[PATCH v2 9/9] netfilter: Replace smp_read_barrier_depends() with lockless_dereference()

Recently lockless_dereference() was added which can be used in place of
hard-coding smp_read_barrier_depends(). The following PATCH makes the change.

Signed-off-by: Pranith Kumar <redacted>
---
 net/ipv4/netfilter/arp_tables.c | 3 +--
 net/ipv4/netfilter/ip_tables.c  | 3 +--
 net/ipv6/netfilter/ip6_tables.c | 3 +--
 3 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/net/ipv4/netfilter/arp_tables.c b/net/ipv4/netfilter/arp_tables.c
index f95b6f9..fc7533d 100644
--- a/net/ipv4/netfilter/arp_tables.c
+++ b/net/ipv4/netfilter/arp_tables.c

@@ -270,12 +270,11 @@ unsigned int arpt_do_table(struct sk_buff *skb,
 
 	local_bh_disable();
 	addend = xt_write_recseq_begin();
-	private = table->private;
 	/*
 	 * Ensure we load private-> members after we've fetched the base
 	 * pointer.
 	 */
-	smp_read_barrier_depends();
+	private = lockless_dereference(table->private);
 	table_base = private->entries[smp_processor_id()];
 
 	e = get_entry(table_base, private->hook_entry[hook]);

diff --git a/net/ipv4/netfilter/ip_tables.c b/net/ipv4/netfilter/ip_tables.c
index 99e810f..e0fd044 100644
--- a/net/ipv4/netfilter/ip_tables.c
+++ b/net/ipv4/netfilter/ip_tables.c

@@ -325,13 +325,12 @@ ipt_do_table(struct sk_buff *skb,
 	IP_NF_ASSERT(table->valid_hooks & (1 << hook));
 	local_bh_disable();
 	addend = xt_write_recseq_begin();
-	private = table->private;
 	cpu        = smp_processor_id();
 	/*
 	 * Ensure we load private-> members after we've fetched the base
 	 * pointer.
 	 */
-	smp_read_barrier_depends();
+	private = lockless_dereference(table->private);
 	table_base = private->entries[cpu];
 	jumpstack  = (struct ipt_entry **)private->jumpstack[cpu];
 	stackptr   = per_cpu_ptr(private->stackptr, cpu);

diff --git a/net/ipv6/netfilter/ip6_tables.c b/net/ipv6/netfilter/ip6_tables.c
index e080fbb..0459d6a 100644
--- a/net/ipv6/netfilter/ip6_tables.c
+++ b/net/ipv6/netfilter/ip6_tables.c

@@ -348,12 +348,11 @@ ip6t_do_table(struct sk_buff *skb,
 
 	local_bh_disable();
 	addend = xt_write_recseq_begin();
-	private = table->private;
 	/*
 	 * Ensure we load private-> members after we've fetched the base
 	 * pointer.
 	 */
-	smp_read_barrier_depends();
+	private = lockless_dereference(table->private);
 	cpu        = smp_processor_id();
 	table_base = private->entries[cpu];
 	jumpstack  = (struct ip6t_entry **)private->jumpstack[cpu];

-- 
1.9.1

On Fri, 2014-11-21 at 10:06 -0500, Pranith Kumar wrote:

Please carefully read the code, before and after your change, then
you'll see this change broke the code.

Problem is that a bug like that can be really hard to diagnose and fix
later, so really you have to be very careful doing these mechanical
changes.

IMO, current code+comment is better than with this
lockless_dereference() which in this particular case obfuscates the
code. more than anything.

In this case we do have a lock (sort of), so lockless_dereference() is
quite misleading.

On Fri, Nov 21, 2014 at 11:12 AM, Eric Dumazet [off-list ref] wrote:

On Fri, 2014-11-21 at 10:06 -0500, Pranith Kumar wrote:

quoted

Recently lockless_dereference() was added which can be used in place of
hard-coding smp_read_barrier_depends(). The following PATCH makes the change.

Signed-off-by: Pranith Kumar <redacted>
---
 net/ipv4/netfilter/arp_tables.c | 3 +--
 net/ipv4/netfilter/ip_tables.c  | 3 +--
 net/ipv6/netfilter/ip6_tables.c | 3 +--
 3 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/net/ipv4/netfilter/arp_tables.c b/net/ipv4/netfilter/arp_tables.c
index f95b6f9..fc7533d 100644
--- a/net/ipv4/netfilter/arp_tables.c
+++ b/net/ipv4/netfilter/arp_tables.c

@@ -270,12 +270,11 @@ unsigned int arpt_do_table(struct sk_buff *skb,

      local_bh_disable();
      addend = xt_write_recseq_begin();
-     private = table->private;
      /*
       * Ensure we load private-> members after we've fetched the base
       * pointer.
       */
-     smp_read_barrier_depends();
+     private = lockless_dereference(table->private);
      table_base = private->entries[smp_processor_id()];

Please carefully read the code, before and after your change, then
you'll see this change broke the code.

Problem is that a bug like that can be really hard to diagnose and fix
later, so really you have to be very careful doing these mechanical
changes.

IMO, current code+comment is better than with this
lockless_dereference() which in this particular case obfuscates the
code. more than anything.

In this case we do have a lock (sort of), so lockless_dereference() is
quite misleading.

