Re: fsync() errors is unsafe and risks data loss

[Jeff Layton <jlayton@redhat.com>]

On Thu, 2018-04-12 at 14:31 -0700, Matthew Wilcox wrote:
> On Thu, Apr 12, 2018 at 05:14:54PM -0400, Jeff Layton wrote:
> > On Thu, 2018-04-12 at 13:28 -0700, Matthew Wilcox wrote:
> > > On Thu, Apr 12, 2018 at 01:13:22PM -0700, Andres Freund wrote:
> > > > I think a per-file or even per-blockdev/fs error state that'd be
> > > > returned by fsync() would be more than sufficient.
> > > 
> > > Ah; this was my suggestion to Jeff on IRC.  That we add a per-
> > > superblock
> > > wb_err and then allow syncfs() to return it.  So you'd open an fd on
> > > a directory (for example), and call syncfs() which would return -EIO
> > > or -ENOSPC if either of those conditions had occurred since you
> > > opened
> > > the fd.
> > 
> > Not a bad idea and shouldn't be too costly. mapping_set_error could
> > flag the superblock one before or after the one in the mapping.
> > 
> > We'd need to define what happens if you interleave fsync and syncfs
> > calls on the same inode though. How do we handle file->f_wb_err in that
> > case? Would we need a second field in struct file to act as the per-sb
> > error cursor?
> 
> Ooh.  I hadn't thought that through.  Bleh.  I don't want to add a field
> to struct file for this uncommon case.
> 
> Maybe O_PATH could be used for this?  It gets you a file descriptor on
> a particular filesystem, so syncfs() is defined, but it can't report
> a writeback error.  So if you open something O_PATH, you can use the
> file's f_wb_err for the mapping's error cursor.
> 

That might work.

It'd be a syscall behavioral change so we'd need to document that well.
It's probably innocuous though -- I doubt we have a lot of callers in
the field opening files with O_PATH and calling syncfs on them.
-- 
Jeff Layton <jlayton@redhat.com>