[PATCH 0/2] file-posix: Add 'accept-read-only-device' option

[PATCH 0/2] file-posix: Add 'accept-read-only-device' option

[Kevin Wolf]

Kevin Wolf (2):
  file-posix: Add 'accept-read-only-device' option
  iotests: Add loop-accept-ro-dev test

 qapi/block-core.json                          | 19 +++-
 block/file-posix.c                            | 23 ++++-
 tests/qemu-iotests/tests/loop-accept-ro-dev   | 87 +++++++++++++++++++
 .../qemu-iotests/tests/loop-accept-ro-dev.out | 19 ++++
 4 files changed, 144 insertions(+), 4 deletions(-)
 create mode 100755 tests/qemu-iotests/tests/loop-accept-ro-dev
 create mode 100644 tests/qemu-iotests/tests/loop-accept-ro-dev.out

-- 
2.56.0

[PATCH 1/2] file-posix: Add 'accept-read-only-device' option

[Kevin Wolf]

The host_device block driver refuses to open a block device read-write
if the BLKROGET ioctl tells it that it's actually read-only. For such
devices it's possible to open a read-write file descriptor, but any
writes to the device would fail.

There are valid use cases for which this check gets in the way, as
explained in the bug report:

    However, this behavior prevents a legitimate enterprise use case:
    running VMs against storage replication target LUNs in a disaster
    recovery configuration. Enterprise storage arrays (Dell Powermax,
    Pure Storage) set the SCSI Write Protect bit on replication target
    LUNs, causing the Linux kernel to mark them read-only via BLKROGET.
    The VMs on the passive DR site need to start with these LUNs
    attached - writes are expected to fail (the application middleware
    knows it's on the passive side), and after a DR role swap the LUN
    becomes writable and writes should succeed immediately without any
    VM restart.

In order to allow such use cases in which a currently read-only device
is deliberately opened read-write, add a new option to the host_device
driver that allows the user to skip the check.

Buglink: https://redhat.atlassian.net/browse/RHEL-186136
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 qapi/block-core.json | 19 ++++++++++++++++++-
 block/file-posix.c   | 23 ++++++++++++++++++++---
 2 files changed, 38 insertions(+), 4 deletions(-)

diff --git a/qapi/block-core.json b/qapi/block-core.json
index 741c8165c06..80272b878c5 100644
--- a/qapi/block-core.json
+++ b/qapi/block-core.json
@@ -3390,6 +3390,23 @@
   'features': [ { 'name': 'dynamic-auto-read-only',
                   'if': 'CONFIG_POSIX' } ] }
 
+##
+# @BlockdevOptionsHostDevice:
+#     Driver specific block device options for the host_device
+#     backend.
+#
+# @accept-read-only-device: allow opening a read-only device as
+#     read-write.  If this is false, opening a host device read-write
+#     requires on Linux that the BLKROGET ioctl considers the device
+#     writable.  If it is true, this check is skipped.  (since: 11.2,
+#     default: false)
+#
+# Since: 11.2
+##
+{ 'struct': 'BlockdevOptionsHostDevice',
+  'base': 'BlockdevOptionsFile',
+  'data': { '*accept-read-only-device': 'bool' } }
+
 ##
 # @BlockdevOptionsNull:
 #     Driver specific block device options for the null backend.
@@ -4777,7 +4794,7 @@
       'ftps':       'BlockdevOptionsCurlFtps',
       'host_cdrom':  { 'type': 'BlockdevOptionsFile',
                        'if': 'HAVE_HOST_BLOCK_DEVICE' },
-      'host_device': { 'type': 'BlockdevOptionsFile',
+      'host_device': { 'type': 'BlockdevOptionsHostDevice',
                        'if': 'HAVE_HOST_BLOCK_DEVICE' },
       'http':       'BlockdevOptionsCurlHttp',
       'https':      'BlockdevOptionsCurlHttps',
diff --git a/block/file-posix.c b/block/file-posix.c
index bad7ccdb733..3318912b5ac 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -177,6 +177,7 @@ typedef struct BDRVRawState {
     bool has_laio_fdsync:1;
     bool use_linux_io_uring:1;
     bool use_mpath:1;
+    bool accept_read_only_device:1;
     int page_cache_inconsistent; /* errno from fdatasync failure */
     bool has_fallocate;
     bool needs_alignment;
@@ -513,7 +514,7 @@ static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
     }
 }
 
