Skip to contentSkip to conversation

bpf: add bpf_get_xdp_hash helper function

PATCH In review Harshitha Ramamurthy 4 participants 7 messages Aug 31, 2020 – Sep 2, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bpf: add bpf_get_xdp_hash helper function

This patch adds a helper function called bpf_get_xdp_hash to calculate the hash for a packet at the XDP layer. In the helper function, we call the kernel flow dissector in non-skb mode by passing the net pointer to calculate the hash.

Changes since RFC:
- accounted for vlans(David Ahern)
- return the correct hash by not using skb_get_hash(David Ahern)
- call __skb_flow_dissect in non-skb mode

---
 include/uapi/linux/bpf.h       |  9 +++++++++
 net/core/filter.c              | 29 +++++++++++++++++++++++++++++
 tools/include/uapi/linux/bpf.h |  9 +++++++++
 3 files changed, 47 insertions(+)
include/uapi/linux/bpf.h +9 -0
index a613750d5515..bffe93b526e7 100644
@@ -3576,6 +3576,14 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of copy_from_user().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u32 bpf_get_xdp_hash(struct xdp_buff *xdp_md)
+ *	Description
+ *		Return the hash for the xdp context passed. This function
+ *		calls skb_flow_dissect in non-skb mode to calculate the
+ *		hash for the packet.
+ *	Return
+ *		The 32-bit hash.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3727,6 +3735,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(get_xdp_hash),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
net/core/filter.c +29 -0
index 47eef9a0be6a..cfb5a6aea6c3 100644
@@ -3765,6 +3765,33 @@ static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
 	.arg3_type      = ARG_ANYTHING,
 };
 
+BPF_CALL_1(bpf_get_xdp_hash, struct xdp_buff *, xdp)
+{
+	void *data_end = xdp->data_end;
+	struct ethhdr *eth = xdp->data;
+	void *data = xdp->data;
+	struct flow_keys keys;
+	u32 ret = 0;
+	int len;
+
+	len = data_end - data;
+	if (len <= 0)
+		return ret;
+	memset(&keys, 0, sizeof(keys));
+	__skb_flow_dissect(dev_net(xdp->rxq->dev), NULL, &flow_keys_dissector,
+			   &keys, data, eth->h_proto, sizeof(*eth), len,
+			   FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);
+	ret = flow_hash_from_keys(&keys);
+	return ret;
+}
+
+static const struct bpf_func_proto bpf_get_xdp_hash_proto = {
+	.func		= bpf_get_xdp_hash,
+	.gpl_only	= false,
+	.ret_type	= RET_INTEGER,
+	.arg1_type	= ARG_PTR_TO_CTX,
+};
+
 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
 				  unsigned long off, unsigned long len)
 {
@@ -6837,6 +6864,8 @@ xdp_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
 		return &bpf_xdp_adjust_tail_proto;
 	case BPF_FUNC_fib_lookup:
 		return &bpf_xdp_fib_lookup_proto;
+	case BPF_FUNC_get_xdp_hash:
+		return &bpf_get_xdp_hash_proto;
 #ifdef CONFIG_INET
 	case BPF_FUNC_sk_lookup_udp:
 		return &bpf_xdp_sk_lookup_udp_proto;
tools/include/uapi/linux/bpf.h +9 -0
index a613750d5515..bffe93b526e7 100644
@@ -3576,6 +3576,14 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of copy_from_user().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u32 bpf_get_xdp_hash(struct xdp_buff *xdp_md)
+ *	Description
+ *		Return the hash for the xdp context passed. This function
+ *		calls skb_flow_dissect in non-skb mode to calculate the
+ *		hash for the packet.
+ *	Return
+ *		The 32-bit hash.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3727,6 +3735,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(get_xdp_hash),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
Signature · 1 line
2.20.1
#1 lore raw

include/uapi/linux/bpf.h union bpf_attr { · line 3580

Quoted 10 lines · Harshitha Ramamurthy
include/uapi/linux/bpf.h +2 -0
index a613750d5515..bffe93b526e7 100644
@@ -3576,6 +3576,14 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of copy_from_user().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u32 bpf_get_xdp_hash(struct xdp_buff *xdp_md)

I thought there was a change recently making the uapi reference xdp_md; xdp_buff is not exported as part of the uapi.

Quoted 27 lines
include/uapi/linux/bpf.h +7 -0
+ *	Description
+ *		Return the hash for the xdp context passed. This function
+ *		calls skb_flow_dissect in non-skb mode to calculate the
+ *		hash for the packet.
+ *	Return
+ *		The 32-bit hash.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3727,6 +3735,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(get_xdp_hash),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which helper
net/core/filter.c +2 -0
index 47eef9a0be6a..cfb5a6aea6c3 100644
@@ -3765,6 +3765,33 @@ static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
 	.arg3_type      = ARG_ANYTHING,
 };
 
+BPF_CALL_1(bpf_get_xdp_hash, struct xdp_buff *, xdp)
+{

net/core/filter.c static const struct bpf_func_proto bpf_xdp_redirect_map_proto = { · line 3779

Quoted 10 lines
net/core/filter.c +10 -0
+	void *data_end = xdp->data_end;
+	struct ethhdr *eth = xdp->data;
+	void *data = xdp->data;
+	struct flow_keys keys;
+	u32 ret = 0;
+	int len;
+
+	len = data_end - data;
+	if (len <= 0)
+		return ret;

you should verify len covers the ethernet header. Looking at __skb_flow_dissect use of hlen presumes it exists.

net/core/filter.c static const struct bpf_func_proto bpf_xdp_redirect_map_proto = { · line 3783

Quoted 4 lines
net/core/filter.c +4 -0
+	memset(&keys, 0, sizeof(keys));
+	__skb_flow_dissect(dev_net(xdp->rxq->dev), NULL, &flow_keys_dissector,
+			   &keys, data, eth->h_proto, sizeof(*eth), len,
+			   FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);

By STOP_AT_FLOW_LABEL I take it you want this to be an L3 hash. Why not add a flags argument to the helper and let the hash be L3 or L4?

you should add test cases and have them cover the permutations - e.g., vlan, Q-in-Q, ipv4, ipv6, non-IP packet for L3 hash and then udp, tcp for L4 hash.

#2 lore raw
Quoted 11 lines

From: [email protected] <[email protected]> On Behalf Of David Ahern
Sent: Monday, August 31, 2020 12:54 PM
To: Ramamurthy, Harshitha <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: Duyck, Alexander H <[email protected]>; Herbert, Tom <[email protected]>
Subject: Re: [PATCH bpf-next] bpf: add bpf_get_xdp_hash helper function

On 8/31/20 1:25 PM, Harshitha Ramamurthy wrote:

Quoted 13 lines

include/uapi/linux/bpf.h union bpf_attr { · line 3580

Quoted 10 lines
include/uapi/linux/bpf.h index renamed from include/uapi/linux/bpf.h +0 -0

a613750d5515..bffe93b526e7 100644

include/uapi/linux/bpf.h +2 -0
@@ -3576,6 +3576,14 @@ union bpf_attr {
  * 		the data in *dst*. This is a wrapper of copy_from_user().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u32 bpf_get_xdp_hash(struct xdp_buff *xdp_md)

I thought there was a change recently making the uapi reference xdp_md; xdp_buff is not exported as part of the uapi.

Not sure what you mean - other xdp related helper functions still use xdp_buff as an argument. Could you point me to an example of what you are referring to?

Quoted 28 lines

net/core/filter.c static const struct bpf_func_proto

Quoted 24 lines
include/uapi/linux/bpf.h +7 -0
+ *	Description
+ *		Return the hash for the xdp context passed. This function
+ *		calls skb_flow_dissect in non-skb mode to calculate the
+ *		hash for the packet.
+ *	Return
+ *		The 32-bit hash.
  */
 #define __BPF_FUNC_MAPPER(FN)		\
 	FN(unspec),			\
@@ -3727,6 +3735,7 @@ union bpf_attr {
 	FN(inode_storage_delete),	\
 	FN(d_path),			\
 	FN(copy_from_user),		\
+	FN(get_xdp_hash),		\
 	/* */
 
 /* integer value in 'imm' field of BPF_CALL instruction selects which

helper diff --git a/net/core/filter.c b/net/core/filter.c index 47eef9a0be6a..cfb5a6aea6c3 100644

net/core/filter.c +0 -0
@@ -3765,6 +3765,33 @@ static const struct bpf_func_proto

bpf_xdp_redirect_map_proto = {

Quoted 3 lines
net/core/filter.c +0 -0
 	.arg3_type      = ARG_ANYTHING,
 };
 
