Re: [PATCH] Fix RCU warning in rt_cache_seq_show

Le jeudi 11 août 2011 à 19:32 -0700, Paul E. McKenney a écrit :

On Thu, Aug 11, 2011 at 06:58:21PM +0200, Eric Dumazet wrote:

quoted

Le mercredi 10 août 2011 à 10:28 +0100, Mark Rutland a écrit :

quoted

quoted

-----Original Message-----
From: Eric Dumazet [mailto:eric.dumazet@gmail.com]
Sent: 09 August 2011 18:19
To: Mark Rutland; Paul E. McKenney
Cc: netdev@vger.kernel.org; David S. Miller; Gergely Kalman
Subject: Re: [PATCH] Fix RCU warning in rt_cache_seq_show

Le mardi 09 août 2011 à 18:02 +0100, Mark Rutland a écrit :

quoted

Commit f2c31e32 ("net: fix NULL dereferences in check_peer_redir()")
added rcu protection to dst neighbour, and updated callsites for
dst_{get,set}_neighbour. Unfortunately, it missed rt_cache_seq_show.

This produces a warning on v3.1-rc1 (on a preemptible kernel, on an
ARM Vexpress A9x4):

===================================================
[ INFO: suspicious rcu_dereference_check() usage. ]
---------------------------------------------------
include/net/dst.h:91 invoked rcu_dereference_check() without

protection!

quoted

other info that might help us debug this:

rcu_scheduler_active = 1, debug_locks = 0
2 locks held by proc01/32159:

stack backtrace:
[<80014880>] (unwind_backtrace+0x0/0xf8) from [<802e5c78>]

(rt_cache_seq_show+0x18c/0x1c4)

quoted

[<802e5c78>] (rt_cache_seq_show+0x18c/0x1c4) from [<800e0c5c>]

(seq_read+0x324/0x4a4)

quoted

[<800e0c5c>] (seq_read+0x324/0x4a4) from [<8010786c>]

(proc_reg_read+0x70/0x94)

quoted

[<8010786c>] (proc_reg_read+0x70/0x94) from [<800c0ba8>]

(vfs_read+0xb0/0x144)

quoted

[<800c0ba8>] (vfs_read+0xb0/0x144) from [<800c0ea8>]

(sys_read+0x40/0x70)

quoted

[<800c0ea8>] (sys_read+0x40/0x70) from [<8000e0c0>]

(ret_fast_syscall+0x0/0x3c)

quoted

This patch adds calls to rcu_read_{lock,unlock} in rt_cache_seq_show,
protecting the dereferenced variable, and clearing the warning.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <redacted>
Cc: Gergely Kalman <redacted>
---
 net/ipv4/route.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index e3dec1c..6699ef7 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c

@@ -419,6 +419,7 @@ static int rt_cache_seq_show(struct seq_file

*seq, void *v)

quoted

 		struct neighbour *n;
 		int len;

+		rcu_read_lock();
 		n = dst_get_neighbour(&r->dst);
 		seq_printf(seq, "%s\t%08X\t%08X\t%8X\t%d\t%u\t%d\t"
 			      "%08X\t%d\t%u\t%u\t%02X\t%d\t%1d\t%08X%n",

@@ -435,6 +436,7 @@ static int rt_cache_seq_show(struct seq_file

*seq, void *v)

quoted

 			-1,
 			(n && (n->nud_state & NUD_CONNECTED)) ? 1 : 0,
 			r->rt_spec_dst, &len);
+		rcu_read_unlock();

 		seq_printf(seq, "%*s\n", 127 - len, "");
 	}

Hmm, I though rcu_read_lock_bh() (done by caller of this function) was
protecting us here.

Aha. Being a bit trigger-happy, I'd had a quick look at the functions
mentioned in the backtrace, and not looked at any possible inlining.

This being my first real exposure to RCU, I wasn't aware of the *_bh
variants. Looking at the documentation (Documentation/RCU/checklist.txt),
I think the real problem is that we should be using rcu_dereference_bh in
this case:

  > read-side critical sections are delimited by rcu_read_lock()
  > and rcu_read_unlock(), or by similar primitives such as
  > rcu_read_lock_bh() and rcu_read_unlock_bh(), in which case
  > the matching rcu_dereference() primitive must be used in order
  > to keep lockdep happy, in this case, rcu_dereference_bh().

Hmm.

I do think dst_get_neighbour() should use rcu_dereference(), because
dst->_neighbour are freed by call_rcu().

The question is : Is following construct [A] safe or not ?

{
rcu_read_lock_bh();
	/* BH are now disabled, and we are not allowed to sleep */
	...

	ptr = rcu_dereference();

This should be:

	ptr = rcu_dereference_bh();

As you say below.  Never mind!  ;-)

quoted

	...
rcu_read_unlock_bh();
}


I dont really understand why lockdep wants [B] instead :

{
rcu_read_lock_bh();
	...

	{
	rcu_read_lock();
	ptr = rcu_dereference();

Here you are protected by both RCU and RCU-bh, so you should be able
to use either rcu_dereference() or rcu_dereference_bh().  A bit
strange to use rcu_dereference_bh(), though.  Except perhaps if a
pointer to a function was passed in from the outer RCU-bh read-side
critical section or something.

quoted

	rcu_read_unlock();
	}
	...
rcu_read_unlock_bh();
}



However, I can understand the other way [C], this is really needed :

{
rcu_read_lock();
	...

	{
	rcu_read_lock_bh();
	ptr = rcu_dereference_bh();
	rcu_read_unlock_bh();
	}
	...
rcu_read_unlock();
}

I believe [A] should be allowed by lockdep.

OK, I'll bite.  Why?

Oh well, I assumed local_bh_disable() disables preemption.

It does since day-0
add_preempt_count(SOFTIRQ_DISABLE_OFFSET);

So following should be safe :

local_bh_disable();
{
ptr = rcu_dereference(...);
use(ptr);
}
local_bh_enable();

Maybe they are longterm plans to break this assumption, I dont know.