[Libguestfs] [PATCH libnbd 3/3] copy: Do not initialize read buffer

[Nir Soffer]

nbdcopy checks pread error now, so we will never leak uninitialized data
from the heap to the destination server. Testing show 3-8% speedup when
copying a real image.

On laptop with 12 cores and 2 consumer NVMe drives:

$ qemu-nbd --read-only --persistent --shared 8 --cache none --aio native \
    --socket /tmp/src.sock --format raw fedora-35-data.raw &

$ qemu-nbd --persistent --shared 8 --cache none --aio native --discard unmap \
    --socket /tmp/dst.sock --format raw dst.raw &

$ hyperfine -p "sleep 5" "nbdcopy $SRC $DST" ".libs/nbdcopy $SRC $DST"

Benchmark 1: nbdcopy nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock
  Time (mean ± σ):      2.065 s ±  0.057 s    [User: 0.296 s, System: 1.414 s]
  Range (min … max):    2.000 s …  2.163 s    10 runs

Benchmark 2: .libs/nbdcopy nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock
  Time (mean ± σ):      1.911 s ±  0.050 s    [User: 0.059 s, System: 1.544 s]
  Range (min … max):    1.827 s …  1.980 s    10 runs

Summary
  '.libs/nbdcopy nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock' ran
    1.08 ± 0.04 times faster than 'nbdcopy nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock'

On server with 80 cores and a fast NVMe drive:

$ hyperfine "./nbdcopy-init $SRC $DST" "./nbdcopy-no-init $SRC $DST"

Benchmark 1: ./nbdcopy-init nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock
  Time (mean ± σ):      2.652 s ±  0.033 s    [User: 0.345 s, System: 1.306 s]
  Range (min … max):    2.619 s …  2.709 s    10 runs

Benchmark 2: ./nbdcopy-no-init nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock
  Time (mean ± σ):      2.572 s ±  0.031 s    [User: 0.055 s, System: 1.409 s]
  Range (min … max):    2.537 s …  2.629 s    10 runs

Summary
  './nbdcopy-no-init nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock' ran
    1.03 ± 0.02 times faster than './nbdcopy-init nbd+unix:///?socket=/tmp/src.sock nbd+unix:///?socket=/tmp/dst.sock'

Signed-off-by: Nir Soffer <nsoffer at redhat.com>
---
 copy/nbd-ops.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/copy/nbd-ops.c b/copy/nbd-ops.c
index adfe4de5..1218d6d0 100644
--- a/copy/nbd-ops.c
+++ b/copy/nbd-ops.c
@@ -52,20 +52,21 @@ static void
 open_one_nbd_handle (struct rw_nbd *rwn)
 {
   struct nbd_handle *nbd;
 
   nbd = nbd_create ();
   if (nbd == NULL) {
     fprintf (stderr, "%s: %s\n", prog, nbd_get_error ());
     exit (EXIT_FAILURE);
   }
 
+  nbd_set_pread_initialize (nbd, false);
   nbd_set_debug (nbd, verbose);
 
   /* Set the handle name for debugging.  We could use rwn->rw.name
    * here but it is usually set to the lengthy NBD URI
    * (eg. "nbd://localhost:10809") which makes debug messages very
    * long.
    */
   if (verbose) {
     char *name;
     const size_t index = rwn->handles.len;
-- 
2.35.1