On 27.05.2012 19:43, Gianluca Gennari wrote:
Hi Antti,
Il 27/05/2012 15:10, Antti Palosaari ha scritto:
On 26.05.2012 01:47, Antti Palosaari wrote:
I was planning make DVB USB be able to switch USB streaming parameters
dynamically. I mean [struct usb_data_stream_properties] parameters.
Currently it reserves USB streaming buffers when device is plugged. Own
buffer is reserved for each frontend, which means currently 1-3
streaming buffers depending on device.
Basically I see USB TS as a DVB USB device property - not property of
frontend. USB TS is interface between computer and USB-bridge and amount
of parallel USB TS or TS configurations depends on USB-bridge
capabilities. Sometimes used USB TS could be configured to fit better
used stream. Smaller buffers for the narrow radio stream and biggest
buffers for the wide DVB-S2 stream.
I was wondering how to resolve that situation? It is not very big
problem currently but I still want to make it better as there is surely
coming new devices that needs better control for the USB streaming
parameters. Currently there is mxl111sf driver which seems to offer 6
different streaming configurations but AFAIK only three is currently
used as there is 3 frontends and each one has own streaming parameters -
and buffers - even only one can be used at the time.
1. Configure streaming parameters (alloc buffers) every time when
streaming is started? IIRC that causes some problems lately for em28xx
as memory goes dis-coherent and buffers cannot allocated.
2. Allocate buffers (streaming configuration) for all needed device use
configurations at very beginning. Then add some logic to map streaming
config to frontend at runtime. That is quite near what mxl111sf does
currently.
After looking existing code carefully and doing some tests I think I
will implement that a way it changes buffers at runtime to fit current
needs.
int usb_urb_submit(struct usb_data_stream *stream) is called when
streaming is started. At that point it is possible to compare already
reserved buffers and buffers needed. If needed buffers are larger it is
possible to reallocate bigger buffers. Of course it is not needed to
decrease buffers even smaller stream configurations is asked, just to
avoid repetitive allocations of large chunks of coherent memory.
I have a patch that enables a similar behavior on the em28xx driver (for
digital devices). I used it to play with the buffer size in order to
determine the minimum working size. I added a couple of parameters to
the em28xx module that can be used to change the buffer number and size
at runtime. If anybody is interested, I can post this patch to the list
for comments.
Maybe there is no idea of merge it but it could be interesting to see still.
1) allocate buffers at the init as now (callback to read stream
properties from the driver)
2) re-allocate bigger buffers in runtime if needed (callback to read
stream properties from the driver)
3) free all buffers when device is disconnected
I think the deciding factor here is the ratio between the default
initial size and the maximum buffer size. If it is something like 2X,
then maybe it's better to keep the code simple and just allocate the
bigger size from the beginning. If the difference is an order of
magnitude or more, then using dynamic allocation is probably worth the
effort.
The problem is here driver author could make buffer size calculation
error easily :) But as it will likely crash very soon or at least some
other effects like broken stream it should never go to the public. And
it is cheap to add check for that too, just few lines of code that
compares buffer sizes when streaming is started.
I think I will go to that as it is very simple.
Another possibility is to allocate a single memory pool of coherent
memory at the init, big enough to cover all the possible modes for the
given device. Then it is possible to map the URB buffers into this
memory pool at runtime, choosing the optimal size for each mode.
The memory pool is freed when the device is disconnected.
This approach avoids any reallocation so it's safer on ARM/MIPS hardware.
I do not see any difference to the situation largest needed buffers are
reserved at the device initialization unless dynamically re-allocate.
But I am going to skip it as it is easier to alloc just as big buffer as
needed at worst case.
For example, the as102 driver allocates a single memory pool in
dev->stream, then it maps the URB buffers into this area (see function
as102_alloc_usb_stream_buffer() ). In this case the URBs are mapped
statically, but nothing prevents to dynamically remap the URB buffers at
runtime if needed.
I should look that driver too. But it sounds a little bit more complex
to implement than alloc own buffer in the loop for every used URB. There
is also drawback that it needs very large, 5-6 times larger, coherent
chunk than per URB buffer.
Basically those USB buffers are type of (buffer count) x (buffer size).
Buffer count is same as used count of URBs, typically ~5, and buffer
size is the payload size of one URB.
Here is some background info (em28xx buffer alloc failures) why to avoid
continuous buffer allocs / frees.
http://www.spinics.net/lists/linux-media/msg44209.html
There is also the old thread about the issue:
http://www.mail-archive.com/linux-media@xxxxxxxxxxxxxxx/msg42776.html
I have one remaining unclear question still. Is there any other drawback
to use more URBs at the same time than it eats memory?
So if I have DVB-T (~22Mbit/sec) frontend which uses 5 URBs in chain and
then I have DVB-C (~44Mbit/sec) frontend which uses same buffersize but
there is 6 URBs in chain. In my understanding I will not lose anything
but memory when I use more URBs in parallel to move data from the device.
regards
Antti
--
http://palosaari.fi/
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
