diff mbox series

[v2] ethdev: fix expand RSS flows

Message ID f48a39a6f2718e82993fe97ec3bd3525ad2ce847.1572960444.git.jackmin@mellanox.com (mailing list archive)
State Accepted, archived
Delegated to: Ferruh Yigit
Headers
Series [v2] ethdev: fix expand RSS flows |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/Intel-compilation fail Compilation issues
ci/iol-intel-Performance success Performance Testing PASS
ci/iol-compilation success Compile Testing PASS
ci/iol-mellanox-Performance success Performance Testing PASS

Commit Message

Xiaoyu Min Nov. 5, 2019, 1:42 p.m. UTC 
  rte_flow_expand_rss expands rte_flow item list based on the RSS
types. In another word, some additional rules are added if the user
specified items are not complete enough according to the RSS type,
for example:

  ... pattern eth / end actions rss type tcp end ...

User only provides item eth but want to do RSS on tcp traffic.
The pattern is not complete enough to filter TCP traffic only.
This will be a problem for some HWs.
So some PMDs use rte_flow_expand_rss to expand above user provided
flow to:

  ... pattern eth / end actions rss types tcp
  ... pattern eth / ipv4 / tcp / end actions rss types tcp ...
  ... pattern eth / ipv6 / tcp / end actions rss types tcp ...

in order to filter TCP traffic only and do RSS correctly.

However the current expansion cannot handle pattern as below, which
provides ethertype or ip next proto instead of providing an item:

  ... pattern eth type is 0x86DD / end actions rss types tcp ...

rte_flow_expand_rss will expand above flow to:

  ... pattern eth type is 0x86DD / ipv4 / tcp end ...

which has conflicting values: 0x86DD vs. ipv4 and some HWs will refuse
to create flow.

This patch will fix above by checking the last item's spec and to
expand RSS flows correctly.

Currently only support to complete item list based on ether type or ip
next proto.

Fixes: 4ed05fcd441b ("ethdev: add flow API to expand RSS flows")
Cc: stable@dpdk.org

Signed-off-by: Xiaoyu Min <jackmin@mellanox.com>
---
 v2:
   * more clear commit message 
---
 lib/librte_ethdev/rte_flow.c | 132 +++++++++++++++++++++++++++++++++--
 1 file changed, 127 insertions(+), 5 deletions(-)

Comments

Ori Kam Nov. 6, 2019, 10:12 a.m. UTC | #1 
> -----Original Message-----
> From: Xiaoyu Min <jackmin@mellanox.com>
> Sent: Tuesday, November 5, 2019 3:43 PM
> To: Ori Kam <orika@mellanox.com>; Slava Ovsiienko
> <viacheslavo@mellanox.com>; Thomas Monjalon <thomas@monjalon.net>;
> Adrien Mazarguil <adrien.mazarguil@6wind.com>; Ferruh Yigit
> <ferruh.yigit@intel.com>; Andrew Rybchenko <arybchenko@solarflare.com>
> Cc: dev@dpdk.org; stable@dpdk.org
> Subject: [PATCH v2] ethdev: fix expand RSS flows
> 
> rte_flow_expand_rss expands rte_flow item list based on the RSS
> types. In another word, some additional rules are added if the user
> specified items are not complete enough according to the RSS type,
> for example:
> 
>   ... pattern eth / end actions rss type tcp end ...
> 
> User only provides item eth but want to do RSS on tcp traffic.
> The pattern is not complete enough to filter TCP traffic only.
> This will be a problem for some HWs.
> So some PMDs use rte_flow_expand_rss to expand above user provided
> flow to:
> 
>   ... pattern eth / end actions rss types tcp
>   ... pattern eth / ipv4 / tcp / end actions rss types tcp ...
>   ... pattern eth / ipv6 / tcp / end actions rss types tcp ...
> 
> in order to filter TCP traffic only and do RSS correctly.
> 
> However the current expansion cannot handle pattern as below, which
> provides ethertype or ip next proto instead of providing an item:
> 
>   ... pattern eth type is 0x86DD / end actions rss types tcp ...
> 
> rte_flow_expand_rss will expand above flow to:
> 
>   ... pattern eth type is 0x86DD / ipv4 / tcp end ...
> 
> which has conflicting values: 0x86DD vs. ipv4 and some HWs will refuse
> to create flow.
> 
> This patch will fix above by checking the last item's spec and to
> expand RSS flows correctly.
> 
> Currently only support to complete item list based on ether type or ip
> next proto.
> 
> Fixes: 4ed05fcd441b ("ethdev: add flow API to expand RSS flows")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Xiaoyu Min <jackmin@mellanox.com>
> ---
>  v2:
>    * more clear commit message
> ---

