Re: Make pipe data structure be a circular list of pages, rather

[Ingo Oeser <ioe-lkml@axxeo.de>]

Hi Linus,

Linus Torvalds wrote:
> +static long do_splice_from(struct inode *pipe, struct file *out, size_t len, unsigned long flags) 
> +static long do_splice_to(struct file *in, struct inode *pipe, size_t len, unsigned long flags) 
> +static long do_splice(struct file *in, struct file *out, size_t len, unsigned long flags) 
> +asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned long flags) 

That part looks quite perfect. As long as they stay like this, I'm totally 
happy. I have even no problem about limiting to a length, since I can use that
to measure progress (e.g. a simple progress bar). 

This way I also keep the process as an "actor" like "linux@horizon.com" pointed out.

It has unnecessary scheduling overhead, but the ability to stop/resume
the transfer by killing the process doing it is worth it, I agree.

So I would put a structure in the inode identifying the special device
and check, whether the "in" and "out" parameters are from devices suitable
for a direct on wire transfer. If they are, I just set up some registers
and wait for the transfer to happen. 

Then I get an interrupt/wakeup, if the requested amount is streamed, increment some 
user space pointers, switch to user space, user space tells me abort or stream
more and I follow. 
 
Continue until abort by user or streaming problems happen.

Just to give you an idea: I debugged such a machine and I had a hard hanging 
kernel with interrupts disabled. It still got data from a tuner, through an
MPEG decoder, an MPEG demultiplexer and played it to the audio card.
Not just a buffer like ALSA/OSS, but as long as I would like and it's end to end
without any CPU intervention.

That behavior would be perfect, but I could also live with a "pushing process".


Regards

Ingo Oeser