched up to use an arch-dependent macro to stop doing it on s390 where it caused a significant regression. Said macro had undergone revisions and was ultimately eliminated later, going back to cpu_relax. While I intended to only remove cpu_relax for x86-64, I got the following comment from Linus: I would actually prefer just removing it entirely and see if somebody else hollers. You have the numbers to prove it hurts on real hardware, and I don't think we have any numbers to the contrary. So I think it's better to trust the numbers and remove it as a failure, than say "let's just remove it on x86-64 and leave everybody else with the potentially broken code" Additionally, Will Deacon (maintainer of the arm64 port, one of the architectures previously benchmarked): So, from the arm64 side of the fence, I'm perfectly happy just removing the cpu_relax() calls from lockref. As such, come back full circle in history and whack it altogether. Signed-off-by: Mateusz Guzik Link: https://lore.kernel.org/all/CAGudoHHx0Nqg6DE70zAVA75eV-HXfWyhVMWZ-aSeOofkA_=WdA@mail.gmail.com/ Acked-by: Tony Luck # ia64 Acked-by: Nicholas Piggin # powerpc Acked-by: Will Deacon # arm64 Acked-by: Peter Zijlstra Signed-off-by: Linus Torvalds Signed-off-by: Sasha Levin --- lib/lockref.c | 1 - 1 file changed, 1 deletion(-) diff --git a/lib/lockref.c b/lib/lockref.c index 45e93ece8ba0..2afe4c5d8919 100644 --- a/lib/lockref.c +++ b/lib/lockref.c @@ -23,7 +23,6 @@ } \ if (!--retry) \ break; \ - cpu_relax(); \ } \ } while (0) -- 2.39.0