On Mon, 2007-05-07 at 16:31 -0700, Andrew Morton wrote:
> On Mon, 7 May 2007 19:14:42 -0400
> Theodore Tso <tytso@xxxxxxx> wrote:
> > On Mon, May 07, 2007 at 03:38:56PM -0700, Andrew Morton wrote:
> > > > Actually, this is a non-issue. The reason that it is handled for
> > > > extent-only
> > > > is that this is the only way to allocate space in the filesystem without
> > > > doing the explicit zeroing. For other filesystems (including ext3 and
> > > > ext4 with block-mapped files) the filesystem should return an error
> > > > (e.g.
> > > > -EOPNOTSUPP) and glibc will do manual zero-filling of the file in
> > > > userspace.
> > >
> > > It can be a bit suboptimal from the layout POV. The reservations code
> > > will
> > > largely save us here, but kernel support might make it a bit better.
> > Actually, the reservations code won't matter, since glibc will fall
> > back to its current behavior, which is it will do the preallocation by
> > explicitly writing zeros to the file.
> No! Reservations code is *critical* here. Without reservations, we get
> disastrously-bad layout if two processes were running a large fallocate()
> at the same time. (This is an SMP-only problem, btw: on UP the timeslice
> lengths save us).
> My point is that even though reservations save us, we could do even-better
In this case, since the number of blocks to preallocate (eg. N=10GB) is
clear, we could improve the current reservation code, to allow callers
explicitly ask for a new window that have the minimum N free blocks for
the blocks-to-preallocated(rather than just have at least 1 free
Before the ext4_fallocate() is called, the right reservation window size
is set with the flag to indicating "please spend time if needed to find
a window covers at least N free blocks".
So for ex4 block mapped files, later when glibc is doing allocation and
zeroing, the ext4 block-mapped allocator will knows to reserve the right
amount of free blocks before allocating and zeroing 10GB space.
I am not sure whether this worth the effort though.
> But then, a smart application would bypass the glibc() fallocate()
> implementation and would tune the reservation window size and would use
> direct-IO or sync_file_range()+fadvise(FADV_DONTNEED).
> > This wlil result in the same
> > layout as if we had done the persistent preallocation, but of course
> > it will mean the posix_fallocate() could potentially take a long time
> > if you're a PVR and you're reserving a gig or two for a two hour movie
> > at high quality. That seems suboptimal, granted, and ideally the
> > application should be warned about this before it calls
> > posix_fallocate(). On the other hand, it's what happens today, all
> > the time, so applications won't be too badly surprised.
> A PVR implementor would take all this over and would do it themselves, for
> > If we think applications programmers badly need to know in advance if
> > posix_fallocate() will be fast or slow, probably the right thing is to
> > define a new fpathconf() configuration option so they can query to see
> > whether a particular file will support a fast posix_fallocate(). I'm
> > not 100% convinced such complexity is really needed, but I'm willing
> > to be convinced.... what do folks think?
> An application could do sys_fallocate(one-byte) to work out whether it's
> supported in-kernel, I guess.