On Wed, Feb 26, 2014 at 03:08:58PM -0800, Hugh Dickins wrote:
> On Wed, 26 Feb 2014, Dave Chinner wrote:
> > On Tue, Feb 25, 2014 at 08:45:15PM -0800, Hugh Dickins wrote:
> > > On Wed, 26 Feb 2014, Dave Chinner wrote:
> > > > On Tue, Feb 25, 2014 at 03:23:35PM -0800, Hugh Dickins wrote:
> > > >
> > > > > I should mention that when "we" implemented this thirty years ago,
> > > > > we had a strong conviction that the system call should be idempotent:
> > > > > that is, the len argument should indicate the final i_size, not the
> > > > > amount being removed from it. Now, I don't remember the grounds for
> > > > > that conviction: maybe it was just an idealistic preference for how
> > > > > to design a good system call. I can certainly see that defining it
> > > > > that way round would surprise many app programmers. Just mentioning
> > > > > this in case anyone on these lists sees a practical advantage to
> > > > > doing it that way instead.
> > > >
> > > > I don't see how specifying the end file size as an improvement. What
> > > > happens if you are collapse a range in a file that is still being
> > > > appended to by the application and so you race with a file size
> > > > update? IOWs, with such an API the range to be collapsed is
> > > > completely unpredictable, and IMO that's a fundamentally broken API.
> > >
> > > That's fine if you don't see the idempotent API as an improvement,
> > > I just wanted to put it on the table in case someone does see an
> > > advantage to it. But I think I'm missing something in your race
> > > example: I don't see a difference between the two APIs there.
> > Userspace can't sample the inode size via stat(2) and then use the value
> > for a
> > syscall atomically. i.e. if you specify the offset you want to
> > collapse at, and the file size you want to have to define the region
> > to collapse, then the length you need to collapse is (current inode
> > size - end file size). If "current inode size" can change between
> > the stat(2) and fallocate() call (and it can), then the length being
> > collapsed is indeterminate....
> Thanks for explaining more, I was just about to acknowledge what a good
> example that is. Indeed, it seems not unreasonable to be editing the
> earlier part of a file while the later part of it is still streaming in.
> But damn, it now occurs to me that there's still a problem at the
> streaming end: its file write offset won't be updated to reflect
> the collapse, so there would be a sparse hole at that end. And
> collapse returns -EPERM if IS_APPEND(inode).
Well, we figure that most applications won't be using append only
inode flags for files that they know they want to edit at random
offsets later on. ;)
However, I can see how DVR apps would use open(O_APPEND) to obtain
the fd they write to because that sets the write position to the EOF
on every write() call (i.e. in generic_write_checks()). And collapse
range should behave sanely with this sort of usage.
e.g. XFS calls generic_write_checks() after it has taken the IO lock
to set the current write position to EOF. Hence it will be correctly
serialised against collapse range calls and so O_APPEND writes will
not leave sparse holes if collapse range calls are interleaved with
the write stream....