From mboxrd@z Thu Jan 1 00:00:00 1970 From: Shamir Rabinovitch Subject: Re: [PATCH v1 2/2] dma-mapping-common: add DMA attribute - DMA_ATTR_IOMMU_BYPASS Date: Sat, 7 Nov 2015 17:06:42 +0200 Message-ID: <20151107150642.GA4405@shamir-ThinkPad-T430> References: <20151102144427.GA2876@suse.de> <20151102173218.GC12484@shamir-ThinkPad-T430> <20151105134206.GD2255@suse.de> <20151105.161121.188431695871972310.davem@davemloft.net> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from aserp1040.oracle.com ([141.146.126.69]:27173 "EHLO aserp1040.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750856AbbKGPHn (ORCPT ); Sat, 7 Nov 2015 10:07:43 -0500 Content-Disposition: inline In-Reply-To: <20151105.161121.188431695871972310.davem@davemloft.net> Sender: linux-arch-owner@vger.kernel.org List-ID: To: David Miller Cc: jroedel@suse.de, dwmw2@infradead.org, benh@kernel.crashing.org, arnd@arndb.de, corbet@lwn.net, linux-doc@vger.kernel.org, linux-arch@vger.kernel.org, luto@kernel.org, borntraeger@de.ibm.com, cornelia.huck@de.ibm.com, sebott@linux.vnet.ibm.com, pbonzini@redhat.com, hch@lst.de, kvm@vger.kernel.org, schwidefsky@de.ibm.com, linux-s390@vger.kernel.org On Thu, Nov 05, 2015 at 04:11:21PM -0500, David Miller wrote: > > And for the record Sowmini fixed a lot of the lock contention: > > commit ff7d37a502022149655c18035b99a53391be0383 > Author: Sowmini Varadhan > Date: Thu Apr 9 15:33:30 2015 -0400 > > Break up monolithic iommu table/lock into finer graularity pools and lock > The poor rds-stress results w/o IOMMU bypass I sent in early post were taken from kernel that has the above patch and that has all the needed changes in arch/sparc to use this new feature. It seems that it worked well for 10G ETH IOMMU lock contention but it still not solving the rds-stress issue. The difference can be from: 1. Lock contention still left with this enhancement <-- zero in bypass 2. Overhead to setup the IOMMU mapping <-- almost zero in bypass (require 1 HV call) 3. Overhead to use the IOMMU mapping <-- not sure how to measure this 4. Overhead to tear the IOMMU mapping <-- zero in bypass