Persistent network rules write out new rules files. When rules change,
we need to kill all workers to update the in-memory copy of the rules.
We need to make sure, that a worker finshes its work for all device
messages it has accepted, before it exits after a SIGTERM from the main
process.
Event processes now get re-used after they handled an event. This reduces
pressure on the CPU significantly because cloned event processes no longer
cause page faults in the main daemon. After the events have settled, the
no longer needed worker processes get killed.
Directory lookups show up in profiling. The queue files are responsible
for a large proportion of file-related system calls in udev coldplug.
Instead of creating a file for each event, append their details to a
log file. The file is periodically rebuilt (garbage-collected) to
prevent it from growing indefinitely.
This single queue file replaces both the queue directory and the
uevent_seqnum file. On desktop systems the file tends not to grow
beyond one page. So it should also save a small amount of memory in
tmpfs.
Tests on a running EeePC indicate average savings of 5% *udevd* cpu time
as measured by oprofile. __link_path_walk is reduced from 1.5% to
1.3%. It is not completely clear where the rest of the gains come from.
In tests running ~400 events, the queue file is rebuilt about 5 times.
Signed-off-by: Alan Jenkins <alan-jenkins@tuffmail.co.uk>
Instead of of our own private monitor socket, we send the
processed event back to our netlink socket, to the multicast
group 2 -- so any number of users can listen to udev events,
just like they can listen to kernel emitted events on group 1.
There's still a slight race condition when using udevadm settle, if the
udev daemon has a pending inotify event but hasn't yet generated the
"change" uevent for it, the kernel and udev sequence numbers will match
and settle will exit.
Now udevadm settle will send a control message to udevd, which will
respond by sending SIGUSR1 back to the waiting udevadm settle once it
has completed the main loop iteration in which it received the control
message.
If there were no pending inotify events, this will simply wake up the
udev daemon and allow settle to continue. If there are pending inotify
events, they are handled first in the main loop so when settle is
continued they will have been turned into uevents and the kernel
sequence number will have been incremented.
Since the inotify event is pending for udevd when the close() system
call returns (it's queued as part of the kernel handling for that system
call), and since the kernel sequence number is incremented by writing to
the uevent file (as udevd does), this solves the race.
When the settle continues, if there were pending inotify events that
udevd had not read, they are now pending uevents which settle can wait
for.
Signed-off-by: Scott James Remnant <scott@ubuntu.com>
This allows you to re-process the rules if the content of the device
has been changed, most useful for block subsystem to cause vol_id to
be run again.
udevd's event_queue_manager loop is pretty sensitive to the
delays of exiting child processes. I found that it helps boot
times if we try to reap children as quickly as possible.
This patch changes event_queue_manager to call sigchilds_waiting
if it finds a signal has been received.
Signed-off-by: Olaf Kirch <okir@suse.de>
In certain cut-down situations such as an installer or inside the
initramfs, we simply don't have any kind of name service. While we
could use rules without OWNER or GROUP, it's better to have the same
rules as a full system and have udevd ignore those parts of the rules.
Adds a --resolve-names=never switch to udevd that has this effect.
udevd uses a rather old-fashioned way of handling signals
while waiting for input through select (ie by using an unnamed
pipe, to which the signal handler writes one byte for every signal
received). This is rather awkward and may potentially even block
if we receive more signals than the kernel's pipe buffer.
This patch replaces all of that with ppoll, which was designed
for this purpose.
It also removes the SA_RESTART flag from all installed signal
handlers, because otherwise the ppoll call would just be restarted
after handling eg a SIGCHLD.
Signed-off-by: Olaf Kirch <okir@suse.de>
This crops up in my threaded udevd profiles from time to time.
It's not consistent - probably due to variations in the number
of concurrent events - but it can hit 4% user time and higher.
The change halves the user time spent in compare_devpath().
Signed-off-by: Alan Jenkins <alan-jenkins@tuffmail.co.uk>
