Ingo Oeser wrote:
> 
> Hi Manfred,
> 
> I'm running your patches [1] with sucess for a while now.
> 
> Did you get any feedback about problems regarding these patches?
>
No feedback yet.

> They seem to work for me, but there seems to be a memleak in
> 2.4.x (x: 0-2), which I'm chasing down.
>
You are right, there's a stupid bug in my poll patch:

p = kmalloc(PAGE_SIZE);
free_page((unsigned long)p);

This causes a memory leak for the slab control structures.

> 
> [1] put on http://www.tu-chemnitz.de/~ioe/fastpipe.patch
>
That's davem's original 
I rewrote it because I didn't like the nested loops. pipe_read() and
_write() were never easy to follow, and adding yet another set of goto &
for(;;) loops is IMHO a bad idea.

The main difference between my patch and davem's patch is that davem
uses zerocopy for exactly PAGE_SIZE sized transfers, I only use it for
transfers larger than PIPE_BUF.

<<<<< davem's patch
+       if (count >= PAGE_SIZE &&
+           !(filp->f_flags & O_NONBLOCK)) {
<<<<<<< my patch
+  if (count > PIPE_BUF && chars == PIPE_SIZE &&
+      (!(filp->f_flags & O_NONBLOCK))) {
<<<<<<<

On i386 PIPE_BUF is 4096, and thus

* davem's patch is faster, since glibc by default writes in 4096 byte
blocks
* davem's patch breaks apps that assume that write(,PIPE_BUF) after
poll(POLLOUT) never blocks, even for blocking pipes.

I've attached the newest versions of both patches.


--
	Manfred