On Thu, Jun 02, 2022 at 10:31:21AM +1000, Dave Chinner wrote:
> From: Dave Chinner <dchinner@xxxxxxxxxx>
>
> LVM/DM has conniptions when you try to use snapshots on a device
> that has DAX capability. It first sets up the underlying device as a
> DAX capable mapping (type 3 or DM_TYPE_DAX_BIO_BASED) but because
> snapshots require COW and shared mappings, it isn't supported on DAX
> capable devices. Hence creating the snapshot device fails because it
> requires a type 1 (DM_TYPE_BIO_BASED) device and DM can't change
> types on a loaded mapping.
>
> Hence we get this obscure error message in the log:
>
> device-mapper: ioctl: can't change device type (old=3 vs new=1) after initial table load.
>
> and these obscure, unhelpful error messages from the LVM command
> outputs:
>
> device-mapper: reload ioctl on (251:0) failed: Invalid argument
> Failed to suspend logical volume vg_081/base_081.
> Device vg_081-base_081-real (251:1) is used by another device.
> Failed to revert logical volume vg_081/base_081.
> Aborting. Manual intervention required.
> Failed to create snapshot
>
> How to turn off DAX capability is not documented in dmsetup or LVM
> man pages, nor is dax mentioned anywhere in
> Documentation/admin/device-mapper/ so I have no idea how to tell
> LVM/DM "don't try to enable DAX support!".
>
> As such, if the uderlying block device is dax capable, skip this
> test.
>
> Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
Seems reasonable to me, who also sees weird problems on the dax vms...
Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
--D
> ---
> common/rc | 39 +++++++++++++++++++++++++++------------
> 1 file changed, 27 insertions(+), 12 deletions(-)
>
> diff --git a/common/rc b/common/rc
> index 2f31ca46..c334cfbf 100644
> --- a/common/rc
> +++ b/common/rc
> @@ -2116,21 +2116,33 @@ _require_sane_bdev_flush()
> # Decide if the scratch filesystem is likely to be mounted in fsdax mode.
> # It goes 3 ways based on mount options::
> # 1. "dax" or "dax=always" means always test using DAX
> -# 2. "dax=never" means we'll never use DAX
> +# 2. "dax=never" means we'll never use DAX.
> # 3. "dax=inode" or nothing means "use scratch dev capability" to
> # determine whether DAX is going to be used.
> #
> -# Returns 0 if DAX will be used, 1 if DAX is not going to be used.
> +# Case 2 and 3 basically mean the same thing for the purpose of
> +# _require_dm_target(). If the fs is not forcing the use of DAX, then DAX
> +# can only be enabled if the underlying block device supports it.
> +#
> +# Returns 0 if the filesytem will use DAX, 1 if it won't.
> __scratch_uses_fsdax()
> {
> local ops=$(_normalize_mount_options "$MOUNT_OPTIONS")
>
> echo $ops | egrep -qw "dax(=always| |$)" && return 0
> - echo $ops | grep -qw "dax=never" && return 1
> + return 1
> +}
>
> +# Determine if the scratch device is DAX capable. Even if the fs is not
> +# using DAX, we still can't use certain device mapper targets if the block
> +# device is DAX capable. Hence the check needs to be separat from the FS
> +# capability.
> +__scratch_dev_has_dax()
> +{
> local sysfs="/sys/block/$(_short_dev $SCRATCH_DEV)"
> test -e "${sysfs}/dax" && return 0
> test "$(cat "${sysfs}/queue/dax" 2>/dev/null)" = "1" && return 0
> +
> return 1
> }
>
> @@ -2147,15 +2159,18 @@ _require_dm_target()
> _require_sane_bdev_flush $SCRATCH_DEV
> _require_command "$DMSETUP_PROG" dmsetup
>
> - if __scratch_uses_fsdax; then
> - case $target in
> - stripe|linear|log-writes)
> - ;;
> - *)
> - _notrun "Cannot run tests with DAX on $target devices."
> - ;;
> - esac
> - fi
> + case $target in
> + stripe|linear|log-writes)
> + ;;
> + *)
> + if __scratch_uses_fsdax; then
> + _notrun "Cannot run tests with fsdax on $target devices."
> + fi
> + if __scratch_dev_has_dax; then
> + _notrun "Cannot use $target devices on DAX capable block devices."
> + fi
> + ;;
> + esac
>
> modprobe dm-$target >/dev/null 2>&1
>
> --
> 2.35.1
>
