Although we added a detection mechanism for the presence of the Kernel
bug earlier, reading from the incorrectly mapped memory might cause a
bus error on some ARM systems.
With the overall performance benefit being rather minimal for the
data rates of rtl-sdr, disable zero-copy by default.
Removed the linked list structure, locking, and malloc/free on each USB callback.
Replaced it with a non blocking ring buffer for performance. This resulted
in a large performance improvement when using an Raspberry Pi or Raspberry Pi Zero W
as a rtl_tcp server. The sample rate could be doubled and instead of wired ethernet
wifi could be used. Frequnecy change lag was reduced greatly also.
Signed-off-by: Stephen Blinick <stephen@stesoft.com>
The Linux Kernel has a bug on ARM/ARM64 systems where the USB CMA
memory is incorrectly mapped to userspace, breaking zerocopy.
When the Kernel allocates the memory, it clears it with memset().
If the mapping worked correctly, we should have zeroed out buffers,
if it doesn't, we get random Kernel memory. We now check for this,
and fall back to buffers in userspace if that's the case.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
With just 'inline', if the compiler decides not to inline them, it isn't
required to emit them at all. For some targets with -Os that is causing
build failures, see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86360.
Perhaps we might consider using '__attribute__((always_inline))' for
GCC builds, but 'static inline' is a good start.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
As pointed out by Carl Laufer on the mailing list,
turning the loop-through output off reduces the
current consumption by 10-20mA which in turn reduces
the heat a bit. So far there seem to be no devices
that have anything connected to the loop-through output.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
This code stayed unchanged for many years, but for some reason
rtl_adsb started hanging upon exit:
*b66116a5164b69281eacc42ae950;
^CSignal caught, exiting!
<------ hangs here forever
Examining it with gdb reveals that the demod thread waits
peacefully on the condition variable, which we're trying to
destroy. Either the signals killed all threads before, or
condition variables were possible to destroy while other
threads still waited on them.
The easiest fix appears to be just cancel the demod thread
and wait for it to exit before proceeding for the door.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
Decreases CPU load especially for embedded machines.
Requires Linux >= 4.6 and libusb >= 1.0.21. If this
is not the case or the allocation fails, we will
fall back to buffers allocated in userspace.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
Since a typo in rtlsdr_set_gpio_output() was fixed,
FC0012 tuners were not detected anymore, as the reset pin
is actually 4, not 5.
Thanks to David Basden et al for reporting the bug.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
At least one distributor of rtl-sdr dongles (rtl-sdr.com) added
a bias-t to their dongles which could be toggled via GPIO P0 of the
RTL2832U chip.
source: http://www.rtl-sdr.com/rtl-sdr-blog-v-3-dongles-user-guide/
Signed-off-by: Steve Markgraf <steve@steve-m.de>
This improves SDR performence for nearby channel interference.
As a sideeffect also improves dynamic range becase ADC is not overloaded
by onwanted singlas.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
In contradiction with the documentation, e4k_reg_write() was
returning the value 2 upon success (the number of bytes written),
not zero as documented.
Patched to return zero upon success.
As a consequence, this corrects also the return value of
e4k_reg_set_mask() and e4k_dc_offset_calibrate().
So far we had 32 * 256KB which was a bit overkill,
15 are more than enough.
15 was chosen instead of 16 because at least on Linux
there seems to be a system-wide limit of 63 transfers
(when they are 256KB large), so 4 dongles can be used
on a single machine without lowering the default transfer
number.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
Otherwise the new transfer status does not propagate
into the xfer->status and we try to cancel all transfers
twice.
Also replace the ifdefs with a single macro for backwards
compatibility.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
libusb < 1.0.9 doesn't have libusb_handle_events_timeout_completed(),
but libusb <= 1.0.8 doesn't have version.h, so we need to check
for the function.
The cmake-code was borrowed from UHD, which also checks
for libusb_error_name(), we add that as well since it might come
handy later on.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
Now we're using libusb_handle_events_timeout_completed()
instead of the deprecated libusb_handle_events_timeout(),
and notify our async state machine when rtlsdr_cancel_async()
is called. This speeds up exiting from any program that uses the
device asynchronously.
Signed-off-by: Steve Markgraf <steve@steve-m.de>
Thanks to Joris van Rantwijk for finding what seems to be
a hardware limitation/bug (bit 28 of the rsamp register being
forced to the value of bit 27).
Signed-off-by: Steve Markgraf <steve@steve-m.de>
rtlsdrblog
Steve Markgraf
Stephen Blinick
David Woodhouse
Pete Zaitcev
Harald Welte
Kacper Michajłow
hayati ayguen
Dimitri Stolnikov
Lucas Teske
Hoernchen
Jiří Pinkava
lcl
Michael Tatarinov
Denys Berkovskyy
Kyle Keen