This didn't use to matter, because none of the BLE-using backends did
retry on timeout until recently.
But Jef started doing packet sending retry for the Mares Icon backend,
and now we should make sure to distinguish the "IO failed" from "IO
timed out" cases.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
See https://www.kdab.com/goodbye-q_foreach/
This is reduced to the places where the container is const or can be made const
without the need to always introduce an extra variable. Sadly qAsConst (Qt 5.7)
and std::as_const (C++17) are not available in all supported setups.
Also do some minor cleanups along the way.
Signed-off-by: Rolf Eike Beer <eike@sf-mail.de>
Some divecomputer backends (ok, right now really only the Aqualung i770R
and i300C) want to know the bluetooth name of the dive computer they
connect to, because the name contains identifying information like the
serial number.
This just adds the support for that to our Qt BLE code.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Because some BLE operations can be very slow (device and service
discovery etc), we have some rather excessive default timeout for BLE
(currently set to 12 seconds).
But once we actually have started doing IO, that long timeout can be a
big performance problem, when the libdivecomputer backend has support
for retry and packet loss.
For that reason, libdivecomputer has a 'set_timeout()' function that
allows the divecomputer backend to say how quickly it expects the dive
computer to answer before the backend will start resending packets.
Let's just implement that for the actual IO side of BLE too. The
default timeout value remains the general BLE timeout, and this only
affects the actual IO phase, but it improves things enormously for the
case where there is packet loss at that point.
For example, on the Aqualung i770R, the timeout for packet loss ends up
now being just one second rather than the full 12 seconds of default BLE
timeout. Which gets the retry going much faster.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When we enable notifications, we actually want to make sure to wait for
that write to have completed before we start communicating with the
device, because otherwise we might lose notification events.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In commit 30fb7bf35c ("qt-ble: set up infrastructure for better
preferred service choice") I moved the service filtering from the
addService() callback into the "select_preferred_service()" function
that picks the right service for the device.
That was nice for debugging, since it meant that we showed the details
of _all_ services, but it also meant that we ended up starting service
discovery on _all_ services, whether they looked at all interesting or
not.
And that can make the BLE device discovery process quite a bit slower.
The debugging advantage is real, but honestly, service discovery can
generally be better done with specialized tools like the Nordic nRF app,
so the debugging advantage of just listing all the details of all the
services is not really worth the discovery slowdown in general.
So move the basic "filter by uuid" back to the service discovery phase,
and don't bother starting service detail discovery for the services that
we can dismiss immediately just based on the service UUID.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The existing BLE dive computers treat BLE as the packetized protocol it
is, and read whole packets at a time.
However, the Mares BlueLink backend treats it as just a basic "serial
over BLE" transport, and for historical reasons reads the reply packets
in smaller chunks.
This allows that kind of IO behavior, where if the divecomputer backend
reads just a part of a packet, we'll split the packet, return the part
the user asked for, and push back the leftover packet onto the received
packet queue.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is perhaps overly verbose, but the timing details helped figure out
some EON Core download issues, and it's nice to see when things actually
happen.
It's also good to see when the data actually enters our queues, and when
we read and write the packets. That might help debug the issues Fabio
is seeing with the Mares Bluelink.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We used to just find all services and connect the characteristics change
signal etc to them all, but we really only care about the actual
preferred service that we'll be using.
So move the qt ble signal connection to after we've selected the
preferred service that we will actually be enabling notifications on and
do the writes to.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We used to just blindly pick "first" and "last" characteristic from the
preferred service, and that was stupid but happened to work for the dive
computers we supported. Note that for some of them, "first" and "last"
was actually the *same* characteristic, since it could be a single one
that supported both.
However, this first/last hack definitely doesn't work for the Mares
BlueLink BLE dongle, and it's really all pretty wrong anyway.
So re-organize the code to actually look at the properties of the
characteristics. I don't have a BlueLink to test with, but my EON Core
and Shearwater Perdix AI are still happy with this, and the code
conceptually makes a lot more sense.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We used to just pick the first non-standard service we found (with a
special case for the Heinrichs Weikamp dive computers that have an
actual registered standard service).
We then waited for that service to finish discovery, and started using
it.
This changes the logic to wait for _all_ services to finish discovery,
and then after that we pick the one we like best. Right now the rule
for picking a preferred service is the same one we had before, but the
difference is that we now have the full discovery data, so we *could* do
something better.
Plus this makes our debug messages a lot more legible, when we don't
have the mix of overlapping service discovery with the actual IO we do
to the preferred service.