-static int check_hdev_writable(int fd)
+static int check_hdev_writable(BlockDriverState *bs, int fd)
 {
 #if defined(BLKROGET)
     /* Linux block devices can be configured "read-only" using blockdev(8).
@@ -523,10 +524,19 @@ static int check_hdev_writable(int fd)
      * bdrv_open() is supposed to fail if the disk is read-only.  Explicitly
      * check for read-only block devices so that Linux block devices behave
      * properly.
+     *
+     * Users can use the accept-read-only-device option to skip the check if
+     * they deliberately open a read-only block device as writable, e.g. because
+     * it may become read-write later.
      */
+    BDRVRawState *s = bs->opaque;
     struct stat st;
     int readonly = 0;
 
+    if (s->accept_read_only_device) {
+        return 0;
+    }
+
     if (fstat(fd, &st)) {
         return -errno;
     }
@@ -620,6 +630,11 @@ static QemuOptsList raw_runtime_opts = {
             .type = QEMU_OPT_BOOL,
             .help = "check that page cache was dropped on live migration (default: off)"
         },
+        {
+            .name = "accept-read-only-device",
+            .type = QEMU_OPT_BOOL,
+            .help = "allow opening a read-only device as read-write (default: off)"
+        },
         { /* end of list */ }
     },
 };
@@ -737,8 +752,10 @@ static int raw_open_common(BlockDriverState *bs, QDict *options,
     s->fd = fd;
 
     /* Check s->open_flags rather than bdrv_flags due to auto-read-only */
+    s->accept_read_only_device =
+        qemu_opt_get_bool(opts, "accept-read-only-device", false);
     if (s->open_flags & O_RDWR) {
-        ret = check_hdev_writable(s->fd);
+        ret = check_hdev_writable(bs, s->fd);
         if (ret < 0) {
             error_setg_errno(errp, -ret, "The device is not writable");
             goto fail;
@@ -1116,7 +1133,7 @@ static int raw_reconfigure_getfd(BlockDriverState *bs, int flags,
     }
 
     if (fd != -1 && (*open_flags & O_RDWR)) {
-        ret = check_hdev_writable(fd);
+        ret = check_hdev_writable(bs, fd);
         if (ret < 0) {
             qemu_close(fd);
             error_setg_errno(errp, -ret, "The device is not writable");
-- 
2.56.0

Re: [PATCH 1/2] file-posix: Add 'accept-read-only-device' option

[Eric Blake]

On Wed, Oct 28, 2026 at 06:10:56PM +0100, Kevin Wolf wrote:
> The host_device block driver refuses to open a block device read-write
> if the BLKROGET ioctl tells it that it's actually read-only. For such
> devices it's possible to open a read-write file descriptor, but any
> writes to the device would fail.
> 
> There are valid use cases for which this check gets in the way, as
> explained in the bug report:
> 
>     However, this behavior prevents a legitimate enterprise use case:
>     running VMs against storage replication target LUNs in a disaster
>     recovery configuration. Enterprise storage arrays (Dell Powermax,
>     Pure Storage) set the SCSI Write Protect bit on replication target
>     LUNs, causing the Linux kernel to mark them read-only via BLKROGET.
>     The VMs on the passive DR site need to start with these LUNs
>     attached - writes are expected to fail (the application middleware
>     knows it's on the passive side), and after a DR role swap the LUN
>     becomes writable and writes should succeed immediately without any
>     VM restart.
> 
> In order to allow such use cases in which a currently read-only device
> is deliberately opened read-write, add a new option to the host_device
> driver that allows the user to skip the check.
> 
> Buglink: https://redhat.atlassian.net/browse/RHEL-186136
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>  qapi/block-core.json | 19 ++++++++++++++++++-
>  block/file-posix.c   | 23 ++++++++++++++++++++---
>  2 files changed, 38 insertions(+), 4 deletions(-)

