On Mon, Aug 13, 2018 at 8:52 PM Eric Blake <eblake@redhat.com> wrote:
On 08/03/2018 02:28 PM, Nir Soffer wrote:
File systems not supporting FALLOC_FL_ZERO_RANGE yet fall back to manual zeroing.
We can avoid this by combining two fallocate calls:
fallocate(FALLOC_FL_PUNCH_HOLE) fallocate(0)
Based on my tests this is much more efficient compared to manual zeroing. The idea came from this qemu patch: https://github.com/qemu/qemu/commit/1cdc3239f1bb
Note: the image is sparse, but nbdkit creates a fully allocated image. This may be a bug in nbdkit or qemu-img.
Calling fallocate(0) forces allocation; so anything explicitly written to 0 won't be sparse when this mode is used.
Sure, this is a poor mans replacement for ZERO_RANGE. I assume that we want sparseness only when may_trim is set, which is the current behavior before this patch. With qemu-img the default is sparse, and using -o preallocation=falloc make the image preallocated. It seems that we don't have a way to select sparseness of the destination image in NBD, or maybe I'm missing something?
There's also a question of whether your source file accurately reports holes to begin with (poor tmpfs SEEK_HOLE performance is still a common problem). But I don't see that as getting in the way of this patch going in.
Nir