Alan Cox wrote:
> > Looking at your patch, I see no reason not to add it. If nobody wants to
> > use it, then fine. But if someone is creating a application/driver pair
> > and has a compelling reason to use it, then there is a standardized way.
>
> read() is surely the standard way
Yes, that I forgot to mention: read() can do userspace dma too in some
cases. For partial reads and non-blocking I/O a kernel bounce buffer is
needed, in other cases the driver can do zerocopy I/O for capture via
read().
BTW: What do the standards (POSIX et. al.) say about the status of a
read buffer after a interrupted system call? undefined? not
modified? i.e. if read() starts the zerocopy I/O, then goes sleep
interruptible until the grabber card is done, it might happen that
the read() system call returns with -EINTR while the DMA transfer
is in progress. Is this ok? Or should I better go sleep
uninterruptible?
Gerd
--
Damn lot people confuse usability and eye-candy.
