Skip to main content

Browse the archive

Revision b97021f85517552ea8a0d2c1680c1ee4beab6d14 authored by Benjamin Herrenschmidt on 15 November 2011, 17:11:27 UTC, committed by Benjamin Herrenschmidt on 17 November 2011, 05:26:07 UTC 
powerpc: Fix atomic_xxx_return barrier semantics
 
The Documentation/memory-barriers.txt document requires that atomic
operations that return a value act as a memory barrier both before
and after the actual atomic operation.

Our current implementation doesn't guarantee this. More specifically,
while a load following the isync can not be issued before stwcx. has
completed, that completion doesn't architecturally means that the
result of stwcx. is visible to other processors (or any previous stores
for that matter) (typically, the other processors L1 caches can still
hold the old value).

This has caused an actual crash in RCU torture testing on Power 7

This fixes it by changing those atomic ops to use new macros instead
of RELEASE/ACQUIRE barriers, called ATOMIC_ENTRY and ATMOIC_EXIT barriers,
which are then defined respectively to lwsync and sync.

I haven't had a chance to measure the performance impact (or rather
what I measured with kernel compiles is in the noise, I yet have to
find a more precise benchmark)

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
1 parent a9a8f77
Raw File
ipcns_notifier.c 
/*
 * linux/ipc/ipcns_notifier.c
 * Copyright (C) 2007 BULL SA. Nadia Derbey
 *
 * Notification mechanism for ipc namespaces:
 * The callback routine registered in the memory chain invokes the ipcns
 * notifier chain with the IPCNS_MEMCHANGED event.
 * Each callback routine registered in the ipcns namespace recomputes msgmni
 * for the owning namespace.
 */

#include <linux/msg.h>
#include <linux/rcupdate.h>
#include <linux/notifier.h>
#include <linux/nsproxy.h>
#include <linux/ipc_namespace.h>

#include "util.h"



static BLOCKING_NOTIFIER_HEAD(ipcns_chain);


static int ipcns_callback(struct notifier_block *self,
				unsigned long action, void *arg)
{
	struct ipc_namespace *ns;

	switch (action) {
	case IPCNS_MEMCHANGED:   /* amount of lowmem has changed */
	case IPCNS_CREATED:
	case IPCNS_REMOVED:
		/*
		 * It's time to recompute msgmni
		 */
		ns = container_of(self, struct ipc_namespace, ipcns_nb);
		/*
		 * No need to get a reference on the ns: the 1st job of
		 * free_ipc_ns() is to unregister the callback routine.
		 * blocking_notifier_chain_unregister takes the wr lock to do
		 * it.
		 * When this callback routine is called the rd lock is held by
		 * blocking_notifier_call_chain.
		 * So the ipc ns cannot be freed while we are here.
		 */
		recompute_msgmni(ns);
		break;
	default:
		break;
	}

	return NOTIFY_OK;
}

int register_ipcns_notifier(struct ipc_namespace *ns)
{
	int rc;

	memset(&ns->ipcns_nb, 0, sizeof(ns->ipcns_nb));
	ns->ipcns_nb.notifier_call = ipcns_callback;
	ns->ipcns_nb.priority = IPCNS_CALLBACK_PRI;
	rc = blocking_notifier_chain_register(&ipcns_chain, &ns->ipcns_nb);
	if (!rc)
		ns->auto_msgmni = 1;
	return rc;
}

int cond_register_ipcns_notifier(struct ipc_namespace *ns)
{
	int rc;

	memset(&ns->ipcns_nb, 0, sizeof(ns->ipcns_nb));
	ns->ipcns_nb.notifier_call = ipcns_callback;
	ns->ipcns_nb.priority = IPCNS_CALLBACK_PRI;
	rc = blocking_notifier_chain_cond_register(&ipcns_chain,
							&ns->ipcns_nb);
	if (!rc)
		ns->auto_msgmni = 1;
	return rc;
}

void unregister_ipcns_notifier(struct ipc_namespace *ns)
{
	blocking_notifier_chain_unregister(&ipcns_chain, &ns->ipcns_nb);
	ns->auto_msgmni = 0;
}

int ipcns_notify(unsigned long val)
{
	return blocking_notifier_call_chain(&ipcns_chain, val, NULL);
}
back to top