diff options
Diffstat (limited to '')
| -rw-r--r-- | docs/LUKS2-locking.txt | 61 |
1 files changed, 61 insertions, 0 deletions
diff --git a/docs/LUKS2-locking.txt b/docs/LUKS2-locking.txt new file mode 100644 index 0000000..e401b61 --- /dev/null +++ b/docs/LUKS2-locking.txt @@ -0,0 +1,61 @@ +LUKS2 device locking overview +============================= + +Why +~~~ + +LUKS2 format keeps two identical copies of metadata stored consecutively +at the head of metadata device (file or bdev). The metadata +area (both copies) must be updated in a single atomic operation to avoid +header corruption during concurrent write. + +While with LUKS1 users may have clear knowledge of when a LUKS header is +being updated (written to) or when it's being read solely the need for +locking with legacy format was not so obvious as it is with the LUKSv2 format. + +With LUKS2 the boundary between read-only and read-write is blurry and what +used to be the exclusively read-only operation (i.e., cryptsetup open command) may +easily become read-update operation silently without user's knowledge. +Major feature of LUKS2 format is resilience against accidental +corruption of metadata (i.e., partial header overwrite by parted or cfdisk +while creating partition on mistaken block device). +Such header corruption is detected early on header read and auto-recovery +procedure takes place (the corrupted header with checksum mismatch is being +replaced by the secondary one if that one is intact). +On current Linux systems header load operation may be triggered without user +direct intervention for example by udev rule or from systemd service. +Such clash of header read and auto-recovery procedure could have severe +consequences with the worst case of having LUKS2 device unaccessible or being +broken beyond repair. + +The whole locking of LUKSv2 device headers split into two categories depending +what backend the header is stored on: + +I) block device +~~~~~~~~~~~~~~~ + +We perform flock() on file descriptors of files stored in a private +directory (by default /run/lock/cryptsetup). The file name is derived +from major:minor couple of affected block device. Note we recommend +that access to private locking directory is supposed to be limited +to superuser only. For this method to work the distribution needs +to install the locking directory with appropriate access rights. + +II) regular files +~~~~~~~~~~~~~~~~~ + +First notable difference between headers stored in a file +vs. headers stored in a block device is that headers in a file may be +manipulated by the regular user unlike headers on block devices. Therefore +we perform flock() protection on file with the luks2 header directly. + +Limitations +~~~~~~~~~~~ + +a) In general, the locking model provides serialization of I/Os targeting +the header only. It means the header is always written or read at once +while locking is enabled. +We do not suppress any other negative effect that two or more concurrent +writers of the same header may cause. + +b) The locking is not cluster aware in any way. |