bug in big pipe code causing performance problems

Hello,

If no one objects to it, I'd like to commit the following ASAP. It
fixes an obvious bug in the big pipe code, a discrepancy between
how free space is calculated in write vs poll. The bug affects
stable as well.

The bug's implications are less obvious: it make write(2) on a
non-blocking pipe return EAGAIN when poll(2) & select(2) return the
descriptor as ready for write. This in turns causes the libc_r code
to busy-wait on pipes, causing a major performance hog.

As an example the patch boosts the multimedia/transcode port (a big
pipes & threads user) by a factor of 2-3.

Index: sys_pipe.c
===================================================================
RCS file: /home/ncvs/src/sys/kern/sys_pipe.c,v
retrieving revision 1.140
diff -u -r1.140 sys_pipe.c
--- sys_pipe.c	30 Jul 2003 18:55:04 -0000	1.140
+++ sys_pipe.c	30 Jul 2003 21:19:41 -0000
_at__at_ -1041,7 +1041,8 _at__at_
 		if ((space < uio->uio_resid) && (orig_resid <= PIPE_BUF))
 			space = 0;
 
-		if (space > 0 && (wpipe->pipe_buffer.cnt < PIPE_SIZE)) {
+		if (space > 0
+		    && wpipe->pipe_buffer.cnt < wpipe->pipe_buffer.size) {
 			if ((error = pipelock(wpipe,1)) == 0) {
 				int size;	/* Transfer size */
 				int segsize;	/* first segment to transfer */
-- 
Pierre Beyssac	      	    pb_at_fasterix.frmug.org pb_at_fasterix.freenix.org
    Free domains: http://www.eu.org/ or mail dns-manager_at_EU.org