From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jesper Dangaard Brouer Date: Sat, 08 Dec 2018 16:06:48 +0000 Subject: Re: [RFC] avoid indirect calls for DMA direct mappings v2 Message-Id: <20181208170648.296fb8a1@redhat.com> List-Id: References: <20181207190720.18517-1-hch@lst.de> In-Reply-To: <20181207190720.18517-1-hch@lst.de> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Christoph Hellwig Cc: iommu@lists.linux-foundation.org, Linus Torvalds , Tariq Toukan , Ilias Apalodimas , Toke =?UTF-8?B?SMO4aWxhbmQtSsO4?= =?UTF-8?B?cmdlbnNlbg==?= , Robin Murphy , Konrad Rzeszutek Wilk , Tony Luck , Fenghua Yu , Marek Szyprowski , Keith Busch , Jonathan Derrick , linux-pci@vger.kernel.org, linux-ia64@vger.kernel.org, x86@kernel.org, linux-kernel@vger.kernel.org, brouer@redhat.com On Fri, 7 Dec 2018 11:07:05 -0800 Christoph Hellwig wrote: > Hi all, > > a while ago Jesper reported major performance regressions due to the > spectre v2 mitigations in his XDP forwarding workloads. A large part > of that is due to the DMA mapping API indirect calls. > > It turns out that the most common implementation of the DMA API is the > direct mapping case, and now that we have merged almost all duplicate > implementations of that into a single generic one is easily feasily to > direct calls for this fast path. > > This series adds consolidate the DMA mapping code by merging the > swiotlb case into the dma direct case, and then treats NULL dma_ops > as an indicator that that we should directly call the direct mapping > case. This recovers a large part of the retpoline induces XDP slowdown. > > This works is based on the dma-mapping tree, so you probably want to > want this git tree for testing: > > git://git.infradead.org/users/hch/misc.git dma-direct-calls.2 > > Gitweb: > > http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/dma-direct-calls.2 You can add my: Tested-by: Jesper Dangaard Brouer or Acked-by: Jesper Dangaard Brouer I'm very happy that you work on this. And I've done micro-benchmark testing of the patchset (and branch dma-direct-calls), which I've made avail here: https://github.com/xdp-project/xdp-project/blob/master/areas/dma/dma01_test_hellwig_direct_dma.org My XDP performance is back, minus the BPF-indirect call, and net_rx_action napi->poll, and net_device->ndo_xdp_xmit calls. I verified that manually disabling retpoline for these remaining netstack retpoline-calls restore the performance full (well minus 1.5 nanosec). -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer