On Sat, 2009-03-28 at 16:00 -0700, Steve Harrington wrote: > In general read(2) is not guaranteed to return a complete record. Especially in non-blocking I/O mode. > For > section reads from the demux device the amount of data returned will > almost certainly be less than requested. I've never seen a section > longer than a single TS packet and the recommended length for a read is > 4096 bytes. > The questions: are the v4l-dvb drivers guaranteed to return > a complete section for each read? Well, the spec says it *tries* to return 1 whole one. >From the Linux DVB API spec, Section 3.2.3 read() (from the demux): "When returning section data the driver always tries to return a complete single section (even though buf would provide buffer space for more data). If the size of the buffer is smaller than the section as much as possible will be returned, and the remaining data will be provided in subsequent calls. The size of the internal buffer is 2 * 4096 bytes (the size of two maximum sized sections) by default." > If not, is there a preferred method > to reassemble split sections? 1. Open the device node in non-blocking IO mode. 2. In a loop select() or poll() for when the device has data available. It may be prudent to set a timeout here, so your program won't hang forever, if something goes wrong. 3. Read() as much data as you can or care to read, when the device is ready. Don't forget trap error conditions: select() timed out, select() was interrupted, read() was interrupted, read() was at the end-of-file/stream, or all the other wonderful errno's on the read() man page. 4. Shovel the data somewhere, keeping track of the bytes you've read in. (buffer management happens here in your program) 5. Call a function to check if you have enough bytes to do something useful (i.e. a section's worth). 6. If you've got enough bytes to do something useful, call the function to drain off some of the bytes you've captured and do something useful with the bytes. (buffer management happens here in your program) 7. Go back to the beginning of the loop (#2) or check some exit condition that you care about and drop out of the loop. Nothing sepcial; just textbook, single-threaded, Unix. Regards, Andy -- 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
