From mboxrd@z Thu Jan 1 00:00:00 1970 From: Christoph Hellwig Date: Thu, 23 Oct 2014 01:51:43 -0700 Subject: [Cluster-devel] [PATCH][try6] VFS: new want_holesize and got_holesize buffer_head flags for fiemap In-Reply-To: <90509064.9314451.1413980933011.JavaMail.zimbra@redhat.com> References: <635673293.9099204.1413939501068.JavaMail.zimbra@redhat.com> <579956475.9102148.1413940195344.JavaMail.zimbra@redhat.com> <20141022060434.GA11237@infradead.org> <90509064.9314451.1413980933011.JavaMail.zimbra@redhat.com> Message-ID: <20141023085143.GC24965@infradead.org> List-Id: To: cluster-devel.redhat.com MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit On Wed, Oct 22, 2014 at 08:28:53AM -0400, Bob Peterson wrote: > Yes, I thought about that. > One of my early prototypes had a separate function used by fiemap. > Function __generic_block_fiemap would call get_block() which > returned an indication of a hole as it does today. When it saw > the hole, fiemap called a new function get_hole_size() that was > passed in like get_block. The problem is: it's grossly inefficient, > since the new function get_hole_size() has to redo most of the work > that get_block just did (at least in the case of GFS2). (Which in the > case of a 1PB sparse file is non-trivial, since it involves several > levels of metadata indirection). Combining it with get_block made it > much more efficient. > > Making a separate get_block_map_fiemap() function just seems like an > exercise in redundancy. I was thinking of replacing get_blocks entirely. We're not actually using a buffer_head in fiemap, so the interface seems somewhat awkward. If it used something like the iomap interface proposed by Dave long time ago we'd have a much saner interface that for example XFS could use as well.