Thread (184 messages) 184 messages, 12 authors, 2022-09-27

Re: [dpdk-dev] [PATCH] RFC: ethdev: add reassembly offload

From: Xu, Rosen <hidden>
Date: 2021-09-08 06:34:21

Hi,
quoted hunk ↗ jump to hunk
-----Original Message-----
From: dev <redacted> On Behalf Of Akhil Goyal
Sent: Monday, August 23, 2021 18:03
To: dev@dpdk.org
Cc: anoobj@marvell.com; Nicolau, Radu <redacted>; Doherty,
Declan [off-list ref]; hemant.agrawal@nxp.com;
matan@nvidia.com; Ananyev, Konstantin [off-list ref];
thomas@monjalon.net; adwivedi@marvell.com; Yigit, Ferruh
[off-list ref]; andrew.rybchenko@oktetlabs.ru; Akhil Goyal
[off-list ref]
Subject: [dpdk-dev] [PATCH] RFC: ethdev: add reassembly offload

Reassembly is a costly operation if it is done in software, however, if it is
offloaded to HW, it can considerably save application cycles.
The operation becomes even more costlier if IP fragmants are encrypted.

To resolve above two issues, a new offload DEV_RX_OFFLOAD_REASSEMBLY
is introduced in ethdev for devices which can attempt reassembly of packets
in hardware.
rte_eth_dev_info is added with the reassembly capabilities which a device
can support.
Now, if IP fragments are encrypted, reassembly can also be attempted while
doing inline IPsec processing.
This is controlled by a flag in rte_security_ipsec_sa_options to enable
reassembly of encrypted IP fragments in the inline path.

The resulting reassembled packet would be a typical segmented mbuf in case
of success.

And if reassembly of fragments is failed or is incomplete (if fragments do not
come before the reass_timeout), the mbuf is updated with an ol_flag
PKT_RX_REASSEMBLY_INCOMPLETE and mbuf is returned as is. Now
application may decide the fate of the packet to wait more for fragments to
come or drop.

Signed-off-by: Akhil Goyal <redacted>
---
 lib/ethdev/rte_ethdev.c     |  1 +
 lib/ethdev/rte_ethdev.h     | 18 +++++++++++++++++-
 lib/mbuf/rte_mbuf_core.h    |  3 ++-
 lib/security/rte_security.h | 10 ++++++++++
 4 files changed, 30 insertions(+), 2 deletions(-)
diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c index
9d95cd11e1..1ab3a093cf 100644
--- a/lib/ethdev/rte_ethdev.c
+++ b/lib/ethdev/rte_ethdev.c
@@ -119,6 +119,7 @@ static const struct {
 	RTE_RX_OFFLOAD_BIT2STR(VLAN_FILTER),
 	RTE_RX_OFFLOAD_BIT2STR(VLAN_EXTEND),
 	RTE_RX_OFFLOAD_BIT2STR(JUMBO_FRAME),
+	RTE_RX_OFFLOAD_BIT2STR(REASSEMBLY),
 	RTE_RX_OFFLOAD_BIT2STR(SCATTER),
 	RTE_RX_OFFLOAD_BIT2STR(TIMESTAMP),
 	RTE_RX_OFFLOAD_BIT2STR(SECURITY),
diff --git a/lib/ethdev/rte_ethdev.h b/lib/ethdev/rte_ethdev.h index
d2b27c351f..e89a4dc1eb 100644
--- a/lib/ethdev/rte_ethdev.h
+++ b/lib/ethdev/rte_ethdev.h
@@ -1360,6 +1360,7 @@ struct rte_eth_conf {
 #define DEV_RX_OFFLOAD_VLAN_FILTER	0x00000200
 #define DEV_RX_OFFLOAD_VLAN_EXTEND	0x00000400
 #define DEV_RX_OFFLOAD_JUMBO_FRAME	0x00000800
+#define DEV_RX_OFFLOAD_REASSEMBLY	0x00001000
 #define DEV_RX_OFFLOAD_SCATTER		0x00002000
 /**
  * Timestamp is set by the driver in
RTE_MBUF_DYNFIELD_TIMESTAMP_NAME @@ -1477,6 +1478,20 @@ struct
rte_eth_dev_portconf {
  */
 #define RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID
	(UINT16_MAX)

+/**
+ * Reassembly capabilities that a device can support.
+ * The device which can support reassembly offload should set
+ * DEV_RX_OFFLOAD_REASSEMBLY
+ */
+struct rte_eth_reass_capa {
+	/** Maximum time in ns that a fragment can wait for further
fragments */
+	uint64_t reass_timeout;
+	/** Maximum number of fragments that device can reassemble */
+	uint16_t max_frags;
+	/** Reserved for future capabilities */
+	uint16_t reserved[3];
+};
IP reassembly occurs at the final recipient of the message, NIC attempts to do it has a fer challenges. The reason is that having NICs need to worry about reassembling fragments would increase their complexity, so most likely it only can handle range length of datagrams. Seems rte_eth_reass_capa miss the max original datagrams length which NIC can support, this features is better to be negotiated between NIC and SW as well.
quoted hunk ↗ jump to hunk
 /**
  * Ethernet device associated switch information
  */
@@ -1582,8 +1597,9 @@ struct rte_eth_dev_info {
 	 * embedded managed interconnect/switch.
 	 */
 	struct rte_eth_switch_info switch_info;
+	/* Reassembly capabilities of a device for reassembly offload */
+	struct rte_eth_reass_capa reass_capa;

-	uint64_t reserved_64s[2]; /**< Reserved for future fields */
 	void *reserved_ptrs[2];   /**< Reserved for future fields */
 };
diff --git a/lib/mbuf/rte_mbuf_core.h b/lib/mbuf/rte_mbuf_core.h index
bb38d7f581..cea25c87f7 100644
--- a/lib/mbuf/rte_mbuf_core.h
+++ b/lib/mbuf/rte_mbuf_core.h
@@ -200,10 +200,11 @@ extern "C" {
 #define PKT_RX_OUTER_L4_CKSUM_BAD	(1ULL << 21)
 #define PKT_RX_OUTER_L4_CKSUM_GOOD	(1ULL << 22)
 #define PKT_RX_OUTER_L4_CKSUM_INVALID	((1ULL << 21) | (1ULL << 22))
+#define PKT_RX_REASSEMBLY_INCOMPLETE	(1ULL << 23)

 /* add new RX flags here, don't forget to update PKT_FIRST_FREE */

-#define PKT_FIRST_FREE (1ULL << 23)
+#define PKT_FIRST_FREE (1ULL << 24)
 #define PKT_LAST_FREE (1ULL << 40)

 /* add new TX flags here, don't forget to update PKT_LAST_FREE  */ diff --git
a/lib/security/rte_security.h b/lib/security/rte_security.h index
88d31de0a6..364eeb5cd4 100644
--- a/lib/security/rte_security.h
+++ b/lib/security/rte_security.h
@@ -181,6 +181,16 @@ struct rte_security_ipsec_sa_options {
 	 * * 0: Disable per session security statistics collection for this SA.
 	 */
 	uint32_t stats : 1;
+
+	/** Enable reassembly on incoming packets.
+	 *
+	 * * 1: Enable driver to try reassembly of encrypted IP packets for
+	 *      this SA, if supported by the driver. This feature will work
+	 *      only if rx_offload DEV_RX_OFFLOAD_REASSEMBLY is set in
+	 *      inline ethernet device.
+	 * * 0: Disable reassembly of packets (default).
+	 */
+	uint32_t reass_en : 1;
 };

 /** IPSec security association direction */
--
2.25.1
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help