Acked-by: Ori Kam <orika@mellanox.com>
Ferruh Yigit Nov. 7, 2019, 5:02 p.m. UTC | #2 
On 11/6/2019 10:12 AM, Ori Kam wrote:
> 
> 
>> -----Original Message-----
>> From: Xiaoyu Min <jackmin@mellanox.com>
>> Sent: Tuesday, November 5, 2019 3:43 PM
>> To: Ori Kam <orika@mellanox.com>; Slava Ovsiienko
>> <viacheslavo@mellanox.com>; Thomas Monjalon <thomas@monjalon.net>;
>> Adrien Mazarguil <adrien.mazarguil@6wind.com>; Ferruh Yigit
>> <ferruh.yigit@intel.com>; Andrew Rybchenko <arybchenko@solarflare.com>
>> Cc: dev@dpdk.org; stable@dpdk.org
>> Subject: [PATCH v2] ethdev: fix expand RSS flows
>>
>> rte_flow_expand_rss expands rte_flow item list based on the RSS
>> types. In another word, some additional rules are added if the user
>> specified items are not complete enough according to the RSS type,
>> for example:
>>
>>   ... pattern eth / end actions rss type tcp end ...
>>
>> User only provides item eth but want to do RSS on tcp traffic.
>> The pattern is not complete enough to filter TCP traffic only.
>> This will be a problem for some HWs.
>> So some PMDs use rte_flow_expand_rss to expand above user provided
>> flow to:
>>
>>   ... pattern eth / end actions rss types tcp
>>   ... pattern eth / ipv4 / tcp / end actions rss types tcp ...
>>   ... pattern eth / ipv6 / tcp / end actions rss types tcp ...
>>
>> in order to filter TCP traffic only and do RSS correctly.
>>
>> However the current expansion cannot handle pattern as below, which
>> provides ethertype or ip next proto instead of providing an item:
>>
>>   ... pattern eth type is 0x86DD / end actions rss types tcp ...
>>
>> rte_flow_expand_rss will expand above flow to:
>>
>>   ... pattern eth type is 0x86DD / ipv4 / tcp end ...
>>
>> which has conflicting values: 0x86DD vs. ipv4 and some HWs will refuse
>> to create flow.
>>
>> This patch will fix above by checking the last item's spec and to
>> expand RSS flows correctly.
>>
>> Currently only support to complete item list based on ether type or ip
>> next proto.
>>
>> Fixes: 4ed05fcd441b ("ethdev: add flow API to expand RSS flows")
>> Cc: stable@dpdk.org
>>
>> Signed-off-by: Xiaoyu Min <jackmin@mellanox.com>
>> ---
>>  v2:
>>    * more clear commit message
>> ---
> 
> Acked-by: Ori Kam <orika@mellanox.com>
> 

Applied to dpdk-next-net/master, thanks.
diff mbox series

Patch

diff --git a/lib/librte_ethdev/rte_flow.c b/lib/librte_ethdev/rte_flow.c
index 2f86d1affc..ef5ce1790d 100644
--- a/lib/librte_ethdev/rte_flow.c
+++ b/lib/librte_ethdev/rte_flow.c
@@ -173,6 +173,67 @@  flow_err(uint16_t port_id, int ret, struct rte_flow_error *error)
 	return ret;
 }
 
