Re: [PATCH] netfilter: iptables: lock free counters, PREEMPT_RCU=y fix

[Ingo Molnar <mingo@elte.hu>]

* Stephen Hemminger <shemminger@vyatta.com> wrote:

> On Thu, 2 Apr 2009 22:12:45 +0200
> Ingo Molnar <mingo@elte.hu> wrote:
> 
> > 
> > Impact: fix log spam under CONFIG_DEBUG_PREEMPT=y
> > 
> > This recent commit:
> > 
> >    7845447: netfilter: iptables: lock free counters
> > 
> > Converted a couple of netfilter codepaths from read_lock() critical 
> > sections to lockless rcu_read_lock(). What it forgot about is that 
> > under CONFIG_PREEMPT=y and CONFIG_PREEMPT_RCU=y these sections can 
> > be preempted.
> > 
> > Under CONFIG_DEBUG_PREEMPT=y this produces such warnings:
> > 
> > BUG: using smp_processor_id() in preemptible [00000000] code: ssh/9115
> > caller is ipt_do_table+0xc8/0x559
> > Pid: 9115, comm: ssh Tainted: G        W  2.6.29-tip-08646-g45ef7c3-dirty #26231
> > Call Trace:
> >  [<c0c0dacf>] ? printk+0x14/0x16
> >  [<c048c2e6>] debug_smp_processor_id+0xa6/0xbc
> >  [<c0adbd03>] ipt_do_table+0xc8/0x559
> >  [<c0c10277>] ? _read_unlock+0x3d/0x49
> >  [<c0ad68c0>] ? fn_hash_lookup+0x94/0xa0
> >  [<c0ad330e>] ? __inet_dev_addr_type+0x56/0x8d
> >  [<c0a87b02>] ? neigh_lookup+0xe5/0x108
> >  [<c0adc2bc>] ipt_local_hook+0x40/0x50
> >  [<c0a93e57>] nf_iterate+0x34/0x80
> >  [<c0aad4b8>] ? dst_output+0x0/0x10
> >  [<c0a93eea>] nf_hook_slow+0x47/0xa4
> >  [<c0aad4b8>] ? dst_output+0x0/0x10
> >  [<c0aaec04>] __ip_local_out+0x78/0x7f
> >  [<c0aad4b8>] ? dst_output+0x0/0x10
> >  [<c0aaec1b>] ip_local_out+0x10/0x20
> >  [<c0aaf416>] ip_queue_xmit+0x2bc/0x332
> >  [<c0aa8fdf>] ? __ip_route_output_key+0x112/0x77b
> >  [<c01397fe>] ? local_bh_enable+0x10/0x12
> >  [<c0abf0e5>] ? tcp_connect+0x32a/0x3bb
> >  [<c0ab15a7>] ? __inet_hash_nolisten+0x97/0xaf
> >  [<c0a78dde>] ? __copy_skb_header+0xe/0x13a
> >  [<c0abf0e5>] ? tcp_connect+0x32a/0x3bb
> >  [<c0abe955>] ? tcp_transmit_skb+0x5a5/0x61c
> >  [<c0abe995>] tcp_transmit_skb+0x5e5/0x61c
> >  [<c0a7b6c0>] ? __alloc_skb+0x54/0x120
> >  [<c0abefca>] ? tcp_connect+0x20f/0x3bb
> >  [<c0abf0e5>] tcp_connect+0x32a/0x3bb
> >  [<c0ac4b68>] tcp_v4_connect+0x466/0x4be
> >  [<c0acf380>] inet_stream_connect+0x8f/0x212
> >  [<c018f4e6>] ? might_fault+0x75/0x77
> >  [<c047f198>] ? copy_from_user+0x2f/0x117
> > BUG: using smp_processor_id() in preemptible [00000000] code: ssh/9114
> > 
> > Since it appears that the tables are RCU freed, and there are no 
> > non-preempt assumptions in the code, the using of 
> > raw_smp_processor_id() is safe.
> > 
> > [ I also audited all of net/netfilter/*.c for smp_processor_id() use,
> >   and fixed all places that used them unsafely. ]
> > 
> > Signed-off-by: Ingo Molnar <mingo@elte.hu>
> > ---
> >  net/ipv4/netfilter/arp_tables.c |    2 +-
> >  net/ipv4/netfilter/ip_tables.c  |    4 ++--
> >  2 files changed, 3 insertions(+), 3 deletions(-)
> > 
> > diff --git a/net/ipv4/netfilter/arp_tables.c b/net/ipv4/netfilter/arp_tables.c
> > index 35c5f6a..30baf3e 100644
> > --- a/net/ipv4/netfilter/arp_tables.c
> > +++ b/net/ipv4/netfilter/arp_tables.c
> > @@ -255,7 +255,7 @@ unsigned int arpt_do_table(struct sk_buff *skb,
> >  
> >  	rcu_read_lock();
> >  	private = rcu_dereference(table->private);
> > -	table_base = rcu_dereference(private->entries[smp_processor_id()]);
> > +	table_base = rcu_dereference(private->entries[raw_smp_processor_id()]);
> >  
> >  	e = get_entry(table_base, private->hook_entry[hook]);
> >  	back = get_entry(table_base, private->underflow[hook]);
> > diff --git a/net/ipv4/netfilter/ip_tables.c b/net/ipv4/netfilter/ip_tables.c
> > index 82ee7c9..eff124e 100644
> > --- a/net/ipv4/netfilter/ip_tables.c
> > +++ b/net/ipv4/netfilter/ip_tables.c
> > @@ -280,7 +280,7 @@ static void trace_packet(struct sk_buff *skb,
> >  	char *hookname, *chainname, *comment;
> >  	unsigned int rulenum = 0;
> >  
> > -	table_base = (void *)private->entries[smp_processor_id()];
> > +	table_base = (void *)private->entries[raw_smp_processor_id()];
> >  	root = get_entry(table_base, private->hook_entry[hook]);
> >  
> >  	hookname = chainname = (char *)hooknames[hook];
> > @@ -341,7 +341,7 @@ ipt_do_table(struct sk_buff *skb,
> >  
> >  	rcu_read_lock();
> >  	private = rcu_dereference(table->private);
> > -	table_base = rcu_dereference(private->entries[smp_processor_id()]);
> > +	table_base = rcu_dereference(private->entries[raw_smp_processor_id()]);
> >  
> >  	e = get_entry(table_base, private->hook_entry[hook]);
> >  
> 
> NAK. The rcu_read_lock() needs to be rcu_read_lock_bh() otherwise RCU
> could corrupt the referenceses.

Indeed - somehow i missed that it originally used read_lock_bh().

I was wondering about that already: how could get_entry() and 
nf-iterator work without worrying about softnet-rx processing, it 
just looked too involved to be trivially preemptible.

Thanks, i'll use Eric's patch.

	Ingo