This work aims to allow userspace to create and destroy block devices
in a race-free way, and to allow them to be exposed to other Xen VMs via
blkback without races.
Changes since v1:
- Several device-mapper fixes added.
- The diskseq is now a separate Xenstore node, rather than being part of
physical-device.
- Potentially backwards-incompatible changes to device-mapper now
require userspace opt-in.
- The code has been tested: I have a block script written in C that uses
these changes to successfully boot a Xen VM.
- The core block layer is almost completely untouched. Instead of
exposing a block device inode directly to userspace, device-mapper
ioctls that create a block device now return that device's diskseq.
Userspace can then use that diskseq to safely open the device.
Furthermore, ioctls that operate on an existing device-mapper device
now accept a diskseq parameter, which can be used to prevent races.
There are a few changes that make device-mapper's table validation
stricter. Two of them are clear-cut fixes for memory safety bugs: one
prevents a misaligned pointer dereference in the kernel, and the other
prevents pointer arithmetic overflow that could cause the kernel to
dereference a userspace pointer, especially on 32-bit systems. One
fixes a double-fetch bug that happens to be harmless right now, but
would make distribution backports of subsequent patches unsafe. The
remaining fixes prevent totally nonsensical tables from being accepted.
This includes parameter strings that overlap the subsequent target spec,
and target specs that overlap the 'struct dm_ioctl' or each other. I
doubt there is any userspace extant that generates such tables.
Finally, one patch forbids device-mapper devices to be named ".", "..",
or "control". Since device-mapper devices are often accessed via
/dev/mapper/NAME, such names would likely greatly confuse userspace. I
consider this to be an extension of the existing check that prohibits
device mapper names or UUIDs from containing '/'.
Demi Marie Obenour (16):
device-mapper: Check that target specs are sufficiently aligned
device-mapper: Avoid pointer arithmetic overflow
device-mapper: do not allow targets to overlap 'struct dm_ioctl'
device-mapper: Better error message for too-short target spec
device-mapper: Target parameters must not overlap next target spec
device-mapper: Avoid double-fetch of version
device-mapper: Allow userspace to opt-in to strict parameter checks
device-mapper: Allow userspace to provide expected diskseq
device-mapper: Allow userspace to suppress uevent generation
device-mapper: Refuse to create device named "control"
device-mapper: "." and ".." are not valid symlink names
device-mapper: inform caller about already-existing device
xen-blkback: Implement diskseq checks
block, loop: Increment diskseq when releasing a loop device
xen-blkback: Minor cleanups
xen-blkback: Inform userspace that device has been opened
block/genhd.c | 1 +
drivers/block/loop.c | 6 +
drivers/block/xen-blkback/blkback.c | 8 +-
drivers/block/xen-blkback/xenbus.c | 147 ++++++++--
drivers/md/dm-core.h | 2 +
drivers/md/dm-ioctl.c | 432 ++++++++++++++++++++++------
drivers/md/dm.c | 5 +-
include/linux/device-mapper.h | 2 +-
include/uapi/linux/dm-ioctl.h | 67 ++++-
9 files changed, 551 insertions(+), 119 deletions(-)
--
Sincerely,
Demi Marie Obenour (she/her/hers)
Invisible Things Lab