+static enum rte_flow_item_type
+rte_flow_expand_rss_item_complete(const struct rte_flow_item *item)
+{
+	enum rte_flow_item_type ret = RTE_FLOW_ITEM_TYPE_VOID;
+	uint16_t ether_type = 0;
+	uint8_t ip_next_proto = 0;
+
+	if (item == NULL || item->spec == NULL)
+		return ret;
+	switch (item->type) {
+	case RTE_FLOW_ITEM_TYPE_ETH:
+		ether_type = ((const struct rte_flow_item_eth *)
+				(item->spec))->type;
+		if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
+			ret = RTE_FLOW_ITEM_TYPE_IPV4;
+		else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
+			ret = RTE_FLOW_ITEM_TYPE_IPV6;
+		else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
+			ret = RTE_FLOW_ITEM_TYPE_VLAN;
+		break;
+	case RTE_FLOW_ITEM_TYPE_VLAN:
+		ether_type = ((const struct rte_flow_item_vlan *)
+				(item->spec))->inner_type;
+		if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV4)
+			ret = RTE_FLOW_ITEM_TYPE_IPV4;
+		else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_IPV6)
+			ret = RTE_FLOW_ITEM_TYPE_IPV6;
+		else if (rte_be_to_cpu_16(ether_type) == RTE_ETHER_TYPE_VLAN)
+			ret = RTE_FLOW_ITEM_TYPE_VLAN;
+		break;
+	case RTE_FLOW_ITEM_TYPE_IPV4:
+		ip_next_proto = ((const struct rte_flow_item_ipv4 *)
+				(item->spec))->hdr.next_proto_id;
+		if (ip_next_proto == IPPROTO_UDP)
+			ret = RTE_FLOW_ITEM_TYPE_UDP;
+		else if (ip_next_proto == IPPROTO_TCP)
+			ret = RTE_FLOW_ITEM_TYPE_TCP;
+		else if (ip_next_proto == IPPROTO_IP)
+			ret = RTE_FLOW_ITEM_TYPE_IPV4;
+		else if (ip_next_proto == IPPROTO_IPV6)
+			ret = RTE_FLOW_ITEM_TYPE_IPV6;
+		break;
+	case RTE_FLOW_ITEM_TYPE_IPV6:
+		ip_next_proto = ((const struct rte_flow_item_ipv6 *)
+				(item->spec))->hdr.proto;
+		if (ip_next_proto == IPPROTO_UDP)
+			ret = RTE_FLOW_ITEM_TYPE_UDP;
+		else if (ip_next_proto == IPPROTO_TCP)
+			ret = RTE_FLOW_ITEM_TYPE_TCP;
+		else if (ip_next_proto == IPPROTO_IP)
+			ret = RTE_FLOW_ITEM_TYPE_IPV4;
+		else if (ip_next_proto == IPPROTO_IPV6)
+			ret = RTE_FLOW_ITEM_TYPE_IPV6;
+		break;
+	default:
+		ret = RTE_FLOW_ITEM_TYPE_VOID;
+		break;
+	}
+	return ret;
+}
+
 /* Get generic flow operations structure from a port. */
 const struct rte_flow_ops *
 rte_flow_ops_get(uint16_t port_id, struct rte_flow_error *error)
@@ -932,6 +993,11 @@  rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 	size_t lsize;
 	size_t user_pattern_size = 0;
 	void *addr = NULL;
+	const struct rte_flow_expand_node *next = NULL;
+	struct rte_flow_item missed_item;
+	int missed = 0;
+	int elt = 0;
+	const struct rte_flow_item *last_item = NULL;
 
