We cannot write to images opened with O_DIRECT unless we allow them to be resized so they are aligned to the sector size: Since 9c60a5d1978, bdrv_node_refresh_perm() ensures that for nodes whose length is not aligned to the request alignment and where someone has taken a WRITE permission, the RESIZE permission is taken, too).
Let qemu-img convert pass the BDRV_O_RESIZE flag (which causes blk_new_open() to take the RESIZE permission) when using cache=none for the target, so that when writing to it, it can be aligned to the target sector size. Without this patch, an error is returned: $ qemu-img convert -f raw -O raw -t none foo.img /mnt/tmp/foo.img qemu-img: Could not open '/mnt/tmp/foo.img': Cannot get 'write' permission without 'resize': Image size is not a multiple of request alignment Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=1994266 Signed-off-by: Hanna Reitz <hre...@redhat.com> --- As I have written in the BZ linked above, I am not sure what behavior we want. It can be argued that the current behavior is perfectly OK because we want the target to have the same size as the source, so if this cannot be done, we should just print the above error (which I think explains the problem well enough that users can figure out they need to resize the source image). OTOH, it is difficult for me to imagine a case where the user would prefer the above error to just having qemu-img align the target image's length. --- qemu-img.c | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/qemu-img.c b/qemu-img.c index 908fd0cce5..d4b29bf73e 100644 --- a/qemu-img.c +++ b/qemu-img.c @@ -2628,6 +2628,14 @@ static int img_convert(int argc, char **argv) goto out; } + if (flags & BDRV_O_NOCACHE) { + /* + * If we open the target with O_DIRECT, it may be necessary to + * extend its size to align to the physical sector size. + */ + flags |= BDRV_O_RESIZE; + } + if (skip_create) { s.target = img_open(tgt_image_opts, out_filename, out_fmt, flags, writethrough, s.quiet, false); -- 2.31.1