Reviewed-by: Eric Blake <eblake@redhat.com>

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.
Virtualization:  qemu.org | libguestfs.org

[PATCH 2/2] iotests: Add loop-accept-ro-dev test

[Kevin Wolf]

Add a test for the accept-read-only-device option of host_device.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 tests/qemu-iotests/tests/loop-accept-ro-dev   | 87 +++++++++++++++++++
 .../qemu-iotests/tests/loop-accept-ro-dev.out | 19 ++++
 2 files changed, 106 insertions(+)
 create mode 100755 tests/qemu-iotests/tests/loop-accept-ro-dev
 create mode 100644 tests/qemu-iotests/tests/loop-accept-ro-dev.out

diff --git a/tests/qemu-iotests/tests/loop-accept-ro-dev b/tests/qemu-iotests/tests/loop-accept-ro-dev
new file mode 100755
index 00000000000..83ca67a0cf0
--- /dev/null
+++ b/tests/qemu-iotests/tests/loop-accept-ro-dev
@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+# group: quick
+#
+# SPDX-License-Identifier: GPL-2.0-or-later
+#
+# Copyright Red Hat, Inc.
+#
+# Test the accept-read-only-device option on a loop device
+
+seq="$(basename $0)"
+echo "QA output created by $seq"
+
+status=1	# failure is the default!
+
+_cleanup() {
+    if [ -n "$loopdev" ]; then
+        sudo losetup --detach "$loopdev"
+    fi
+
+    _cleanup_test_img
+}
+
+trap "_cleanup; exit \$status" 0 1 2 3 15
+
+# get standard environment, filters and checks
+cd ..
+. ./common.rc
+. ./common.filter
+
+_supported_fmt raw
+_supported_proto file
+_supported_os Linux
+
+do_run_qemu()
+{
+    echo Testing: "$@"
+    $QEMU -nographic -qmp stdio -nodefaults "$@"
+    echo
+}
+
+run_qemu()
+{
+    echo -e '{"execute":"qmp_capabilities"}\n{"execute":"quit"}' |
+    do_run_qemu "$@" 2>&1 | _filter_testdir | _filter_qemu | _filter_qmp |
+        _filter_generated_node_ids | _filter_imgfmt |
+        _filter_actual_image_size | _filter_img_info
+}
+
+if ! sudo -n losetup &>/dev/null; then
+    _notrun "sudo losetup not available"
+fi
+
+if ! command -v blockdev >/dev/null 2>&1; then
+    _notrun "blockdev tool not available"
+fi
+
+echo
+echo "=== Start QEMU on a loop device after blockdev --setro ==="
+echo
+
+_make_test_img -f $IMGFMT 1M
+
+loopdev=$(sudo losetup --find --show "$TEST_IMG")
+if [ -z "$loopdev" ]; then
+    _fail
+fi
+
+sudo blockdev --setro "$loopdev"
+sudo chmod a+rw "$loopdev"
+
+run_qemu \
+    -blockdev host_device,node-name=disk,filename="$loopdev" \
+    -device virtio-blk,drive=disk
+
+run_qemu \
+    -blockdev host_device,node-name=disk,filename="$loopdev",accept-read-only-device=off \
+    -device virtio-blk,drive=disk
+
+run_qemu \
+    -blockdev host_device,node-name=disk,filename="$loopdev",accept-read-only-device=on \
+    -device virtio-blk,drive=disk
+
+# success, all done
+echo
+echo '*** done'
+rm -f $seq.full
+status=0
diff --git a/tests/qemu-iotests/tests/loop-accept-ro-dev.out b/tests/qemu-iotests/tests/loop-accept-ro-dev.out
new file mode 100644
index 00000000000..dbf3de9fdf6
--- /dev/null
+++ b/tests/qemu-iotests/tests/loop-accept-ro-dev.out
@@ -0,0 +1,19 @@
+QA output created by loop-accept-ro-dev
+
+=== Start QEMU on a loop device after blockdev --setro ===
+
+Formatting 'TEST_DIR/t.IMGFMT', fmt=IMGFMT size=1048576
+Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1 -device virtio-blk,drive=disk
+QEMU_PROG: -blockdev host_device,node-name=disk,filename=/dev/loop1: The device is not writable: Permission denied
+
+Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1,accept-read-only-device=off -device virtio-blk,drive=disk
+QEMU_PROG: -blockdev host_device,node-name=disk,filename=/dev/loop1,accept-read-only-device=off: The device is not writable: Permission denied
+
+Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1,accept-read-only-device=on -device virtio-blk,drive=disk
+QMP_VERSION
+{"return": {}}
+{"timestamp": {"seconds":  TIMESTAMP, "microseconds":  TIMESTAMP}, "event": "SHUTDOWN", "data": {"guest": false, "reason": "host-qmp-quit"}}
+{"return": {}}
+
+
+*** done
-- 
2.56.0