 	lsize = offsetof(struct rte_flow_expand_rss, entry) +
 		elt_n * sizeof(buf->entry[0]);
@@ -942,8 +1008,8 @@  rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 		addr = buf->entry[0].pattern;
 	}
 	for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
-		const struct rte_flow_expand_node *next = NULL;
-
+		if (item->type != RTE_FLOW_ITEM_TYPE_VOID)
+			last_item = item;
 		for (i = 0; node->next && node->next[i]; ++i) {
 			next = &graph[node->next[i]];
 			if (next->type == item->type)
@@ -964,6 +1030,41 @@  rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 	/* Start expanding. */
 	memset(flow_items, 0, sizeof(flow_items));
 	user_pattern_size -= sizeof(*item);
+	/*
+	 * Check if the last valid item has spec set
+	 * and need complete pattern.
+	 */
+	missed_item.type = rte_flow_expand_rss_item_complete(last_item);
+	if (missed_item.type != RTE_FLOW_ITEM_TYPE_VOID) {
+		next = NULL;
+		missed = 1;
+		for (i = 0; node->next && node->next[i]; ++i) {
+			next = &graph[node->next[i]];
+			if (next->type == missed_item.type) {
+				flow_items[0].type = missed_item.type;
+				flow_items[1].type = RTE_FLOW_ITEM_TYPE_END;
+				break;
+			}
+			next = NULL;
+		}
+	}
+	if (next && missed) {
+		elt = 2; /* missed item + item end. */
+		node = next;
+		lsize += elt * sizeof(*item) + user_pattern_size;
+		if ((node->rss_types & types) && lsize <= size) {
+			buf->entry[buf->entries].priority = 1;
+			buf->entry[buf->entries].pattern = addr;
+			buf->entries++;
+			rte_memcpy(addr, buf->entry[0].pattern,
+				   user_pattern_size);
+			addr = (void *)(((uintptr_t)addr) + user_pattern_size);
+			rte_memcpy(addr, flow_items, elt * sizeof(*item));
+			addr = (void *)(((uintptr_t)addr) +
+					elt * sizeof(*item));
+		}
+	}
+	memset(flow_items, 0, sizeof(flow_items));
 	next_node = node->next;
 	stack[stack_pos] = next_node;
 	node = next_node ? &graph[*next_node] : NULL;
@@ -976,21 +1077,24 @@  rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 			 * When the stack_pos is 0, there are 1 element in it,
 			 * plus the addition END item.
 			 */
-			int elt = stack_pos + 2;
-
+			elt = stack_pos + 2;
 			flow_items[stack_pos + 1].type = RTE_FLOW_ITEM_TYPE_END;
 			lsize += elt * sizeof(*item) + user_pattern_size;
 			if (lsize <= size) {
 				size_t n = elt * sizeof(*item);
 
 				buf->entry[buf->entries].priority =
-					stack_pos + 1;
+					stack_pos + 1 + missed;
 				buf->entry[buf->entries].pattern = addr;
 				buf->entries++;
 				rte_memcpy(addr, buf->entry[0].pattern,
 					   user_pattern_size);
 				addr = (void *)(((uintptr_t)addr) +
 						user_pattern_size);
+				rte_memcpy(addr, &missed_item,
+					   missed * sizeof(*item));
+				addr = (void *)(((uintptr_t)addr) +
+					missed * sizeof(*item));
 				rte_memcpy(addr, flow_items, n);
 				addr = (void *)(((uintptr_t)addr) + n);
 			}
@@ -1015,5 +1119,23 @@  rte_flow_expand_rss(struct rte_flow_expand_rss *buf, size_t size,
 		}
 		node = *next_node ? &graph[*next_node] : NULL;
 	};
+	/* no expanded flows but we have missed item, create one rule for it */
+	if (buf->entries == 1 && missed != 0) {
+		elt = 2;
+		lsize += elt * sizeof(*item) + user_pattern_size;
+		if (lsize <= size) {
+			buf->entry[buf->entries].priority = 1;
+			buf->entry[buf->entries].pattern = addr;
+			buf->entries++;
+			flow_items[0].type = missed_item.type;
+			flow_items[1].type = RTE_FLOW_ITEM_TYPE_END;
+			rte_memcpy(addr, buf->entry[0].pattern,
+				   user_pattern_size);
+			addr = (void *)(((uintptr_t)addr) + user_pattern_size);
+			rte_memcpy(addr, flow_items, elt * sizeof(*item));
+			addr = (void *)(((uintptr_t)addr) +
+					elt * sizeof(*item));
+		}
+	}
 	return lsize;
 }