Skip to contentSkip to conversation

src: add set optimization options

RFC PATCH Superseded Arturo Borrero Gonzalez 3 participants 5 messages Aug 13, 2014 – Sep 12, 2014 netfilter

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 src: add set optimization options

This patch adds options to choose set optimization mechanisms.

The syntax is one of:

nft add set filter set1 { type ipv4_addr size 1024 ; }
nft add set filter set1 { type ipv4_addr policy memory ; } nft add set filter set1 { type ipv4_addr policy performance ; } nft add set filter set1 { type ipv4_addr policy memory size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy memory ; } nft add set filter set1 { type ipv4_addr policy performance size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy performance ; }

Also valid for maps:

nft add map filter map1 { type ipv4_addr : verdict policy performace ; } [...]

This is the output format, which can be imported later with `nft -f':

table filter {
	set set1 {
		type ipv4_addr policy memory size 1024
	}
}

---

This is my proposal for set internal mechanism selection in nft.

My idea is: given the kernel uses optional arguments to choose the set mechanism, in userspace the configuration should be also optional.

The patch is not fully tested, there are still some issues; It seems that the kernel is lacking some details.
For example, it doesn't dump back to userspace the policy configuration.

In my opinion, we should be able to inform userspace of the configuration, in a way that userspace can differentiate between default values and fixed ones. For example, NFT_SET_POL_PERFORMANCE seems to be the default in kernel, but we don't want every set to be printed with "policy performance".

Please, feel free to comment.

 include/linux/netfilter/nf_tables.h |   29 +++++++++++++++++++++
 include/rule.h                      |    8 ++++++
 src/netlink.c                       |   19 ++++++++++++++
 src/parser.y                        |   49 ++++++++++++++++++++++++++++++++++-
 src/rule.c                          |   21 +++++++++++++++
 src/scanner.l                       |    5 ++++
 6 files changed, 129 insertions(+), 2 deletions(-)
include/linux/netfilter/nf_tables.h +29 -0
index a5f8ec0..fce770c 100644
@@ -209,6 +209,29 @@ enum nft_set_flags {
 };
 
 /**
+ * enum nft_set_policies - set selection policy
+ *
+ * @NFT_SET_POL_PERFORMANCE: prefer high performance over low memory use
+ * @NFT_SET_POL_MEMORY: prefer low memory use over high performance
+ */
+enum nft_set_policies {
+	NFT_SET_POL_PERFORMANCE,
+	NFT_SET_POL_MEMORY,
+};
+
+/**
+ * enum nft_set_desc_attributes - set element description
+ *
+ * @NFTA_SET_DESC_SIZE: number of elements in set (NLA_U32)
+ */
+enum nft_set_desc_attributes {
+	NFTA_SET_DESC_UNSPEC,
+	NFTA_SET_DESC_SIZE,
+	__NFTA_SET_DESC_MAX
+};
+#define NFTA_SET_DESC_MAX	(__NFTA_SET_DESC_MAX - 1)
+
+/**
  * enum nft_set_attributes - nf_tables set netlink attributes
  *
  * @NFTA_SET_TABLE: table name (NLA_STRING)
@@ -218,6 +241,9 @@ enum nft_set_flags {
  * @NFTA_SET_KEY_LEN: key data length (NLA_U32)
  * @NFTA_SET_DATA_TYPE: mapping data type (NLA_U32)
  * @NFTA_SET_DATA_LEN: mapping data length (NLA_U32)
+ * @NFTA_SET_POLICY: selection policy (NLA_U32)
+ * @NFTA_SET_DESC: set description (NLA_NESTED)
+ * @NFTA_SET_ID: uniquely identifies a set in a transaction (NLA_U32)
  */
 enum nft_set_attributes {
 	NFTA_SET_UNSPEC,
@@ -228,6 +254,9 @@ enum nft_set_attributes {
 	NFTA_SET_KEY_LEN,
 	NFTA_SET_DATA_TYPE,
 	NFTA_SET_DATA_LEN,
+	NFTA_SET_POLICY,
+	NFTA_SET_DESC,
+	NFTA_SET_ID,
 	__NFTA_SET_MAX
 };
 #define NFTA_SET_MAX		(__NFTA_SET_MAX - 1)
include/rule.h +8 -0
index db91406..646e634 100644
@@ -180,6 +180,7 @@ enum set_flags {
  * @datatype:	mapping data type
  * @datalen:	mapping data len
  * @init:	initializer
+ * @mechanism:	internal mechanism
  */
 struct set {
 	struct list_head	list;
@@ -192,6 +193,13 @@ struct set {
 	const struct datatype	*datatype;
 	unsigned int		datalen;
 	struct expr		*init;
+	struct {
+		uint32_t	flags;
+		uint32_t	policy;
+		struct {
+			uint32_t	size;
+		} desc;
+	} mechanism;
 };
 
 extern struct set *set_alloc(const struct location *loc);
src/netlink.c +19 -0
index e149215..e8ae324 100644
@@ -1044,6 +1044,17 @@ static struct set *netlink_delinearize_set(struct netlink_ctx *ctx,
 		set->datalen = data_len * BITS_PER_BYTE;
 	}
 
+	if (nft_set_attr_is_set(nls, NFT_SET_ATTR_POLICY)) {
+		set->mechanism.policy = nft_set_attr_get_u32(nls, NFT_SET_ATTR_POLICY);
+		set->mechanism.flags |= (1 << NFT_SET_ATTR_POLICY);
+	}
+
+	if (nft_set_attr_is_set(nls, NFT_SET_ATTR_DESC_SIZE)) {
+		set->mechanism.desc.size = nft_set_attr_get_u32(nls,
+						     NFT_SET_ATTR_DESC_SIZE);
+		set->mechanism.flags |= (1 << NFT_SET_ATTR_DESC_SIZE);
+	}
+
 	return set;
 }
 
@@ -1102,6 +1113,14 @@ static int netlink_add_set_batch(struct netlink_ctx *ctx,
 	}
 	set->handle.set_id = ++set_id;
 	nft_set_attr_set_u32(nls, NFT_SET_ATTR_ID, set->handle.set_id);
+
+	if (set->mechanism.flags & (1 << NFT_SET_ATTR_POLICY))
+		nft_set_attr_set_u32(nls, NFT_SET_ATTR_POLICY,
+				     set->mechanism.policy);
+	if (set->mechanism.flags & (1 << NFT_SET_ATTR_DESC_SIZE))
+		nft_set_attr_set_u32(nls, NFT_SET_ATTR_DESC_SIZE,
+				     set->mechanism.desc.size);
+
 	netlink_dump_set(nls);
 
 	err = mnl_nft_set_batch_add(nf_sock, nls, NLM_F_EXCL, ctx->seqnum);
src/parser.y +47 -2

91 lines not on this page · raw message

src/rule.c +21 -0
index 1e54526..9195a01 100644
@@ -75,6 +75,7 @@ void set_free(struct set *set)
 	if (--set->refcnt > 0)
 		return;
 	handle_free(&set->handle);
+
 	xfree(set);
 }
 
@@ -90,6 +91,7 @@ struct set *set_clone(const struct set *set)
 	newset->datatype = set->datatype;
 	newset->datalen = set->datalen;
 	newset->init = expr_clone(set->init);
+	newset->mechanism = set->mechanism;
 
 	return newset;
 }
@@ -134,6 +136,18 @@ struct print_fmt_options {
 	const char	*stmt_separator;
 };
 
+static const char *set_policy2str(uint32_t policy)
+{
+	switch (policy) {
+	case NFT_SET_POL_PERFORMANCE:
+		return "performance";
+	case NFT_SET_POL_MEMORY:
+		return "memory";
+	default:
+		return "unknown";
+	}
+}
+
 static void do_set_print(const struct set *set, struct print_fmt_options *opts)
 {
 	const char *delim = "";
@@ -153,6 +167,13 @@ static void do_set_print(const struct set *set, struct print_fmt_options *opts)
 	printf("%s%stype %s", opts->tab, opts->tab, set->keytype->name);
 	if (set->flags & SET_F_MAP)
 		printf(" : %s", set->datatype->name);
+
+	if (set->mechanism.flags & NFT_SET_ATTR_POLICY)
+		printf(" policy %s",
+		       set_policy2str(set->mechanism.policy));
+	if (set->mechanism.flags & NFT_SET_ATTR_DESC_SIZE)
+		printf(" size %d", set->mechanism.desc.size);
+
 	printf("%s", opts->stmt_separator);
 
 	if (set->flags & (SET_F_CONSTANT | SET_F_INTERVAL)) {
src/scanner.l +5 -0
index 4eec92f..846fc34 100644
@@ -267,6 +267,11 @@ addrstring	({macaddr}|{ip4addr}|{ip6addr})
 "interval"		{ return INTERVAL; }
 "elements"		{ return ELEMENTS; }
 
+"policy"		{ return POLICY; }
+"size"			{ return SIZE; }
+"performance"		{ return PERFORMANCE; }
+"memory"		{ return MEMORY; }
+
 "counter"		{ return COUNTER; }
 "packets"		{ return PACKETS; }
 "bytes"			{ return BYTES; }
#1 lore raw

Cc'ing Patrick.

