Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753339AbdGSJJd (ORCPT ); Wed, 19 Jul 2017 05:09:33 -0400 Received: from szxga03-in.huawei.com ([45.249.212.189]:9374 "EHLO szxga03-in.huawei.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752582AbdGSJJ1 (ORCPT ); Wed, 19 Jul 2017 05:09:27 -0400 Subject: Re: [PATCH 0/6] Cache coherent device memory (CDM) with HMM v5 To: Jerome Glisse References: <20170713211532.970-1-jglisse@redhat.com> <2d534afc-28c5-4c81-c452-7e4c013ab4d0@huawei.com> <20170718153816.GA3135@redhat.com> <20170719022537.GA6911@redhat.com> CC: , , John Hubbard , David Nellans , Dan Williams , Balbir Singh , "Michal Hocko" From: Bob Liu Message-ID: Date: Wed, 19 Jul 2017 17:09:04 +0800 User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:45.0) Gecko/20100101 Thunderbird/45.8.0 MIME-Version: 1.0 In-Reply-To: <20170719022537.GA6911@redhat.com> Content-Type: text/plain; charset="windows-1252" Content-Transfer-Encoding: 8bit X-Originating-IP: [10.142.83.150] X-CFilter-Loop: Reflected X-Mirapoint-Virus-RAPID-Raw: score=unknown(0), refid=str=0001.0A020203.596F21C4.0126,ss=1,re=0.000,recu=0.000,reip=0.000,cl=1,cld=1,fgs=0, ip=0.0.0.0, so=2014-11-16 11:51:01, dmn=2013-03-21 17:37:32 X-Mirapoint-Loop-Id: 23c4a0a9227bec3849d7b96e7b05592b Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4126 Lines: 87 On 2017/7/19 10:25, Jerome Glisse wrote: > On Wed, Jul 19, 2017 at 09:46:10AM +0800, Bob Liu wrote: >> On 2017/7/18 23:38, Jerome Glisse wrote: >>> On Tue, Jul 18, 2017 at 11:26:51AM +0800, Bob Liu wrote: >>>> On 2017/7/14 5:15, J?r?me Glisse wrote: >>>>> Sorry i made horrible mistake on names in v4, i completly miss- >>>>> understood the suggestion. So here i repost with proper naming. >>>>> This is the only change since v3. Again sorry about the noise >>>>> with v4. >>>>> >>>>> Changes since v4: >>>>> - s/DEVICE_HOST/DEVICE_PUBLIC >>>>> >>>>> Git tree: >>>>> https://cgit.freedesktop.org/~glisse/linux/log/?h=hmm-cdm-v5 >>>>> >>>>> >>>>> Cache coherent device memory apply to architecture with system bus >>>>> like CAPI or CCIX. Device connected to such system bus can expose >>>>> their memory to the system and allow cache coherent access to it >>>>> from the CPU. >>>>> >>>>> Even if for all intent and purposes device memory behave like regular >>>>> memory, we still want to manage it in isolation from regular memory. >>>>> Several reasons for that, first and foremost this memory is less >>>>> reliable than regular memory if the device hangs because of invalid >>>>> commands we can loose access to device memory. Second CPU access to >>>>> this memory is expected to be slower than to regular memory. Third >>>>> having random memory into device means that some of the bus bandwith >>>>> wouldn't be available to the device but would be use by CPU access. >>>>> >>>>> This is why we want to manage such memory in isolation from regular >>>>> memory. Kernel should not try to use this memory even as last resort >>>>> when running out of memory, at least for now. >>>>> >>>> >>>> I think set a very large node distance for "Cache Coherent Device Memory" >>>> may be a easier way to address these concerns. >>> >>> Such approach was discuss at length in the past see links below. Outcome >>> of discussion: >>> - CPU less node are bad >>> - device memory can be unreliable (device hang) no way for application >>> to understand that >> >> Device memory can also be more reliable if using high quality and expensive memory. > > Even ECC memory does not compensate for device hang. When your GPU lockups > you might need to re-init GPU from scratch after which the content of the > device memory is unreliable. During init the device memory might not get > proper clock or proper refresh cycle and thus is susceptible to corruption. > >> >>> - application and driver NUMA madvise/mbind/mempolicy ... can conflict >>> with each other and no way the kernel can figure out which should >>> apply >>> - NUMA as it is now would not work as we need further isolation that >>> what a large node distance would provide >>> >> >> Agree, that's where we need spend time on. >> >> One drawback of HMM-CDM I'm worry about is one more extra copy. >> In the cache coherent case, CPU can write data to device memory >> directly then start fpga/GPU/other accelerators. > > There is not necessarily an extra copy. Device driver can pre-allocate > virtual address range of a process with device memory. Device page fault Okay, I get your point. But the typical use case is CPU allocate a memory and prepare/write data then launch GPU "cuda kernel". How to control the allocation go to device memory e.g HBM or system DDR at the beginning without user explicit advise? If goes to DDR by default, there is an extra copy. If goes to HBM by default, the HBM may be waste. > can directly allocate device memory. Once allocated CPU access will use > the device memory. > Then it's more like replace the numa node solution(CDM) with ZONE_DEVICE(type MEMORY_DEVICE_PUBLIC). But the problem is the same, e.g how to make sure the device memory say HBM won't be occupied by normal CPU allocation. Things will be more complex if there are multi GPU connected by nvlink(also cache coherent) in a system, each GPU has their own HBM. How to decide allocate physical memory from local HBM/DDR or remote HBM/DDR? If using numa(CDM) approach there are NUMA mempolicy and autonuma mechanism at least. Thanks, Bob