Quoted 14 lines

net/core/filter.c static const struct bpf_func_proto bpf_xdp_redirect_map_proto = { · line 3779

Quoted 11 lines
net/core/filter.c +11 -0
+BPF_CALL_1(bpf_get_xdp_hash, struct xdp_buff *, xdp) {
+	void *data_end = xdp->data_end;
+	struct ethhdr *eth = xdp->data;
+	void *data = xdp->data;
+	struct flow_keys keys;
+	u32 ret = 0;
+	int len;
+
+	len = data_end - data;
+	if (len <= 0)
+		return ret;

you should verify len covers the ethernet header. Looking at __skb_flow_dissect use of hlen presumes it exists.

Yes, I will make sure to return if len < sizeof(struct ethhdr)

Quoted 9 lines

net/core/filter.c static const struct bpf_func_proto · line 3780

Quoted 3 lines
net/core/filter.c +2 -0
+	memset(&keys, 0, sizeof(keys));
+	__skb_flow_dissect(dev_net(xdp->rxq->dev), NULL,

&flow_keys_dissector,

net/core/filter.c static const struct bpf_func_proto bpf_xdp_redirect_map_proto = { · line 3783

Quoted 2 lines
net/core/filter.c +2 -0
+			   &keys, data, eth->h_proto, sizeof(*eth), len,
+			   FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);

By STOP_AT_FLOW_LABEL I take it you want this to be an L3 hash. Why not add a flags argument to the helper and let the hash be L3 or L4?

I wrote this exactly how skb_get_hash calls skb_flow_dissect - with the same flag STOP_AT_FLOW_LABEL. So it should already cover L3 and L4 hash, right? From what I understand STOP_AT_FLOW_LABEL flag is used to only stop parsing when a flow label is seen in ipv6 packets.

Quoted 4 lines

you should add test cases and have them cover the permutations - e.g., vlan, Q-in-Q, ipv4, ipv6, non-IP packet for L3 hash and then udp, tcp for L4 hash.

Sure, I will add test cases using this helper function.

Thanks for the feedback!
Harshitha

#3 lore raw
Quoted 15 lines · Harshitha
Quoted 6 lines · Harshitha Ramamurthy
Quoted 5 lines
include/uapi/linux/bpf.h index renamed from include/uapi/linux/bpf.h +0 -0

a613750d5515..bffe93b526e7 100644

include/uapi/linux/bpf.h +0 -0
@@ -3576,6 +3576,14 @@ union bpf_attr {
Quoted 8 lines

include/uapi/linux/bpf.h union bpf_attr { · line 3580

Quoted 5 lines
include/uapi/linux/bpf.h +2 -0
  * 		the data in *dst*. This is a wrapper of copy_from_user().
  * 	Return
  * 		0 on success, or a negative error in case of failure.
+ *
+ * u32 bpf_get_xdp_hash(struct xdp_buff *xdp_md)

I thought there was a change recently making the uapi reference xdp_md; xdp_buff is not exported as part of the uapi.

Quoted 1 line

Not sure what you mean - other xdp related helper functions still use xdp_buff as an argument. Could you point me to an example of what you are referring to?

a patch from Jakub that is apparently not committed yet.

Quoted 11 lines
Quoted 9 lines

include/uapi/linux/bpf.h union bpf_attr { · line 3582

Quoted 3 lines
include/uapi/linux/bpf.h +2 -0
+	memset(&keys, 0, sizeof(keys));
+	__skb_flow_dissect(dev_net(xdp->rxq->dev), NULL,

&flow_keys_dissector,

include/uapi/linux/bpf.h union bpf_attr { · line 3584

Quoted 2 lines
include/uapi/linux/bpf.h +2 -0
+			   &keys, data, eth->h_proto, sizeof(*eth), len,
+			   FLOW_DISSECTOR_F_STOP_AT_FLOW_LABEL);

By STOP_AT_FLOW_LABEL I take it you want this to be an L3 hash. Why not add a flags argument to the helper and let the hash be L3 or L4?

I wrote this exactly how skb_get_hash calls skb_flow_dissect - with the same flag STOP_AT_FLOW_LABEL. So it should already cover L3 and L4 hash, right? From what I understand STOP_AT_FLOW_LABEL flag is used to only stop parsing when a flow label is seen in ipv6 packets.

right; missed that. That means this new helper will always do an L4 hash for tcp/udp. Adding a flags argument now for the uapi allows a later change to make it an L3 hash.

maintainer
#4 lore raw
In reply to Harshitha Ramamurthy
Quoted 4 lines · Harshitha Ramamurthy

This patch adds a helper function called bpf_get_xdp_hash to calculate the hash for a packet at the XDP layer. In the helper function, we call the kernel flow dissector in non-skb mode by passing the net pointer to calculate the hash.

So this commit msg says 'what' the patch does, but says nothing about 'why' it is needed especially given there's the 1 mio insn limit in place where it should be easy to write that up in BPF anyway. The commit msg needs to have a clear rationale which describes the motivation behind this helper.. why it cannot be done in BPF itself?

Quoted 5 lines

Changes since RFC:
- accounted for vlans(David Ahern)
- return the correct hash by not using skb_get_hash(David Ahern)
- call __skb_flow_dissect in non-skb mode

#5 lore raw
Quoted 15 lines

From: Daniel Borkmann <[email protected]>
Sent: Monday, August 31, 2020 1:33 PM
To: Ramamurthy, Harshitha <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; Duyck, Alexander H <[email protected]>; Herbert, Tom <[email protected]>
Subject: Re: [PATCH bpf-next] bpf: add bpf_get_xdp_hash helper function

Quoted 4 lines · Harshitha Ramamurthy

This patch adds a helper function called bpf_get_xdp_hash to calculate the hash for a packet at the XDP layer. In the helper function, we call the kernel flow dissector in non-skb mode by passing the net pointer to calculate the hash.

Quoted 5 lines

So this commit msg says 'what' the patch does, but says nothing about 'why' it is needed especially given there's the 1 mio insn limit in place where it should be easy to write that up in BPF anyway. The commit msg needs to have a clear rationale which describes the motivation behind this helper.. why it cannot be done in BPF itself?

Okay, will add a rationale in the commit message in the next version for use-case.

Thanks,
Harshitha

Quoted 6 lines
Quoted 6 lines

Changes since RFC:
- accounted for vlans(David Ahern)
- return the correct hash by not using skb_get_hash(David Ahern)
- call __skb_flow_dissect in non-skb mode

#6 lore raw
In reply to Daniel Borkmann
Quoted 7 lines · Daniel Borkmann
Quoted 4 lines · Harshitha Ramamurthy

This patch adds a helper function called bpf_get_xdp_hash to calculate the hash for a packet at the XDP layer. In the helper function, we call the kernel flow dissector in non-skb mode by passing the net pointer to calculate the hash.

Quoted 6 lines

So this commit msg says 'what' the patch does, but says nothing about 'why' it is needed especially given there's the 1 mio insn limit in place where it should be easy to write that up in BPF anyway. The commit msg needs to have a clear rationale which describes the motivation behind this helper.. why it cannot be done in BPF itself?

Daniel,

We already have a fully functional, well tested, and very complete parser supporting many protocols and packet hash functions in the kernel in skb_flow_dissect and friends. I suppose we could replicate all that code in eBPF but I don't think it's fair to say it's easy to get to the same level in eBPF. eBPF does solve the problem of extensibility of kernel code, however IMO if someone is looking for an easy way to get a packet hash then a simple helper function makes sense and results in a much simpler XDP program.

There's some nice potential follow on work also. If the hardware provides a quality L4 hash that could be passed into the XDP program to avoid having to call the helper. If we do invoke the helper it would make sense to return the hash to the driver so that it can set in the skbuff to avoid having the stack call skb_flow_dissect again. We can also have an flow_diessct helper that invokes skb_flow_dissect and returns the meta data based on input keys. This would be useful for filtering in XDP etc.

Tom

Quoted 5 lines
Quoted 5 lines

Changes since RFC:
- accounted for vlans(David Ahern)
- return the correct hash by not using skb_get_hash(David Ahern)
- call __skb_flow_dissect in non-skb mode