Quoted 11 lines · Arturo Borrero Gonzalez

This patch adds options to choose set optimization mechanisms.

The syntax is one of:

nft add set filter set1 { type ipv4_addr size 1024 ; }
nft add set filter set1 { type ipv4_addr policy memory ; } nft add set filter set1 { type ipv4_addr policy performance ; } nft add set filter set1 { type ipv4_addr policy memory size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy memory ; } nft add set filter set1 { type ipv4_addr policy performance size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy performance ; }

@Patrick: Does this syntax look reasonable to you?

Quoted 26 lines

Also valid for maps:

nft add map filter map1 { type ipv4_addr : verdict policy performace ; } [...]

This is the output format, which can be imported later with `nft -f':

table filter {
	set set1 {
		type ipv4_addr policy memory size 1024
	}
}

---

This is my proposal for set internal mechanism selection in nft.

My idea is: given the kernel uses optional arguments to choose the set mechanism, in userspace the configuration should be also optional.

The patch is not fully tested, there are still some issues; It seems that the kernel is lacking some details.
For example, it doesn't dump back to userspace the policy configuration.

Quoted 6 lines

In my opinion, we should be able to inform userspace of the configuration, in a way that userspace can differentiate between default values and fixed ones. For example, NFT_SET_POL_PERFORMANCE seems to be the default in kernel, but we don't want every set to be printed with "policy performance".

Please, feel free to comment.

Just one minor comment:

     struct {                                                                 
             uint32_t        flags;                                           
             uint32_t        policy;                                          
             struct {                                                         
                     uint32_t        size;                                    
             } desc;                                                          
     } mechanism;                                                             
};

I prefer if you map 1:1 what we have in the kernel:

        struct set {
                ...
                uint32_t policy;
                struct {
                        uint32_t size;
                        uint32_t flags;
                } desc;
        };

Thanks!

#2 lore raw
Quoted 15 lines · Pablo Neira Ayuso

Cc'ing Patrick.

Quoted 11 lines · Arturo Borrero Gonzalez

This patch adds options to choose set optimization mechanisms.

The syntax is one of:

nft add set filter set1 { type ipv4_addr size 1024 ; }
nft add set filter set1 { type ipv4_addr policy memory ; } nft add set filter set1 { type ipv4_addr policy performance ; } nft add set filter set1 { type ipv4_addr policy memory size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy memory ; } nft add set filter set1 { type ipv4_addr policy performance size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy performance ; }

Quoted 1 line

@Patrick: Does this syntax look reasonable to you?

I think I would prefer having statements instead of everything combined.

Quoted 14 lines
Quoted 14 lines

Also valid for maps:

nft add map filter map1 { type ipv4_addr : verdict policy performace ; } [...]

This is the output format, which can be imported later with `nft -f':

table filter {
	set set1 {
		type ipv4_addr policy memory size 1024
	}
}
Quoted 7 lines
Quoted 7 lines

---

This is my proposal for set internal mechanism selection in nft.

My idea is: given the kernel uses optional arguments to choose the set mechanism, in userspace the configuration should be also optional.

Agreed.

Quoted 24 lines
Quoted 10 lines

The patch is not fully tested, there are still some issues; It seems that the kernel is lacking some details.
For example, it doesn't dump back to userspace the policy configuration.

In my opinion, we should be able to inform userspace of the configuration, in a way that userspace can differentiate between default values and fixed ones. For example, NFT_SET_POL_PERFORMANCE seems to be the default in kernel, but we don't want every set to be printed with "policy performance".

Please, feel free to comment.

Just one minor comment:

     struct {                                                                 
             uint32_t        flags;                                           
             uint32_t        policy;                                          
             struct {                                                         
                     uint32_t        size;                                    
             } desc;                                                          
     } mechanism;                                                             
};

I prefer if you map 1:1 what we have in the kernel:

Quoted 8 lines
        struct set {
                ...
                uint32_t policy;
                struct {
                        uint32_t size;
                        uint32_t flags;
                } desc;
        };

This is what I have in my patchset:

+struct set_desc {

+       unsigned int            size;
+       struct expr             *range;

+};

Range is another way to express the maximum size and at the same time a common prefix.

