svn commit: r192094 - head/sys/kern

[Peter Holm]

On Fri, May 15, 2009 at 12:48:52PM +0300, Kostik Belousov wrote:
> On Fri, May 15, 2009 at 10:06:13AM +0200, Peter Holm wrote:
> > On Fri, May 15, 2009 at 09:02:39AM +0200, Ed Schouten wrote:
> > > Hi Kostik,
> > > 
> > > * Konstantin Belousov <kib at FreeBSD.org> wrote:
> > > > Log:
> > > >   Do not advance req->oldidx when sysctl_old_user returning an
> > > >   error due to copyout failure or short buffer.
> > > >   
> > > >   The later breaks the usermode iterators of the sysctl results that pack
> > > >   arbitrary number of variable-sized structures. Iterator expects that
> > > >   kernel filled exactly oldlen bytes, and tries to interpret half-filled
> > > >   or garbage structure at the end of the buffer. In particular,
> > > >   kinfo_getfile(3) segfaulted.
> > > >   
> > > >   Reported and tested by:	pho
> > > >   MFC after:	3 weeks
> > > 
> > > Is it possible that this change introduces a regression? Right now
> > > `pstat -t' gets stuck in an infinite loop. I've added the following
> > > printf:
> > > 
> > > | Index: pstat.c
> > > | ===================================================================
> > > | --- pstat.c	(revision 192128)
> > > | +++ pstat.c	(working copy)
> > > | @@ -263,6 +263,7 @@
> > > |  		if (errno != ENOMEM)
> > > |  			err(1, "sysctlbyname()");
> > > |  		len *= 2;
> > > | +		printf("Going to %zu\n", len);
> > > |  		if ((xttys = realloc(xttys, len)) == NULL)
> > > |  			err(1, "realloc()");
> > > |  	}
> > > 
> > > pstat on -CURRENT prints:
> > > 
> > > |       LINE   INQ  CAN  LIN  LOW  OUTQ  USE  LOW   COL  SESS  PGID STATE
> > > | Going to 0
> > > | Going to 0
> > > | Going to 0
> > > | ...
> > > 
> > > If I use the same patch on RELENG_6, I get the expected result:
> > > 
> > > |      LINE RAW CAN OUT IHIWT ILOWT OHWT LWT     COL STATE  SESS      PGID DISC
> > > | Going to 272
> > > | Going to 544
> > > | Going to 1088
> > > | Going to 2176
> > > | Going to 4352
> > > | Going to 8704
> > > |   sysmouse  0   0   0     0     0    0   0       0 -             0     0 term
> > > | ...
> > > 
> > > So the problem is that sysctl overwrites the len argument with 0, even
> > > if it returns back to userspace with ENOMEM.
> > > 
> > > I see we have two changes in sysctl. In theory it could also be related
> > > to jhb@'s changes to sysctl locking, but I suspect it's less likely.
> > > 
> > 
> > I can confirm that it is r192094 that triggers the loop.
> 
> Yes, this is what I mean when talked about a breakage.
> 
> Below is the reversal of r192094 + the change to keep the old, ugly
> behaviour of sysctl kern.proc.filedesc to return 0 on ENOMEM, but with
> oldlen chopped at the end of the last completely written struct kern_info
> instead of the middle of partially-written one.
> 
> Peter, could you, please, retest ?

Right away.

- Peter