Re: [PATCH 2/2] iotests: Add loop-accept-ro-dev test

[Eric Blake]

On Wed, Oct 28, 2026 at 06:10:57PM +0100, Kevin Wolf wrote:
> Add a test for the accept-read-only-device option of host_device.
> 
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>  tests/qemu-iotests/tests/loop-accept-ro-dev   | 87 +++++++++++++++++++
>  .../qemu-iotests/tests/loop-accept-ro-dev.out | 19 ++++
>  2 files changed, 106 insertions(+)
>  create mode 100755 tests/qemu-iotests/tests/loop-accept-ro-dev
>  create mode 100644 tests/qemu-iotests/tests/loop-accept-ro-dev.out
> 

> +
> +run_qemu()
> +{
> +    echo -e '{"execute":"qmp_capabilities"}\n{"execute":"quit"}' |

'printf %s' is slightly more portable than 'echo -e'; even though this
script uses /bin/bash, you can use 'shopt -s xpg_echo' to break the
latter. On the other hand, this is copy-paste from other tests, so not
a showstopper.

> +    do_run_qemu "$@" 2>&1 | _filter_testdir | _filter_qemu | _filter_qmp |
> +        _filter_generated_node_ids | _filter_imgfmt |
> +        _filter_actual_image_size | _filter_img_info
> +}
> +
> +if ! sudo -n losetup &>/dev/null; then
> +    _notrun "sudo losetup not available"
> +fi

Huh - we seem to have several outstanding patches related to probing
whether losetup is usable:

https://lists.nongnu.org/archive/html/qemu-devel/2026-10/msg02963.html
https://lists.nongnu.org/archive/html/qemu-devel/2026-10/msg01918.html

Is this a common enough pattern in our qemu-iotests that we should
factor it out?

> +++ b/tests/qemu-iotests/tests/loop-accept-ro-dev.out
> @@ -0,0 +1,19 @@
> +QA output created by loop-accept-ro-dev
> +
> +=== Start QEMU on a loop device after blockdev --setro ===
> +
> +Formatting 'TEST_DIR/t.IMGFMT', fmt=IMGFMT size=1048576
> +Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1 -device virtio-blk,drive=disk
> +QEMU_PROG: -blockdev host_device,node-name=disk,filename=/dev/loop1: The device is not writable: Permission denied
> +
> +Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1,accept-read-only-device=off -device virtio-blk,drive=disk

/dev/loop1 might already be busy, at which point losetup will pick
/dev/loop2 or similar. This output needs to be filtered.