I can dust off my patch and send it later so we combine what we have.

#3 lore raw
Quoted 16 lines · Patrick McHardy
Quoted 15 lines · Pablo Neira Ayuso

Cc'ing Patrick.

Quoted 11 lines · Arturo Borrero Gonzalez

This patch adds options to choose set optimization mechanisms.

The syntax is one of:

nft add set filter set1 { type ipv4_addr size 1024 ; }
nft add set filter set1 { type ipv4_addr policy memory ; } nft add set filter set1 { type ipv4_addr policy performance ; } nft add set filter set1 { type ipv4_addr policy memory size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy memory ; } nft add set filter set1 { type ipv4_addr policy performance size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy performance ; }

Quoted 4 lines
Quoted 1 line

@Patrick: Does this syntax look reasonable to you?

I think I would prefer having statements instead of everything combined.

Could you please elaborate on this?

Quoted 64 lines
Quoted 14 lines
Quoted 14 lines

Also valid for maps:

nft add map filter map1 { type ipv4_addr : verdict policy performace ; } [...]

This is the output format, which can be imported later with `nft -f':

table filter {
    set set1 {
            type ipv4_addr policy memory size 1024
    }
}
Quoted 7 lines
Quoted 7 lines

---

This is my proposal for set internal mechanism selection in nft.

My idea is: given the kernel uses optional arguments to choose the set mechanism, in userspace the configuration should be also optional.

Agreed.

Quoted 24 lines
Quoted 10 lines

The patch is not fully tested, there are still some issues; It seems that the kernel is lacking some details.
For example, it doesn't dump back to userspace the policy configuration.

In my opinion, we should be able to inform userspace of the configuration, in a way that userspace can differentiate between default values and fixed ones. For example, NFT_SET_POL_PERFORMANCE seems to be the default in kernel, but we don't want every set to be printed with "policy performance".

Please, feel free to comment.

Just one minor comment:

     struct {
             uint32_t        flags;
             uint32_t        policy;
             struct {
                     uint32_t        size;
             } desc;
     } mechanism;
};

I prefer if you map 1:1 what we have in the kernel:

Quoted 8 lines
        struct set {
                ...
                uint32_t policy;
                struct {
                        uint32_t size;
                        uint32_t flags;
                } desc;
        };

This is what I have in my patchset:

(quoted diff) +4 -0
+struct set_desc {
+       unsigned int            size;
+       struct expr             *range;
+};
Quoted 4 lines

Range is another way to express the maximum size and at the same time a common prefix.

I can dust off my patch and send it later so we combine what we have.

Would you send that patch?

thanks.

Signature · 5 lines
Arturo Borrero González
--
To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#4 lore raw
Quoted 17 lines · Arturo Borrero Gonzalez
Quoted 16 lines · Patrick McHardy
Quoted 15 lines · Pablo Neira Ayuso

Cc'ing Patrick.

Quoted 11 lines · Arturo Borrero Gonzalez

This patch adds options to choose set optimization mechanisms.

The syntax is one of:

nft add set filter set1 { type ipv4_addr size 1024 ; }
nft add set filter set1 { type ipv4_addr policy memory ; } nft add set filter set1 { type ipv4_addr policy performance ; } nft add set filter set1 { type ipv4_addr policy memory size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy memory ; } nft add set filter set1 { type ipv4_addr policy performance size 1024 ; } nft add set filter set1 { type ipv4_addr size 1024 policy performance ; }

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

@Patrick: Does this syntax look reasonable to you?

I think I would prefer having statements instead of everything combined.

Could you please elaborate on this?

Something like this:

{
	type ipve_addr
	size 1024
	policy performance
}

or seperate by colons in case they are on a single line.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

Also valid for maps:

nft add map filter map1 { type ipv4_addr : verdict policy performace ; }

This gets a bit unreadable so { type ipv4_addr : verdict; policy performance; } seems better to me.

Quoted 7 lines
Quoted 7 lines

This is what I have in my patchset:

(quoted diff) +4 -0
+struct set_desc {
+       unsigned int            size;
+       struct expr             *range;
+};
Quoted 6 lines
Quoted 4 lines

Range is another way to express the maximum size and at the same time a common prefix.

I can dust off my patch and send it later so we combine what we have.

Would you send that patch?

Yep, just started to work on my tree five minutes ago. I'll forward-port everything and will send the patch once I'm done.