From: Pádraig Brady <P@draigBrady.com>
Date: Fri, 30 Dec 2022 19:34:27 +0000 (+0000)
Subject: copy: attempt copy offload with sparse files by default
X-Git-Tag: v9.2~100
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=879d2180d6b58e7a83312681fbce9a1e841c2ae4;p=thirdparty%2Fcoreutils.git

copy: attempt copy offload with sparse files by default

This was seen to vastly improve performance
on NFS 4.2 systems by allowing server side copies,
with partially sparse files (avidemux generated mp4 files).

* src/copy.c (lseek_copy): Also set hole_size to 0,
i.e. enable copy_file_range(), with --sparse=auto (the default),
to enable copy offload in this case, as we've strong signal
from SEEK_DATA that we're operating on actual data and not holes here.
* NEWS: Mention the improvement.
Fixes https://bugs.gnu.org/60416
---

diff --git a/NEWS b/NEWS
index 7fb8c5363f..0beaef5066 100644
--- a/NEWS
+++ b/NEWS
@@ -69,6 +69,12 @@ GNU coreutils NEWS                                    -*- outline -*-
 
 ** Improvements
 
+  cp --sparse=auto (the default), mv, and install,
+  will use the copy_file_range syscall now also with sparse files.
+  This may be more efficient, by avoiding user space copies,
+  and possibly employing copy offloading or reflinking,
+  for the non sparse portion of such sparse files.
+
   date --debug now diagnoses if multiple --date or --set options are
   specified, as only the last specified is significant in that case.
 
diff --git a/src/copy.c b/src/copy.c
index e465271efa..7407517f16 100644
--- a/src/copy.c
+++ b/src/copy.c
@@ -526,12 +526,12 @@ lseek_copy (int src_fd, int dest_fd, char **abuf, size_t buf_size,
       last_ext_len = ext_len;
 
       /* Copy this extent, looking for further opportunities to not
-         bother to write zeros unless --sparse=never, since SEEK_HOLE
+         bother to write zeros if --sparse=always, since SEEK_HOLE
          is conservative and may miss some holes.  */
       off_t n_read;
       bool read_hole;
       if ( ! sparse_copy (src_fd, dest_fd, abuf, buf_size,
-                          sparse_mode == SPARSE_NEVER ? 0 : hole_size,
+                          sparse_mode != SPARSE_ALWAYS ? 0 : hole_size,
                           true, allow_reflink, src_name, dst_name,
                           ext_len, &n_read, &read_hole))
         return false;