[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20130522095839.GC3578@linux.vnet.ibm.com>
Date: Wed, 22 May 2013 02:58:39 -0700
From: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To: Eric Dumazet <eric.dumazet@...il.com>
Cc: Roman Gushchin <klamm@...dex-team.ru>,
Dipankar Sarma <dipankar@...ibm.com>, zhmurov@...dex-team.ru,
linux-kernel@...r.kernel.org, netdev@...r.kernel.org,
"David S. Miller" <davem@...emloft.net>,
Alexey Kuznetsov <kuznet@....inr.ac.ru>,
James Morris <jmorris@...ei.org>,
Hideaki YOSHIFUJI <yoshfuji@...ux-ipv6.org>,
Patrick McHardy <kaber@...sh.net>
Subject: Re: [PATCH v2] rcu: fix a race in hlist_nulls_for_each_entry_rcu
macro
On Tue, May 21, 2013 at 07:01:20PM -0700, Eric Dumazet wrote:
> On Tue, 2013-05-21 at 22:12 +0400, Roman Gushchin wrote:
> > > If other rcu accessors have the same problem, a more complete patch is
> > > needed.
> >
> > [PATCH] rcu: fix a race in rcu lists traverse macros
> >
> > Some network functions (udp4_lib_lookup2(), for instance) use the
> > rcu lists traverse macros (hlist_nulls_for_each_entry_rcu, for instance)
> > in a way that assumes restarting of a loop. In this case, it is strictly
> > necessary to reread the head->first value from the memory before each scan.
> > Without additional hints, gcc caches this value in a register. In this case,
> > if a cached node is moved to another chain during the scan, we can loop
> > forever getting wrong nulls values and restarting the loop uninterruptedly.
> >
> > Signed-off-by: Roman Gushchin <klamm@...dex-team.ru>
> > Reported-by: Boris Zhmurov <zhmurov@...dex-team.ru>
> > ---
> > include/linux/compiler.h | 6 ++++++
> > include/linux/rculist.h | 6 ++++--
> > include/linux/rculist_nulls.h | 3 ++-
> > 3 files changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/compiler.h b/include/linux/compiler.h
> > index 92669cd..0e05d7c 100644
> > --- a/include/linux/compiler.h
> > +++ b/include/linux/compiler.h
> > @@ -351,6 +351,12 @@ void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect);
> > */
> > #define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
> >
> > +/*
> > + * Prevent the compiler from optimizing accesses to structure member.
> > + */
> > +#define GET_STRUCT_FIELD_VOLATILE(struct_ptr, field) \
> > + (((volatile typeof(*struct_ptr) *)struct_ptr)->field)
> > +
> > /* Ignore/forbid kprobes attach on very low level functions marked by this attribute: */
> > #ifdef CONFIG_KPROBES
> > # define __kprobes __attribute__((__section__(".kprobes.text")))
> > diff --git a/include/linux/rculist.h b/include/linux/rculist.h
> > index 8089e35..6c3eea7 100644
> > --- a/include/linux/rculist.h
> > +++ b/include/linux/rculist.h
> > @@ -282,7 +282,8 @@ static inline void list_splice_init_rcu(struct list_head *list,
> > * as long as the traversal is guarded by rcu_read_lock().
> > */
> > #define list_for_each_entry_rcu(pos, head, member) \
> > - for (pos = list_entry_rcu((head)->next, typeof(*pos), member); \
> > + for (pos = list_entry_rcu(GET_STRUCT_FIELD_VOLATILE(head, next), \
> > + typeof(*pos), member); \
> > &pos->member != (head); \
> > pos = list_entry_rcu(pos->member.next, typeof(*pos), member))
> >
> > @@ -348,7 +349,8 @@ static inline void hlist_replace_rcu(struct hlist_node *old,
> > /*
> > * return the first or the next element in an RCU protected hlist
> > */
> > -#define hlist_first_rcu(head) (*((struct hlist_node __rcu **)(&(head)->first)))
> > +#define hlist_first_rcu(head) (*((struct hlist_node __rcu **) \
> > + (&GET_STRUCT_FIELD_VOLATILE(head, first))))
> > #define hlist_next_rcu(node) (*((struct hlist_node __rcu **)(&(node)->next)))
> > #define hlist_pprev_rcu(node) (*((struct hlist_node __rcu **)((node)->pprev)))
> >
> > diff --git a/include/linux/rculist_nulls.h b/include/linux/rculist_nulls.h
> > index 2ae1371..b04fd0a 100644
> > --- a/include/linux/rculist_nulls.h
> > +++ b/include/linux/rculist_nulls.h
> > @@ -38,7 +38,8 @@ static inline void hlist_nulls_del_init_rcu(struct hlist_nulls_node *n)
> > }
> >
> > #define hlist_nulls_first_rcu(head) \
> > - (*((struct hlist_nulls_node __rcu __force **)&(head)->first))
> > + (*((struct hlist_nulls_node __rcu __force **) \
> > + &(GET_STRUCT_FIELD_VOLATILE(head, first))))
> >
> > #define hlist_nulls_next_rcu(node) \
> > (*((struct hlist_nulls_node __rcu __force **)&(node)->next))
Now that I am more awake...
The RCU list macros assume that the list header is either statically
allocated (in which case no ACCESS_ONCE() or whatever is needed) or
that the caller did whatever was necessary to protect the list header,
whether that be holding the right lock, using rcu_dereference() when
traversing the pointer to the list header, or whatever.
Maybe I need to document this assumption...
Thanx, Paul
> Please use ACCESS_ONCE(), which is the standard way to deal with this,
> and remove the rcu_dereference_raw() in
> hlist_nulls_for_each_entry_rcu()
>
> something like : (for the nulls part only)
>
> diff --git a/include/linux/rculist_nulls.h b/include/linux/rculist_nulls.h
> index 2ae1371..1d485d3 100644
> --- a/include/linux/rculist_nulls.h
> +++ b/include/linux/rculist_nulls.h
> @@ -40,6 +40,9 @@ static inline void hlist_nulls_del_init_rcu(struct hlist_nulls_node *n)
> #define hlist_nulls_first_rcu(head) \
> (*((struct hlist_nulls_node __rcu __force **)&(head)->first))
>
> +#define hlist_nulls_first_rcu_once(head) \
> + ACCESS_ONCE(*((struct hlist_nulls_node __rcu __force **)&(head)->first))
> +
> #define hlist_nulls_next_rcu(node) \
> (*((struct hlist_nulls_node __rcu __force **)&(node)->next))
>
> @@ -107,7 +110,7 @@ static inline void hlist_nulls_add_head_rcu(struct hlist_nulls_node *n,
> *
> */
> #define hlist_nulls_for_each_entry_rcu(tpos, pos, head, member) \
> - for (pos = rcu_dereference_raw(hlist_nulls_first_rcu(head)); \
> + for (pos = hlist_nulls_first_rcu_once(head); \
> (!is_a_nulls(pos)) && \
> ({ tpos = hlist_nulls_entry(pos, typeof(*tpos), member); 1; }); \
> pos = rcu_dereference_raw(hlist_nulls_next_rcu(pos)))
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists