This fixes 2 races wrt probing v4l2 devices:
1. Before this change there was a window where a new udev device can get
added between the udev_enumerate_scan_devices() call in enum_devices() and
the udev_monitor_enable_receiving(this->umonitor); call. If this window was
hit then enum_devices() would not see the device and no udev-event for it
would be received either causing the device to not be seen.
Enabling udev event monitoring before calling udev_enumerate_scan_devices()
fixes this. Note that the code is already prepared to deal with getting
multiple add/change events for the same udev device, so hitting the new
race window where PipeWire may receive both an add- or change-event and
also sees + probes the device from enum_devices() is not a problem.
2. Before this change devices added by enum_devices() would not have
inotify monitoring activated right away because notify.fd = -1 at this
time turning start_watching_device() into a no-op.
These devices without inotify monitoring would then have their access
checked by process_device() calling check_access().
Then after all devices have been enumerated start_monitor() would call
start_inotify() which calls start_watching_device() for all devices added
by enum_devices(). This leaves a window where the ACL can change without
there being an inotify watch for it.
Calling start_monitor() before enum_devices() puts start_inotify()
notify before enum_devices() so that the add_device() calls done
by enum_devices() will now successfully call start_watching_device()
closing this window.
PipeWire is somewhat likely to not notify ACL changes because of this
because PipeWire is part of the systemd user default.target, where as
logind only starts applying the ACLs after GNOME has created the seat
for the GNOME session. So on first login we have PipeWire starting
and logind applying the ACLs at the same time, which allows for the ACL
change to hit the small race window where PipeWire is not monitoring
for ACL changes. Fixing this second race should hopefully resolve
issue #3960.
Closes: #3960
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Previously both `impl_on_notify_events()` and `process_{device,card}()`
called `check_access()`. Avoid that by merging `ACTION_ADD` and
`ACTION_DISABLE` into a single `ACTION_CHANGE` and let `process_{device,card}()`
call `check_access()` and decide what to do.
Split up `process_{device,card}()` to have a separate function that does
the lookup based on the udev device, and only use that when there is
no available reference to the actual device/card object.
Instead of watching /dev, use a separate watch for each device.
This is supposed to achieve the same result as the now reverted
88f0dbd6fc ("v4l2: don't set inotify on /dev"):
Doing inotify on /dev is not a good idea because we will be woken up by
a lot of unrelated events.
There is a report of a performance regression on some IO benchmark
because of lock contention within the fsnotify subsystem due to this.
Instead, just watch for attribute changes on the /dev/videoX files
directly. We are only interested in attribute changes, udev should
notify us when the file is added or removed.
Fixes#3439
Doing inotify on /dev is not a good idea because we will be woken up by
a lot of unrelated events.
There is a report of a performance regression on some IO benchmark
because of lock contention within the fsnotify subsystem due to this.
Instead, just watch for attribute changes on the /dev/videoX files
directly. We are only interested in attribute changes, udev should
notify us when the file is added or removed.
udev/kernel sometimes give "Video Capture N" as ID_V4L_PRODUCT, which is
bogus as a device product name. The ID_MODEL* field seem to always have
a sensible product name.
Get device.product.name always from ID_MODEL*, and use ID_V4L_PRODUCT
only as the last fallback.
The inotify_event field has a variable sized type, so it's recommended to store
it at the end of the storage unit. Fixes gnu-variable-sized-type-not-at-end warning.
When we add a new listener to an object, it will emit the full state
of the object. For this it temporarily sets the change_mask to all
changes. Restore the previous state after this or else we might not
emit the right change_mask for the next listener.
Consider the case where one there are two listeners on an object.
The object emits a change and the first listener wants to enumerate the
changed params. For this is adds a new listener and then triggers the
enumeration. If we set the change_mask to 0 after adding the listener,
the second listener would get a 0 change_mask and fail to update
its state.
udev's ID_MODEL_ID and ID_VENDOR_ID are inconsistent: always 4-digit hex but
sound devices are prefixed with 0x, v4l devices are not. Depending on the
actual ID, the value will look like decimal (1234) or hex (a234).
pw-dump will then print those as either decimal integers (i.e. 0x1234 becomes
decimal 1234) or double (i.e. a234 becomes 41524.00).
Make this consistent by converting the string from hex do decimal where we
get it.
SPA_MEMBER is misleading, all we're doing here is pointer+offset and a
type-casting the result. Rename to SPA_PTROFF which is more expressive (and
has the same number of characters so we don't need to re-indent).
This is more in line with wayland and it allows us to create new
interfaces in modules without having to add anything to the type
enum. It also removes some lookups to map type_id to readable
name in debug.
Remove the monitor API, we can use the device API for it. Make sure
we support creating devices (like alsa) from another device (udev).
Use new object.id to store the object id in the object properties. Use
the port.id/node.id etc to make relations to other objects.
2019-09-20 13:04:14 +02:00
Renamed from spa/plugins/v4l2/v4l2-monitor.c (Browse further)
