When exporting sparse raw image, qemu-nbd reports unallocated area as zero: $ qemu-nbd --persistent --socket /tmp/nbd.sock --read-only --shared=10 \ --format raw empty-6g.raw --cache=none --aio=native $ nbdinfo --map nbd+unix:///?socket=/tmp/nbd.sock 0 6442450944 2 zero When using qcow2 image, it reports unallocated areas as a hole: $ qemu-nbd --persistent --socket /tmp/nbd.sock --read-only --shared=10 \ --format qcow2 empty-6g.qcow2 --cache=none --aio=native $ nbdinfo --map nbd+unix:///?socket=/tmp/nbd.sock 0 6442450944 3 hole,zero Since nbdcopy is ignoring the ZERO flag and using only the HOLE flag, coping raw images is extremely slow: $ hyperfine -w3 "./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null:" Benchmark #1: ./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null: Time (mean ± σ): 1.595 s ± 0.034 s [User: 2.284 s, System: 3.565 s] Range (min … max): 1.522 s … 1.630 s 10 runs This is 69 times slower than qemu-img: $ hyperfine -w3 "qemu-img convert -n nbd+unix:///?socket=/tmp/nbd.sock \ 'json:{\"file.driver\":\"null-co\",\"file.size\":\"6g\"}'" Benchmark #1: qemu-img convert -n nbd+unix:///?socket=/tmp/nbd.sock 'json:{"file.driver":"null-co","file.size":"6g"}' Time (mean ± σ): 23.1 ms ± 0.5 ms [User: 6.3 ms, System: 16.5 ms] Range (min … max): 22.6 ms … 25.5 ms 124 runs Using ZERO instead of HOLE, nbdcopy does not read zero extents from the server so it can copy this image 165 times faster: $ hyperfine -w3 "./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null:" Benchmark #1: ./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null: Time (mean ± σ): 9.8 ms ± 0.8 ms [User: 6.7 ms, System: 5.2 ms] Range (min … max): 9.2 ms … 15.4 ms 287 runs Real images show smaller speedup, only 2 times faster: $ qemu-nbd --persistent --socket /tmp/nbd.sock --read-only --shared=10 \ --format raw fedora-32.raw --cache=none --aio=native Before: $ hyperfine -w3 "./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null:" Benchmark #1: ./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null: Time (mean ± σ): 1.613 s ± 0.181 s [User: 1.843 s, System: 2.820 s] Range (min … max): 1.407 s … 1.829 s 10 runs After: $ hyperfine -w3 "./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null:" Benchmark #1: ./nbdcopy nbd+unix:///?socket=/tmp/nbd.sock null: Time (mean ± σ): 795.5 ms ± 78.7 ms [User: 198.3 ms, System: 743.1 ms] Range (min … max): 743.3 ms … 1012.0 ms 10 runs For reference, copying same image with qemu-img: $ hyperfine -w3 "qemu-img convert -n nbd+unix:///?socket=/tmp/nbd.sock \ 'json:{\"file.driver\":\"null-co\",\"file.size\":\"6g\"}'" Benchmark #1: qemu-img convert -n nbd+unix:///?socket=/tmp/nbd.sock 'json:{"file.driver":"null-co","file.size":"6g"}' Time (mean ± σ): 1.046 s ± 0.028 s [User: 122.3 ms, System: 354.5 ms] Range (min … max): 1.026 s … 1.121 s 10 runs This issue does not exist when copying from file, since in this case we detect unallocated areas as holes. Signed-off-by: Nir Soffer <nsoffer(a)redhat.com&gt; --- copy/file-ops.c | 4 ++-- copy/main.c | 2 +- copy/multi-thread-copying.c | 4 ++-- copy/nbd-ops.c | 10 ++++++++-- copy/nbdcopy.h | 2 +- copy/synch-copying.c | 2 +- 6 files changed, 15 insertions(+), 9 deletions(-) diff --git a/copy/file-ops.c b/copy/file-ops.c index f61b67e..2a239d0 100644 --- a/copy/file-ops.c +++ b/copy/file-ops.c @@ -250,7 +250,7 @@ file_get_extents (struct rw *rw, uintptr_t index, if (pos > offset) { e.offset = offset; e.length = pos - offset; - e.hole = true; + e.zero = true; if (extent_list_append (ret, e) == -1) { perror ("realloc"); exit (EXIT_FAILURE); @@ -271,7 +271,7 @@ file_get_extents (struct rw *rw, uintptr_t index, if (pos > offset) { e.offset = offset; e.length = pos - offset; - e.hole = false; + e.zero = false; if (extent_list_append (ret, e) == -1) { perror ("realloc"); exit (EXIT_FAILURE); diff --git a/copy/main.c b/copy/main.c index cfecb32..68a6030 100644 --- a/copy/main.c +++ b/copy/main.c @@ -667,7 +667,7 @@ default_get_extents (struct rw *rw, uintptr_t index, e.offset = offset; e.length = count; - e.hole = false; + e.zero = false; if (extent_list_append (ret, e) == -1) { perror ("realloc"); exit (EXIT_FAILURE); diff --git a/copy/multi-thread-copying.c b/copy/multi-thread-copying.c index 4576119..98b4056 100644 --- a/copy/multi-thread-copying.c +++ b/copy/multi-thread-copying.c @@ -157,8 +157,8 @@ worker_thread (void *indexp) char *data; size_t len; - if (exts.ptr[i].hole) { - /* The source is a hole so we can proceed directly to + if (exts.ptr[i].zero) { + /* The source is zero so we can proceed directly to * skipping, trimming or writing zeroes at the destination. */ command = calloc (1, sizeof *command); diff --git a/copy/nbd-ops.c b/copy/nbd-ops.c index f7dc37c..0bcf29b 100644 --- a/copy/nbd-ops.c +++ b/copy/nbd-ops.c @@ -190,8 +190,14 @@ add_extent (void *vp, const char *metacontext, e.offset = offset; e.length = entries[i]; - /* Note we deliberately don't care about the ZERO flag. */ - e.hole = (entries[i+1] & LIBNBD_STATE_HOLE) != 0; + + /* + * Note we deliberately don't care about the HOLE flag. There is no need to + * read extent that reads as zeroes. We will convert to it to a hole or + * allocated extents based on the command line arguments. + */ + e.zero = (entries[i+1] & LIBNBD_STATE_ZERO) != 0; + if (extent_list_append (ret, e) == -1) { perror ("realloc"); exit (EXIT_FAILURE); diff --git a/copy/nbdcopy.h b/copy/nbdcopy.h index f586fc5..69fac2a 100644 --- a/copy/nbdcopy.h +++ b/copy/nbdcopy.h @@ -100,7 +100,7 @@ struct command { struct extent { uint64_t offset; uint64_t length; - bool hole; + bool zero; }; DEFINE_VECTOR_TYPE(extent_list, struct extent); diff --git a/copy/synch-copying.c b/copy/synch-copying.c index 043893f..2712c10 100644 --- a/copy/synch-copying.c +++ b/copy/synch-copying.c @@ -68,7 +68,7 @@ synch_copying (void) for (i = 0; i < exts.size; ++i) { assert (exts.ptr[i].length <= count); - if (exts.ptr[i].hole) { + if (exts.ptr[i].zero) { if (!dst.ops->synch_trim (&dst, offset, exts.ptr[i].length) && !dst.ops->synch_zero (&dst, offset, exts.ptr[i].length)) { /* If neither trimming nor efficient zeroing are possible, -- 2.26.2