From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Message-ID: <552BB4CA.6060808@plexistor.com> Date: Mon, 13 Apr 2015 15:21:30 +0300 From: Boaz Harrosh MIME-Version: 1.0 Subject: Re: [GIT PULL] PMEM driver for v4.1 References: <20150413093309.GA30219@gmail.com> <20150413093541.GA5147@lst.de> <20150413104531.GB30556@gmail.com> In-Reply-To: <20150413104531.GB30556@gmail.com> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org To: Ingo Molnar , Christoph Hellwig Cc: Linus Torvalds , linux-kernel@vger.kernel.org, linux-nvdimm@lists.01.org, Ross Zwisler , Dan Williams , Matthew Wilcox List-ID: On 04/13/2015 01:45 PM, Ingo Molnar wrote: > > * Christoph Hellwig wrote: > >> On Mon, Apr 13, 2015 at 11:33:09AM +0200, Ingo Molnar wrote: >>> Limitations: this is a regular block device, and since the pmem areas >>> are not struct page backed, they are invisible to the rest of the >>> system (other than the block IO device), so direct IO to/from pmem >>> areas, direct mmap() or XIP is not possible yet. The page cache will >>> also shadow and double buffer pmem contents, etc. >> >> Unless you use the DAX support in ext2/4 and soon XFS, in which case >> we avoid that double buffering when doing read/write and mmap > > Indeed, I missed that DAX support just went upstream in v4.0 - nice! > > DAX may have some other limitations though that comes from not having > struct page * backing and using VM_MIXEDMAP, the following APIs might > not work on DAX files: > > - splice splice works fine. Also I sent a cleanup in this area to Andrew it will be in for 4.1 > - zero copy O_DIRECT into DAX areas. DAX is always O_DIRECT. What does not work is mmap of DAX file and use that pointer in an O_DIRECT operation of another device. (unless it is a DAX device) Also mmap of DAX file and RDMA or direct-networking. Will need a copy. All this is fixable by applying my page-struct patch for pmem > - futexes > > - ( AFAICS hugetlbs won't work on DAX mmap()s yet - although with > the current nocache mapping that's probable the least of the > performance issues for now. ) > > Btw., what's the future design plan here? Enable struct page backing, > or provide special codepaths for all DAX uses like the special pte > based approach for mmap()s? > I'm hopping for struct page, 4k pages at first and 2M pages later on, which needs more work in IO stacks, where I need this most. > Thanks, > Ingo > Thanks Boaz