On Wed, Mar 08, 2017 at 08:49:56PM +0000, Trond Myklebust wrote: > On Wed, 2017-03-08 at 15:32 -0500, bfields@xxxxxxxxxxxx wrote: > > On Wed, Mar 08, 2017 at 08:18:31PM +0000, Trond Myklebust wrote: > > > On Wed, 2017-03-08 at 15:00 -0500, Olga Kornievskaia wrote: > > > > > On Mar 8, 2017, at 2:53 PM, J. Bruce Fields <bfields@fieldses.o > > > > > rg> > > > > > wrote: > > > > > > > > > > On Wed, Mar 08, 2017 at 12:32:12PM -0500, Olga Kornievskaia > > > > > wrote: > > > > > > > > > > > > > On Mar 8, 2017, at 12:25 PM, Christoph Hellwig <hch@infrade > > > > > > > ad.o > > > > > > > rg> > > > > > > > wrote: > > > > > > > > > > > > > > On Wed, Mar 08, 2017 at 12:05:21PM -0500, J. Bruce Fields > > > > > > > wrote: > > > > > > > > Since copy isn't atomic that check is never going to be > > > > > > > > reliable. > > > > > > > > > > > > > > That's true for everything that COPY does. By that logic > > > > > > > we > > > > > > > should > > > > > > > not implement it at all (a logic that I'd fully support) > > > > > > > > > > > > If you were to only keep CLONE then you’d lose a huge > > > > > > performance > > > > > > gain > > > > > > you get from server-to-server COPY. > > > > > > > > > > Yes. Also, I think copy-like copy implementations have > > > > > reasonable > > > > > semantics that are basically the same as read: > > > > > > > > > > - copy can return successfully with less copied than > > > > > requested. > > > > > - it's fine for the copied range to start and/or end > > > > > past end > > > > > of > > > > > file, it'll just return a short read. > > > > > - A copy of more than 0 bytes returning 0 means you're > > > > > at end > > > > > of > > > > > file. > > > > > > > > > > The particular problem here is that that doesn't fit how clone > > > > > works at > > > > > all. > > > > > > > > > > It feels like what happened is that copy_file_range() was made > > > > > mainly > > > > > for the clone case, with the idea that copy might be > > > > > reluctantly > > > > > accepted as a second-class implementation. > > > > > > Historically? No... Christoph added clone as a valid implementation > > > of > > > copy_file_range() almost a year after Zach and Anna defined the > > > semantics of vfs_copy_file_range(). git blame is your friend... > > > > Yeah, I know. It still feels to me like the interface was originally > > designed with clone in mind, but that's my vague impression from the > > man > > pages and half-remembered conversations. > > > > Though the lack of a "just copy the whole file regardless of size" > > case > > is weird for clone. All you can do is stat the file and then hope it > > doesn't change before you issue the copy_file_range. But I'd think > > it'd > > be easy for an atomic clone implementation to handle, say, getting a > > snapshot of a log file while it's getting continuously appended to. > > It really isn't that interesting in the continuously appended case > (what difference does it make if you only get data from just a few > moments ago), but I can see it being an issue in the case of random > writes where the file size is being extended. Bah, yes, apologies for the bad example. > The thing is that in both those cases, the copy_file_range() semantics > are worse, since they don't even guarantee a time-consistent copy. > > > > > > But the performance gain of copy offload is too big to just > > > > > ignore, > > > > > and > > > > > in fact it's what copy_file_range does on every filesystem but > > > > > btrfs and > > > > > ocfs2 (and maybe cifs?), so I don't think we can just ignore > > > > > it. > > > > > > > > > > If we had separate copy_file_range and clone_file_range, I > > > > > *think* > > > > > it > > > > > could all be made sensible. Am I missing something? > > > > > > > > > > > > > How would the application (cp) know when to call the > > > > clone_file_range > > > > and when to call copy_file_range? > > > > > > cp can probably call copy_file_range(), but any application that > > > needs > > > atomic semantics (i.e. a binary operation success/fail) must call > > > clone_file_range(). > > > > I don't believe there's a clone_file_range(). I see the vfs > > interface, > > but no system call. > > There is a standard FICLONERANGE ioctl() that can be used on all > filesystems that support the vfs interface. Oh, thanks, I forgot about that. So I don't understand why it needed to be added to copy_file_range(). The copy and clone semantics are different enough that I think callers want to know which they're getting. > > And implementing a simple cp is harder than it should be when you > > don't > > know whether it's implemented as copy or clone. You have to stat for > > the file size first, retry if you got it wrong, and also retry if you > > get a short read. The example in the clone_file_range() man page is > > incomplete. > > As I said, you shouldn't be using copy_file_range() either in the case > where the file is being modified. Don't we want it to have more or less the same behavior as a read-write loop? People are probably running backup programs that depend on just simple copies, and maybe the results are good enough for their purposes, or maybe they're actually corrupting parts of their backups and don't know, but we can't suddenly start aborting their backups with errors and tell users it's for their own good. So copy_file_range() callers will need to handle EINVAL on changing files somehow. --b.