Hi Eric,

Thanks for looking at this patch.

I've been scratching my head since morning trying to find out what was
so obviously wrong with this patch. Alas, I don't see what you do.

Could you point it out and show me how incompetent I am, please?

Thanks!
-- 
Pranith

On Fri, 2014-11-21 at 16:57 -0500, Pranith Kumar wrote:

Hi Eric,

Thanks for looking at this patch.

I've been scratching my head since morning trying to find out what was
so obviously wrong with this patch. Alas, I don't see what you do.

Could you point it out and show me how incompetent I am, please?

Thanks!

Well, even it the code is _not_ broken, I don't see any value with this
patch.

If I use git blame on current code, line containing
smp_read_barrier_depends() exactly points to the relevant commit [1]

After your change, it will point to some cleanup, which makes little
sense to me, considering you did not change the smp_wmb() in
xt_replace_table().

I, as a netfilter contributor would like to keep current code as is,
because it is how I feel safe with it.

We have a proliferation of interfaces, but this does not help to
understand the issues and code maintenance.

smp_read_barrier_depends() better documents the read barrier than 
lockless_dereference().

The point of having a lock or not is irrelevant here.

[1]
http://git.kernel.org/cgit/linux/kernel/git/davem/net.git/commit/?id=b416c144f46af1a30ddfa4e4319a8f077381ad63

Hi,

On 2014-11-21 16:57:00 -0500, Pranith Kumar wrote:

On Fri, Nov 21, 2014 at 11:12 AM, Eric Dumazet [off-list ref] wrote:

quoted

On Fri, 2014-11-21 at 10:06 -0500, Pranith Kumar wrote:

quoted

Recently lockless_dereference() was added which can be used in place of
hard-coding smp_read_barrier_depends(). The following PATCH makes the change.

Signed-off-by: Pranith Kumar <redacted>
---
 net/ipv4/netfilter/arp_tables.c | 3 +--
 net/ipv4/netfilter/ip_tables.c  | 3 +--
 net/ipv6/netfilter/ip6_tables.c | 3 +--
 3 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/net/ipv4/netfilter/arp_tables.c b/net/ipv4/netfilter/arp_tables.c
index f95b6f9..fc7533d 100644
--- a/net/ipv4/netfilter/arp_tables.c
+++ b/net/ipv4/netfilter/arp_tables.c

@@ -270,12 +270,11 @@ unsigned int arpt_do_table(struct sk_buff *skb,

      local_bh_disable();
      addend = xt_write_recseq_begin();
-     private = table->private;
      /*
       * Ensure we load private-> members after we've fetched the base
       * pointer.
       */
-     smp_read_barrier_depends();
+     private = lockless_dereference(table->private);
      table_base = private->entries[smp_processor_id()];

Please carefully read the code, before and after your change, then
you'll see this change broke the code.

Problem is that a bug like that can be really hard to diagnose and fix
later, so really you have to be very careful doing these mechanical
changes.

IMO, current code+comment is better than with this
lockless_dereference() which in this particular case obfuscates the
code. more than anything.

In this case we do have a lock (sort of), so lockless_dereference() is
quite misleading.

Hi Eric,

Thanks for looking at this patch.

I've been scratching my head since morning trying to find out what was
so obviously wrong with this patch. Alas, I don't see what you do.

Afaics the read_barrier_depends protected the load from private->entries[x]
earlier, not the load of table->private itself.

Greetings,

Andres Freund

On Fri, Nov 21, 2014 at 7:05 PM, Eric Dumazet [off-list ref] wrote:

On Fri, 2014-11-21 at 16:57 -0500, Pranith Kumar wrote:

quoted

Hi Eric,

Thanks for looking at this patch.

I've been scratching my head since morning trying to find out what was
so obviously wrong with this patch. Alas, I don't see what you do.

Could you point it out and show me how incompetent I am, please?

Thanks!

Well, even it the code is _not_ broken, I don't see any value with this
patch.

Phew. Not being broken itself is a win :)

If I use git blame on current code, line containing
smp_read_barrier_depends() exactly points to the relevant commit [1]

And that is an opinion I will respect. I don't want to muck the git
history where it is significant.

This effort is to eventually replace the uses of
smp_read_barrier_depends() and to use either rcu or
lockless_dereference() as documented in memory-barriers.txt.

After your change, it will point to some cleanup, which makes little
sense to me, considering you did not change the smp_wmb() in
xt_replace_table().

That does not need to change as it is fine as it is. It still pairs
with the smp_read_barrier_depends() in lockless_dereference().

I, as a netfilter contributor would like to keep current code as is,
because it is how I feel safe with it.

We have a proliferation of interfaces, but this does not help to
understand the issues and code maintenance.

smp_read_barrier_depends() better documents the read barrier than
lockless_dereference().

I think this is a matter of opinion. But in the current effort I've
seen cases where it is not clear what the barrier is actually
guaranteeing. I am glad that the current code is not one of those and
it has reasonable comments.

lockless_dereference() on the other hand makes the dependency explicit.

The point of having a lock or not is irrelevant here.

[1]
http://git.kernel.org/cgit/linux/kernel/git/davem/net.git/commit/?id=b416c144f46af1a30ddfa4e4319a8f077381ad63

Thanks!
-- 
Pranith