This fixes a bug introduced by 7be4b23649.
The function `efi_loader_get_device_part_uuid()` handles NULL
gracefully, and it is called with NULL in gpt-auto-generator.
Fixes#22862.
Previously, if a block device is locked by another process, then the
corresponding worker skip to process the corresponding event, and does
not broadcast the uevent to libudev listners. This causes several issues:
- During a period of a device being locked by a process, if a user trigger
an event with `udevadm trigger --settle`, then it never returned.
- When there is a delay between close and unlock in a process, then the
synthesized events triggered by inotify may not be processed. This can
happens easily by wrapping mkfs with flock. This causes severe issues
e.g. new devlinks are not created, or old devlinks are not removed.
This commit makes events are requeued with a tiny delay when the corresponding
block devices are locked by other processes. With this way, the triggered
uevent may be delayed but is always processed by udevd. Hence, the above
issues can be solved. Also, it is not necessary to watch a block device
unconditionally when it is already locked. Hence, the logic is dropped.
Previously, `event_run()` was called repeatedly in one `event_queue_start()`
invocation. Hence, the SELinux label database is reloaded many times needlessly.
Other settings, e.g. udev rules or hwdata, are tried to be reloaded in the
beginning of `event_queue_start()`. Let's also do so for the SELinux database.
Previously, if udevd failed to resolve event dependency, the event is
ignored and libudev listeners did not receive the event. This is
inconsistent with the case when a worker failed to process a event,
in that case, the original uevent sent by the kernel is broadcasted to
listeners.
The ENOENT, ENXIO, and ENODEV error can happen easily when a block
device appears and soon removed. So, it is reasonable to ignore the
error. But other errors should not occur here, and hence let's handle
them as critical.
When the last queued event is processed, information about subsequent
events may be already queued in the netlink socket of sd-device-monitor.
In that case, previously we once removed /run/udev/queue and touch the
file soon later, and `udevadm settle` mistakenly considered all events
are processed.
To mitigate such situation, this makes /run/udev/queue removed in on_post().
The commit 1cf4ed142d makes the IPv4 ACD
enabled unconditionally for IPv4 link-local addresses even if users
explicitly disable ACD.
This makes the IPv4 ACD is enabled by default, but honor user setting.
Fixes#22763.
This addresses the comment by Lennart
https://github.com/systemd/systemd/pull/22662#discussion_r829799863:
> /sys/subsystem is preparation for a future that never came.
> And given that the main proponent of this left Linux kernel
> development (Kay), I doubt this will ever come. So maybe we
> should start dropping references to /sys/subsystem/ given it's
> unlikely to materialize anytime soon.
and devices are sorted when the iteration started.
Previously, devices added by udev_enumerate_add_syspath() ->
device_enumerator_add_device() are not sorted. This fixes the issue.
And enable cgroup delegation for udevd.
Then, processes invoked through ExecReload= are assigned .control
subcgroup, and they are not killed by cg_kill().
Fixes#16867 and #22686.
If the prefix is only referenced by sd_radv, then the returned pointer
is already freed.
networkd does not uses the returned value. Let's voidify the function.
It seems there exists a short time period that we cannot see the
loopback device after `losetup` is finished:
```
testsuite-58.sh[367]: ++ losetup -b 1024 -P --show -f /tmp/testsuite-58-sector-1024.img
kernel: loop1: detected capacity change from 0 to 204800
testsuite-58.sh[285]: + LOOP=/dev/loop1
testsuite-58.sh[285]: + systemd-repart --pretty=yes --definitions=/tmp/testsuite-58-sector/ --seed=750b6cd5c4ae4012a15e7be3c29e6a47 --empty=require --dry-run=no /dev/loop1
testsuite-58.sh[368]: Device '/dev/loop1' has no dm-crypt/dm-verity device, no need to look for underlying block device.
testsuite-58.sh[368]: Failed to determine canonical path for '/dev/loop1': No such file or directory
testsuite-58.sh[368]: Failed to open file or determine backing device of /dev/loop1: No such file or directory
```
The Ubuntu CI on ppc64el seems to have a issue on tmpfs, and files
may not be fsynced. See c10caebb98.
For safety, let's use /var/tmp to store disk images.
This makes Request object takes hash, compare, free, and process functions.
With this change, the logic in networkd-queue.c can be mostly
independent of the type of the request or the object (e.g. Address) assigned
to the request, and it becomes simpler.
The condition should be satisfied only when users request to reconfigure
the link, and in that case, all request will be cancelled. Hence, it is
not necessary to process the request.
In most netlink handlers, we do the following,
1. decrease the message counter,
2. check the link state,
3. error handling,
4. update link state via e.g. link_check_ready().
The first two steps are mostly common, hence let's extract it.
Moreover, this is not only extracting the common logic, but provide a
strong advantage; `request_call_netlink_async()` assigns the relevant
Request object to the userdata of the netlink slot, and the request object
has full information about the message we sent. Hence, in the future,
netlink handler can print more detailed error message. E.g. when
an address is failed to configure, then currently we only show an
address is failed to configure, but with this commit, potentially we can
show which address is failed explicitly.
This does not change such error handling yet. But let's do that later.
Previously, even though all Request object are owned by Manager, they
do not have direct reference to Manager, but through Link or NetDev
object. But, as Link or NetDev can be NULL, we need to conditionalize
how to access Manager from Request with the type of the request.
This makes the way simpler, as now Request object has direct reference
to Manager.
This also rename request_drop() -> request_detach(), as in the previous
commit, the reference counter is introduced, so even if a reference of
a Request object from Manager is dropped, the object may still alive.
The naming `request_drop()` sounds the object will freed by the
function. But it may not. And `request_detach()` suggests the object
will not be managed by Manager any more, and I think it is more
appropreate.
This is just a cleanup, and should not change any behavior.
Currently, all Request object are always owned by Manager, and freed
when it is processed, especially, soon after a netlink message is sent.
So, it is not necessary to introduce the reference counter.
In a later commit, the Request object will _not_ be freed at the time
when a netlink message is sent, but assigned to the relevant netlink
slot as a userdata, and will be freed when a reply is received. So, the
owner of the Request object is changed in its lifetime. In that case, it
is convenient that the object has reference counter to avoid memleak or
double free.
This also renames e.g. request_process_address() -> address_process_request().
Also, this drops type checks such as `assert(req->type == REQUEST_TYPE_ADDRESS)`,
as in the later commits, the function of processing request, e.g.
`address_process_request()`, will be assigned to the Request object when
it is created. And the request type will be used to distinguish and to
avoid deduplicating requests which do not have any assigned objects,
like REQUEST_TYPE_DHCP4_CLIENT. Hence, the type checks in process functions
are mostly not necessary and redundant.
This is mostly cleanups and preparation for later commits, and should
not change any behavior.
This should not change any behavior, as req->netlink_handler is always
qdisc_handler or tclass_handler.
This is just a preparation for a later commit which introduces
request_call_netlink_async().
The string will be used when the client load additional config file to
boot, and it must be a valid path or url. Hence, let's refuse non-safe or
non-characters.
On Ubuntu CI on ppc64el, the test randomly fails when /run/udev is not
synced before checking its contents (see #22357). But /run/udev is a
tmpfs and fsync on tmpfs is noop (see `struct shmem_file_operations` in
mm/shmem.c of the kernel), hence, it is not necessary to call fsync on
/run/udev in general. This should be a testing emvironment issue (I
guess it is an issue on nested KVM on ppc64el), instead of an issue on
udev.
C.f. #22357.
Fixes#21603.
Yu Watanabe