From: Christoph Hellwig Subject: Re: [PATCH v5 2/8] drivers/pmem: Allow pmem_clear_poison() to accept arbitrary offset and len Date: Fri, 28 Feb 2020 06:05:08 -0800 Message-ID: <20200228140508.GA2987@infradead.org> References: <20200220215707.GC10816@redhat.com> <20200221201759.GF25974@redhat.com> <20200223230330.GE10737@dread.disaster.area> <20200224153844.GB14651@redhat.com> <20200227030248.GG10737@dread.disaster.area> <20200228013054.GO10737@dread.disaster.area> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Content-Disposition: inline In-Reply-To: Sender: linux-fsdevel-owner@vger.kernel.org To: Dan Williams Cc: Dave Chinner , Vivek Goyal , Jeff Moyer , linux-fsdevel , linux-nvdimm , Christoph Hellwig , device-mapper development List-Id: dm-devel.ids On Thu, Feb 27, 2020 at 07:28:41PM -0800, Dan Williams wrote: > "don't perform generic memory-error-handling, there's an fs that owns > this daxdev and wants to disposition errors". The fs/dax.c > infrastructure that sets up ->index and ->mapping would still need to > be there for ext4 until its ready to take on the same responsibility. > Last I checked the ext4 reverse mapping implementation was not as > capable as XFS. This goes back to why the initial design avoided > relying on not universally available / stable reverse-mapping > facilities and opted for extending the generic mm/memory-failure.c > implementation. The important but is that we stop using ->index and ->mapping when XFS is used as that enables using DAX+reflinks, which actually is the most requested DAX feature on XFS (way more than silly runtime flag switches for example).