Re: r247839: broken pipe - for top, sudo and ports

On Thu, Mar 07, 2013 at 04:54:01AM -0100, Jan Beich wrote:
> Jilles Tjoelker <jilles_at_stack.nl> writes:

> > On Tue, Mar 05, 2013 at 08:59:09PM +0100, Hartmann, O. wrote:

> >> A "truss top" reveals this, is this of help?

> >> [...]
> >> stat("/etc/nsswitch.conf",{ mode=-rw-r--r--
> >> ,inode=162310,size=1007,blksize=32768 }) = 0 (0x0)
> >> stat("/etc/nsswitch.conf",{ mode=-rw-r--r--
> >> ,inode=162310,size=1007,blksize=32768 }) = 0 (0x0)
> >> stat("/etc/nsswitch.conf",{ mode=-rw-r--r--
> >> ,inode=162310,size=1007,blksize=32768 }) = 0 (0x0)
> >> stat("/etc/nsswitch.conf",{ mode=-rw-r--r--
> >> ,inode=162310,size=1007,blksize=32768 }) = 0 (0x0)
> >> stat("/etc/nsswitch.conf",{ mode=-rw-r--r--
> >> ,inode=162310,size=1007,blksize=32768 }) = 0 (0x0)
> >> socket(PF_LOCAL,SOCK_STREAM,0)                   = 4 (0x4)
> >> connect(4,{ AF_UNIX "/var/run/nscd" },15)        = 0 (0x0)
> >> fcntl(4,F_SETFL,O_NONBLOCK)                      = 0 (0x0)
> >> kqueue(0x80183b000,0x80122fc58,0x10,0x80062b308,0x80183b010,0x2) = 5 (0x5)
> >> kevent(5,{0x4,EVFILT_WRITE,EV_ADD,0,0x0,0x0},1,0x0,0,0x0) = 0 (0x0)
> >> kqueue(0x5,0x7fffffffd2e0,0x1,0x0,0x0,0x0)       = 6 (0x6)
> >> kevent(6,{0x4,EVFILT_READ,EV_ADD,0,0x0,0x0},1,0x0,0,0x0) = 0 (0x0)
> >> kevent(5,{0x4,EVFILT_WRITE,EV_ADD,1,0x4,0x0},1,0x0,0,0x0) = 0 (0x0)
> >> kevent(5,0x0,0,{0x4,EVFILT_WRITE,EV_EOF,0,0x2000,0x0},1,0x0) = 1 (0x1)
> >> sendmsg(0x4,0x7fffffffd290,0x0,0x1,0x1,0x0)      ERR#32 'Broken pipe'
> >> SIGNAL 13 (SIGPIPE)
> >> process exit, rval = 0

> > Apparently there is a bug that causes nscd to close the connection
> > immediately but even then it is wrong that this terminates the calling
> > program with SIGPIPE.

> > The below patch prevents the SIGPIPE but cannot revive the connection to
> > nscd. This may cause numeric UIDs in top or increase the load on the
> > directory server. It is compile tested only.
> [...]

> The patch seems to fix the issue in a world after r247804. I don't see
> numeric UIDs in top but without the patch top crashes with SIGPIPE a lot
> less frequently than sudo or make install (in base/ports) for me.

> In my case shutting down nscd helped, too. Compared to stock
> nsswitch.conf I only have "cache" added.

Can you find what causes nscd to close the connection quickly, such as
using ktrace?

-- 
Jilles Tjoelker