media: ngene: compensate for TS buffer offset shifts

A possible hardware bug was discovered when using CA addon hardware
attached to the ngene hardware, in that the TS input buffer much likely
will shift and thus become unaligned to 188 byte blocks (a full TS frame)
when things like CA module initialisation (which happens via differing
communication paths) take place. This causes the TS NULL removal in
tsin_exchange() to fail to detect this previously inserted data and thus
causes userspace applications to receive data they didn't sent beforehand
and ultimately cause troubles.

On driver load with an inserted CAM, buffers are fine at first (note that
the driver has to keep the communication running from/to the card by
inserting TS NULL frames, this is done in tsout_exchange() via
FillTSBuffer() - that data is simply sent back by the hardware):

  offset | 0    1   2   3   4   5 .... 188 189 190 191 192 193 .... 376
  data   | 47  1f  ff  10  6f  6f ....  47  1f  ff  10  6f  6f ....  47

After a few seconds, the CA module is recognised and initialised, which is
signalled by

  dvb_ca_en50221: dvb_ca adapter X: DVB CAM detected and initialised successfully

This is where the first shift happens (this is always four bytes), buffer
becomes like this:

  offset | 0    1   2   3   4   5 .... 188 189 190 191 192 193 .... 376
  data   | 6f  6f  6f  6f  47  1f ....  6f  6f  6f  6f  47  1f ....  6f

Next, VDR, TVHeadend or any other CI aware application is started, buffers
will shift by even more bytes. It is believed this is due to the hardware
not handling control and data bytes properly distinct, and control data
having an influence on the actual data stream, which we cannot properly
detect at the driver level.

Workaround this hardware quirk by adding a detection for the TS sync byte
0x47 before each TS frame copy, scan for a new SYNC byte and a TS NULL
packet if buffers become unaligned, take note of that offset and apply
that when copying data to the DVB ring buffers. The last <188 bytes from
the hardware buffers are stored in a temp buffer (tsin_buffer), for which
the remainder will be in the beginning of the next hardware buffer (next
iteration of tsin_exchange()). That remainder will be appended to the
temp buffer and finally sent to the DVB ring buffer. The resulting TS
stream is perfectly fine, and the TS NULL packets inserted by the driver
which are sent back are properly removed. The resulting offset is being
clamped to 188 byte segments (one TS packet). Though this can result in
a repeated TS packet if the overall offset grows beyond this (and it
will grow only on CA initialisation), this is still way better than
unaligned TS frames and data sent to userspace that just isn't supposed
to be there.

This compensation can be toggled by the ci_tsfix modparam, which defaults
to 1 (enabled). In the case of problems, this can be turned off by setting
the parameter to 0 to restore the old behaviour.

Signed-off-by: Daniel Scheller <d.scheller@gmx.net>
Signed-off-by: Mauro Carvalho Chehab <mchehab@s-opensource.com>

1 files changed, 3 insertions, 0 deletions

diff --git a/drivers/media/pci/ngene/ngene.h b/drivers/media/pci/ngene/ngene.h
index 66d8eaa28549..01d9f1b58fcb 100644
--- a/drivers/media/pci/ngene/ngene.h
+++ b/drivers/media/pci/ngene/ngene.h
@@ -732,6 +732,9 @@ struct ngene_channel {
 #endif
 
 	int running;
+
+	int tsin_offset;
+	u8  tsin_buffer[188];
 };