The rest of the test looks okay.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.
Virtualization:  qemu.org | libguestfs.org

Re: [PATCH 2/2] iotests: Add loop-accept-ro-dev test

[Kevin Wolf]

Am 28.10.2026 um 19:10 hat Eric Blake geschrieben:
> On Wed, Oct 28, 2026 at 06:10:57PM +0100, Kevin Wolf wrote:
> > Add a test for the accept-read-only-device option of host_device.
> > 
> > Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> > ---
> >  tests/qemu-iotests/tests/loop-accept-ro-dev   | 87 +++++++++++++++++++
> >  .../qemu-iotests/tests/loop-accept-ro-dev.out | 19 ++++
> >  2 files changed, 106 insertions(+)
> >  create mode 100755 tests/qemu-iotests/tests/loop-accept-ro-dev
> >  create mode 100644 tests/qemu-iotests/tests/loop-accept-ro-dev.out
> > 
> 
> > +
> > +run_qemu()
> > +{
> > +    echo -e '{"execute":"qmp_capabilities"}\n{"execute":"quit"}' |
> 
> 'printf %s' is slightly more portable than 'echo -e'; even though this
> script uses /bin/bash, you can use 'shopt -s xpg_echo' to break the
> latter. On the other hand, this is copy-paste from other tests, so not
> a showstopper.

Given that I have to fix the test anyway, I'll change it.

> > +    do_run_qemu "$@" 2>&1 | _filter_testdir | _filter_qemu | _filter_qmp |
> > +        _filter_generated_node_ids | _filter_imgfmt |
> > +        _filter_actual_image_size | _filter_img_info
> > +}
> > +
> > +if ! sudo -n losetup &>/dev/null; then
> > +    _notrun "sudo losetup not available"
> > +fi
> 
> Huh - we seem to have several outstanding patches related to probing
> whether losetup is usable:
> 
> https://lists.nongnu.org/archive/html/qemu-devel/2026-10/msg02963.html
> https://lists.nongnu.org/archive/html/qemu-devel/2026-10/msg01918.html
> 
> Is this a common enough pattern in our qemu-iotests that we should
> factor it out?

Maybe. There seem to be four tests (including this one) that use
losetup. This can be done in a separate patch, though.

> > +++ b/tests/qemu-iotests/tests/loop-accept-ro-dev.out
> > @@ -0,0 +1,19 @@
> > +QA output created by loop-accept-ro-dev
> > +
> > +=== Start QEMU on a loop device after blockdev --setro ===
> > +
> > +Formatting 'TEST_DIR/t.IMGFMT', fmt=IMGFMT size=1048576
> > +Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1 -device virtio-blk,drive=disk
> > +QEMU_PROG: -blockdev host_device,node-name=disk,filename=/dev/loop1: The device is not writable: Permission denied
> > +
> > +Testing: -blockdev host_device,node-name=disk,filename=/dev/loop1,accept-read-only-device=off -device virtio-blk,drive=disk
> 
> /dev/loop1 might already be busy, at which point losetup will pick
> /dev/loop2 or similar. This output needs to be filtered.

Oops, I'm too used to qemu-iotests filtering out all the filenames, but
obviously that doesn't work for something outside the scratch directory.
Thanks!

Kevin

Re: [PATCH 0/2] file-posix: Add 'accept-read-only-device' option

[Stefan Hajnoczi]

On Wed, Oct 28, 2026 at 06:10:55PM +0100, Kevin Wolf wrote:
> Kevin Wolf (2):
>   file-posix: Add 'accept-read-only-device' option
>   iotests: Add loop-accept-ro-dev test
> 
>  qapi/block-core.json                          | 19 +++-
>  block/file-posix.c                            | 23 ++++-
>  tests/qemu-iotests/tests/loop-accept-ro-dev   | 87 +++++++++++++++++++
>  .../qemu-iotests/tests/loop-accept-ro-dev.out | 19 ++++
>  4 files changed, 144 insertions(+), 4 deletions(-)
>  create mode 100755 tests/qemu-iotests/tests/loop-accept-ro-dev
>  create mode 100644 tests/qemu-iotests/tests/loop-accept-ro-dev.out
> 
> -- 
> 2.56.0
> 

Modulo Eric's comment about /dev/loop1 being hardcoded in the test
output:

Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>