Date: Sun, 11 Nov 2001 00:36:41 GMT From: Tor.Egge@cvsup.no.freebsd.org To: ohartman@klima.physik.uni-mainz.de Cc: freebsd-stable@FreeBSD.ORG Subject: Re: FBSD4.4-STABLE SMP broken! Message-ID: <20011111003641B.tegge@cvsup.no.freebsd.org> In-Reply-To: <20011110172005.J999-100000@klima.physik.uni-mainz.de> References: <20011110172005.J999-100000@klima.physik.uni-mainz.de>
next in thread | previous in thread | raw e-mail | index | archive | help
[-- Attachment #1 --]
> I did a cvsupdate today, the target machine was running
> stable for the last four days.
>
> After the update and make world today, the system run a few minutes and then
> get stuck ... no keyboard input, nothing, no network response.
Sleeping while holding a simplelock can be fatal on an SMP machine.
If the other CPU tries to obtain the simplelock then you get a hang.
If the same CPU tries to obtain the simplelock then you get a panic.
One example is vinvalbuf() holding the vnode interlock while calling
vm_object_page_remove() (which might block), cf. PR 26224.
For 4.4-STABLE, it might make sense to just define some simplelock
operations to nops also under SMP. All relevant code is already
serialized by mp_lock and further lock pushdown is unlikely to occur
on that branch. The only effect of these simplelock operations under
4.4-STABLE SMP is a hang or panic when the locking protocol is
violated.
The enclosed patch will break 3rd party device drivers that use fast
interrupts and simple_lock() to serialize access to shared resources.
I don't know about any such device drivers.
You'll need to add
options SIMPLELOCK_NULL
to the relevant kernel config file and recompile the kernel for the
patch to have any effect.
- Tor Egge
[-- Attachment #2 --]
Index: sys/conf/options
===================================================================
RCS file: /home/ncvs/src/sys/conf/options,v
retrieving revision 1.191.2.37
diff -u -r1.191.2.37 options
--- sys/conf/options 3 Nov 2001 01:41:07 -0000 1.191.2.37
+++ sys/conf/options 10 Nov 2001 23:29:19 -0000
@@ -361,7 +363,8 @@
INVARIANT_SUPPORT opt_global.h
INVARIANTS opt_global.h
SIMPLELOCK_DEBUG opt_global.h
+SIMPLELOCK_NULL opt_global.h
VFS_BIO_DEBUG opt_global.h
# These are VM related options
VM_KMEM_SIZE opt_vm.h
Index: sys/i386/include/lock.h
===================================================================
RCS file: /home/ncvs/src/sys/i386/include/Attic/lock.h,v
retrieving revision 1.11.2.2
diff -u -r1.11.2.2 lock.h
--- sys/i386/include/lock.h 30 Sep 2000 02:49:34 -0000 1.11.2.2
+++ sys/i386/include/lock.h 10 Nov 2001 23:29:19 -0000
@@ -170,6 +170,18 @@
extern struct simplelock mcount_lock;
#if !defined(SIMPLELOCK_DEBUG) && MAXCPU > 1
+
+#ifdef SIMPLELOCK_NULL
+
+/* On SMP systems, mp_lock is used for the serialization */
+
+#define NULL_SIMPLELOCKS
+#define simple_lock_init(alp)
+#define simple_lock(alp)
+#define simple_lock_try(alp) (1) /* always succeeds */
+#define simple_unlock(alp)
+
+#else
/*
* This set of defines turns on the real functions in i386/isa/apic_ipl.s.
*/
@@ -177,6 +189,8 @@
#define simple_lock(alp) s_lock(alp)
#define simple_lock_try(alp) s_lock_try(alp)
#define simple_unlock(alp) s_unlock(alp)
+
+#endif
#endif /* !SIMPLELOCK_DEBUG && MAXCPU > 1 */
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20011111003641B.tegge>