NOTE! This doesn't much matter for most of the dive computers that we
currently support BLE for. They don't tend to have a lot of odd
services.
But at least both the Mares BlueLink and the Garmin Descent both have
multiple services and it's not obvious which one to use, and this will
make it not only easier to debug those, it will make it easier to pick
the right preferred service descriptor to use.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is not only much clearer (and smaller code), but it also lowers the
latency for the waiting, since we don't always wait for the full 100ms.
Get rid of the now unused "waitfor()" function that just unconditionally
waited for 100ms.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We had a special-case for the Shearwater case, let's just make it
slightly more generic and add Garmin to the list of vendors that want a
random BLE address rather than a static one.
The Bluez model of having to state this explicitly - but not giving the
information to the user - is completely broken and this is all very
annoying, credit goes to Wojciech Więckowski for pointing this out.
Of course, right now we don't actually know how to parse the BLE stream
from the Garmin Descent, but with this (and some libdivecomputer
hackery) I actually get connected and start receiving data. That we
then can't parse, but that's hopefully just a libdivecomputer update
away.
Pointed-out-by: Wojciech Więckowski <xplwowi@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
This should never happen, since our interface is bassically synchronous,
but it could happen with delayed replies that came in just after we
decided to re-transmit a command.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Our model of waiting for 100ms before re-checking if we got a packet
over BLE resulted in potentially horrendously bad latency for received
packets.
That isn't just a possible performance issue, it actually seems to cause
IO errors with my Suunto EON Core. I'm not entirely sure why, but it
might simply be some timing interaction, particularly since the IO
errors seemed to primarily happen when the dive computer itself was also
busy updating the screen (ie if you pressed buttons on the dive computer
to switch to compass mode, for example).
So replace the silly hardcoded 100ms "waitFor()" function with a
WAITFOR() macro that checks the provided expression every time through
the loop, which gets us a much lower latency (we basically check every
ten milliseconds).
The macro is not beautiful, but it WorksForMe(tm).
This makes a huge difference to the reliability of the download for me,
and might matter for some other dive computers too.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
the postEvent is only called when downloading from a dc
with bluetooth, so in most it does not have an effect
on the deleteLater() in the code.
there are no reason to do special cleanup while waiting
for bluetooth
QEvent::DeferredDelete is not supported on iOS.
Signed-off-by: Jan Iversen <jani@apache.org>
This converts our old custom IO model to the new model that
libdivecomputer introduced. This is partly based on Jef's rough patch
to make things build, with further work by me.
The FTDI code is temporarily disabled here, because it will need to be
integrated with the new way of opening devices.
The ble_serial code goes away entirely, since now libdivecomputer knows
about BLE transport natively, and doesn't need to have any serial
wrapper around it.
Signed-off-by: Jef Driesen <jef@libdivecomputer.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Remove a semicolon after Q_OBJECT and a few others after the closing
braces of while loops.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Replace constructs of the kind
s.toUtf8().data(),
s.toUtf8().constData(),
s.toLocal8Bit().data(),
s.toLocal8Bit.constData() or
qUtf8Printable(s)
by
qPrintable(s).
This is concise, consistent and - in principle - more performant than
the .data() versions.
Sadly, owing to a suboptimal implementation, qPrintable(s) currently
is a pessimization compared to s.toUtf8().data(). A fix is scheduled for
new Qt versions: https://codereview.qt-project.org/#/c/221331/
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
Each callsite of saveBtDeviceInfo() has a QString, which is converted
to a C-string, passed and immediately converted back. Remove these
conversions by taking a reference to QString directly.
getBtDeviceInfo() is not as clear. Here, the callsite has a C-string
handed down from libdivecomputer. Nevertheless, pass a reference of
QString here as well. Firstly, for reasons of symmetry. Secondly,
to avoid multiple conversions in the getBtDeviceInfo() functions.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
On BLE connection timeout a weird error-message was shown, because
the controller was still in connecting state and no error string was
set. Therefore, handle the timeout case with a special case label.
Moreover, remove three unnecessary calls to disconnectFromDevice(),
which is called in the destructor of the controller anyway (verified
by looking at Qt source).
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
1) Destroy QLowEnergyService objects in destructor of BLEObject.
2) Let BLE object take ownership of the controller so that the
latter can be destroyed in the destructor of the former. This
introduces a certain ownership subtlety, which could be solved by
allocating the controller object in the BLE object. But let's
first do the less intrusive thing.
