1
0
mirror of https://github.com/systemd/systemd.git synced 2024-11-02 02:21:44 +03:00
systemd/docs/BLOCK_DEVICE_LOCKING.md
Michael Prokop d238709c14 docs: fix typos and duplicate words
s/and and/and/
s/explicity/explicitly/
s/that that/that/
s/the the/the/
s/is is/it is/
s/overriden/overridden/
2019-06-27 10:43:21 +02:00

68 lines
3.7 KiB
Markdown

---
title: Locking Block Device Access
---
# Locking Block Device Access
*TL;DR: Use BSD file locks
[(`flock(2)`)](http://man7.org/linux/man-pages/man2/flock.2.html) on block
device nodes to synchronize access for partitioning and file system formatting
tools.*
`systemd-udevd` probes all block devices showing up for file system superblock
and partition table information (utilizing `libblkid`). If another program
concurrently modifies a superblock or partition table this probing might be
affected, which is bad in itself, but also might in turn result in undesired
effects in programs subscribing to `udev` events.
Applications manipulating a block device can temporarily stop `systemd-udevd`
from processing rules on it — and thus bar it from probing the device — by
taking a BSD file lock on the block device node. Specifically, whenever
`systemd-udevd` starts processing a block device it takes a `LOCK_SH|LOCK_NB`
lock using [`flock(2)`](http://man7.org/linux/man-pages/man2/flock.2.html) on
the main block device (i.e. never on any partition block device, but on the
device the partition belongs to). If this lock cannot be taken (i.e. `flock()`
returns `EBUSY`), it refrains from processing the device. If it manages to take
the lock it is kept for the entire time the device is processed.
Note that `systemd-udevd` also watches all block device nodes it manages for
`inotify()` `IN_CLOSE` events: whenever such an event is seen, this is used as
trigger to re-run the rule-set for the device.
These two concepts allow tools such as disk partitioners or file system
formatting tools to safely and easily take exclusive ownership of a block
device while operating: before starting work on the block device, they should
take an `LOCK_EX` lock on it. This has two effects: first of all, in case
`systemd-udevd` is still processing the device the tool will wait for it to
finish. Second, after the lock is taken, it can be sure that
`systemd-udevd` will refrain from processing the block device, and thus all
other client applications subscribed to it won't get device notifications from
potentially half-written data either. After the operation is complete the
partitioner/formatter can simply close the device node. This has two effects:
it implicitly releases the lock, so that `systemd-udevd` can process events on
the device node again. Secondly, it results an `IN_CLOSE` event, which causes
`systemd-udevd` to immediately re-process the device — seeing all changes the
tool made — and notify subscribed clients about it.
Besides synchronizing block device access between `systemd-udevd` and such
tools this scheme may also be used to synchronize access between those tools
themselves. However, do note that `flock()` locks are advisory only. This means
if one tool honours this scheme and another tool does not, they will of course
not be synchronized properly, and might interfere with each other's work.
Note that the file locks follow the usual access semantics of BSD locks: since
`systemd-udevd` never writes to such block devices it only takes a `LOCK_SH`
*shared* lock. A program intending to make changes to the block device should
take a `LOCK_EX` *exclusive* lock instead. For further details, see the
`flock(2)` man page.
And please keep in mind: BSD file locks (`flock()`) and POSIX file locks
(`lockf()`, `F_SETLK`, …) are different concepts, and in their effect
orthogonal. The scheme discussed above uses the former and not the latter,
because these types of locks more closely match the required semantics.
Summarizing: it is recommended to take `LOCK_EX` BSD file locks when
manipulating block devices in all tools that change file system block devices
(`mkfs`, `fsck`, …) or partition tables (`fdisk`, `parted`, …), right after
opening the node.