[src/trunk]: src/sys/net Add a little comment on how bpf can be made unloadab...

[pooka <pooka%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/b426a3f3a54d
branches:  trunk
changeset: 753950:b426a3f3a54d
user:      pooka <pooka%NetBSD.org@localhost>
date:      Wed Apr 14 13:31:33 2010 +0000

description:
Add a little comment on how bpf can be made unloadable, per pointer from ad.

diffstat:

 sys/net/bpf.c |  24 ++++++++++++++++++++----
 1 files changed, 20 insertions(+), 4 deletions(-)

diffs (45 lines):

diff -r bae5b230ea8d -r b426a3f3a54d sys/net/bpf.c
--- a/sys/net/bpf.c     Wed Apr 14 13:07:19 2010 +0000
+++ b/sys/net/bpf.c     Wed Apr 14 13:31:33 2010 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: bpf.c,v 1.157 2010/04/05 07:22:22 joerg Exp $  */
+/*     $NetBSD: bpf.c,v 1.158 2010/04/14 13:31:33 pooka Exp $  */
 
 /*
  * Copyright (c) 1990, 1991, 1993
@@ -39,7 +39,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: bpf.c,v 1.157 2010/04/05 07:22:22 joerg Exp $");
+__KERNEL_RCSID(0, "$NetBSD: bpf.c,v 1.158 2010/04/14 13:31:33 pooka Exp $");
 
 #if defined(_KERNEL_OPT)
 #include "opt_bpf.h"
@@ -1976,8 +1976,24 @@
 
        case MODULE_CMD_FINI:
                /*
-                * bpf_ops is not (yet) referenced in the callers before
-                * attach.  maybe other issues too.  "safety first".
+                * While there is no reference counting for bpf callers,
+                * unload could at least in theory be done similarly to 
+                * system call disestablishment.  This should even be
+                * a little simpler:
+                * 
+                * 1) replace op vector with stubs
+                * 2) post update to all cpus with xc
+                * 3) check that nobody is in bpf anymore
+                *    (it's doubtful we'd want something like l_sysent,
+                *     but we could do something like *signed* percpu
+                *     counters.  if the sum is 0, we're good).
+                * 4) if fail, unroll changes
+                *
+                * NOTE: change won't be atomic to the outside.  some
+                * packets may be not captured even if unload is
+                * not succesful.  I think packet capture not working
+                * is a perfectly logical consequence of trying to
+                * disable packet capture.
                 */
                error = EOPNOTSUPP;
                break;