Bug 23099

Summary: tail -f from a fifo goes into a busy loop after first read
Product: Base System Reporter: Nick Johnson <freebsd>
Component: kernAssignee: jlemon
Status: Closed FIXED    
Severity: Affects Only Me    
Priority: Normal    
Version: 4.1.1-STABLE   
Hardware: Any   
OS: Any   

Description Nick Johnson 2000-11-25 22:30:00 UTC
When tail -f is reading from a FIFO, after the first read, it goes
into a busy loop between kevent and read.  In the case where I've observed
it, one process is writing to the FIFO with the descriptor set 
non-blocking.  Following is a truss from tail -f showing what happens,
minus the startup junk common to any process:

sigaction(SIGILL,0xbfbffaa0,0xbfbffa88)		 = 0 (0x0)
sigprocmask(0x1,0x0,0x2805d73c)			 = 0 (0x0)
sigaction(SIGILL,0xbfbffa88,0x0)		 = 0 (0x0)
sigprocmask(0x1,0x2805d700,0xbfbffac8)		 = 0 (0x0)
sigprocmask(0x3,0x2805d710,0x0)			 = 0 (0x0)
open("mlog",0,0666)				 = 3 (0x3)
fstat(3,0xbfbffac4)				 = 0 (0x0)
readlink("/etc/malloc.conf",0xbfbff904,63)	 ERR#2 'No such file or directory'
mmap(0x0,4096,0x3,0x1002,-1,0x0)		 = 672112640 (0x280fa000)
break(0x804d000)				 = 0 (0x0)
break(0x804e000)				 = 0 (0x0)
fstat(3,0xbfbff874)				 = 0 (0x0)
break(0x804f000)				 = 0 (0x0)
read(0x3,0x804e000,0x400)			 = 53 (0x35)
break(0x8050000)				 = 0 (0x0)
read(0x3,0x804e000,0x400)			 = 0 (0x0)
24.142.183.133 GET /
write(1,0x804f000,21)				 = 21 (0x15)
	Lynx/2.8.2rel.1 libwww-FM/2.14
write(1,0x804f200,32)				 = 32 (0x20)
kqueue()					 = 4 (0x4)
kevent(0x4,0xbfbffa5c,0x1,0x0,0x0,0xbfbff9f4)	 = 0 (0x0)
read(0x3,0x804e000,0x400)			 = 0 (0x0)
kevent(0x4,0x0,0x0,0xbfbffa5c,0x1,0x0)		 = 1 (0x1)
read(0x3,0x804e000,0x400)			 = 0 (0x0)
kevent(0x4,0x0,0x0,0xbfbffa5c,0x1,0x0)		 = 1 (0x1)
read(0x3,0x804e000,0x400)			 = 0 (0x0)
(...)
kevent(0x4,0x0,0x0,0xbfbffa5c,0x1,0x0)		 = 1 (0x1)
read(0x3,0x804e000,0x400)			 = 0 (0x0)
SIGNAL 2
SIGNAL 2
process exit, rval = 2

Fix: 

Sorry, I don't have a patch, as I'm not yet familiar with kqueue.
How-To-Repeat: 
mkfifo foo
tail -f foo
start a process to open foo for non-blocking writing
write something to foo
Comment 1 bill fumerola freebsd_committer freebsd_triage 2000-11-25 22:38:00 UTC
Responsible Changed
From-To: gnats-admin->jlemon

jlemon wrote kqueue and loves bug reports about it.
Comment 2 Nick Johnson 2000-11-26 22:23:18 UTC
For some reason, trying to fix the trouble by passing in a change with
EV_CLEAR set doesn't work, but freopen & re-adding the events seems to
work fine:

--- forward.c.orig	Sun Nov 26 13:02:38 2000
+++ forward.c	Sun Nov 26 14:15:14 2000
@@ -218,6 +218,18 @@
 		}

 		case USE_KQUEUE:
+			if ((ev->flags & EV_EOF) > 0) {
+				/* Maybe EOF on a fifo; reopen it
+				 * the original is closed so the events
+				 * are deleted
+				 */
+				if ((fp = freopen(fname, "r", fp)) == NULL)
+					ierr();
+				else
+					action = ADD_EVENTS;
+				break;
+			}
+
 			if (kevent(kq, NULL, 0, ev, 1, NULL) < 0)
 				err(1, "kevent");

There may be a more elegant fix than this of course.

-- 
"The reasonable man adapts himself to the world; the unreasonable one
persists in trying to adapt the world to himself.  Therefore all
progress depends on the unreasonable man."
	-- George Bernard Shaw
Nick Johnson, version 2.0b1                   http://www.spatula.net/
Comment 3 jlemon freebsd_committer freebsd_triage 2001-06-15 22:03:38 UTC
State Changed
From-To: open->closed

Fix for the problem has been committed.