3) Destroy the BLE object for two error conditions.
Signed-off-by: Berthold Stoeger <bstoeger@mail.tuwien.ac.at>
While this interface is deprecated, too much in our existing code depends
on being able to create the QLowEnergyController with just the address.
Additionally, createCentral() is new in Qt 5.7 and therefor this broke
builds on Linux distros that are still on 5.6.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
The ordering on Mac appears to be random, but after looking through the
various successful logs of BLE downloads, it seems we always wrote to the
ClientCharacteristicConfiguration descriptor. So try to find that one first,
and only grab the first descriptor in the list if we didn't find a
ClientCharacteristicConfiguration descriptor.
Signed-off-by: Dirk Hohndel <dirk@hohndel.org>
In the serial api for libdivecomputer is ok to send NULL as the int
pointer actual, if you dont't care about how many bytes that where
actually read or written.
This makes sure we don't crash if the ble backend where ever used with
such a backend.
Signed-off-by: Anton Lundin <glance@acc.umu.se>
Commit 709c1df2af introduced a hard blocking read for BLE devices.
This did break BLE reads from multiple DCs, and (in hindsight) was not
a correct implementation. It would require, for example, dynamic
read buffers as especially profile data grows with dive time, and
in addition, and more importantly, also the OSTC libdc parser cannot
process the entire profile of a dive at once (but likes to receive
it in 1K blocks). So, basically, it introduced issues, and did not
solve the OSTC read.
This commit reverts this hard blocking read (and as such will break
OSTC BLE reads). But it enables removal of the special cases for
the EON Steel and G2.
A next commit will solve OSTC BLE reads.
Signed-off-by: Jan Mulder <jlmulder@xs4all.nl>
Handle credits. Do not just ask for maximum credits all the time as this
will stop the download. Also do not let the credits go back to 0 (while
this might work, this is not tested). Getting back the 0 credits stops
the download, and even when it can be restarted, it is less efficient
(and not needed). Notice also that it takes some time before a grant
request is honoured. During testing I saw reception of up to 25 packets
between request and grant. So a lower bound for the request of
32 packets seems resonable.
One aspect the Telit/Stollmann TIO puzzeled me. Sections 4.1 and 4.2
both talk about credits, but my hyphothesis is that there are two
credits counters in play. One for traffic either way. This commit
only deals with credits granted by Subsurface to the OSTC to send
data. Credits granted by the OSTC to allow Subsurface to send new
commands is NOT part of this commit, and is seemingly not needed
in our scenario. As we only send new commands to the OSTC when
a previous one is finished (per HW's interface spec), the OSTC
does not run out of credits to receive commands.
Signed-off-by: Jan Mulder <jlmulder@xs4all.nl>
The current BLE read reads just one 20 bype packet. That packet size is set
in ble_serial_ops, so, without being able to test on anything other than
a OSTC3, I assume that this holds for other BLE DCs too. So, I think is
is weird that those interfaces work with the current read() of just one
packet at the time.
As we need a blocking read (at least for the OSTC parser), just read all
data that is available on the input. And when we think we are done, give
the QtEventloop control to see if there is more, and process that incoming
data as well. All this basically implements a blocking read.
CAVEAT 1: This might break the reading from the currently working BLE devices.
CAVEAT 2: With this, I still cannot read the OSTC3 completely. For
developers familiar with the HW transfer protocol: it just stops while
reading the first full dive (header + profile) command 0x66, despite
correctly reading about 5Kb of data before. For some
reason, I do not believe that this is related to this commit.
CAVEAT 3: All above tested on Linux Desktop with bluez stack, and
confirmed NOT to work on Android 7.1.2, build with Qt 5.9.0, And
yes, I know 5.9.1 recommended.
Signed-off-by: Jan Mulder <jlmulder@xs4all.nl>
1) As the OSTC sends data to the BLE central role (the SSRF client) over 2
characteristics, we have to filter the administrative credit data from
the actual dive data that it received. The characteristcStateChanged
function is adapted for this.
2) We have to be sure that the Terminal Client I/O is fully defined during
opening the connecton to the OSTC. From 6d505b24f0c15 we can see
that the last step in setting up the terminal interface is the grant
of credits. This is done by writing to the proper (the only one, with
id = 0x2902) descriptor of the credits RX characteristic. The here
added slot is triggered on the completion of write of credits marking
the final stage of the setup.
Signed-off-by: Jan Mulder <jlmulder@xs4all.nl>
