| ibv_flow_action_esp(3) | ibv_flow_action_esp(3) |
ibv_flow_action_esp - Flow action esp for verbs
#include <infiniband/verbs.h> struct ibv_flow_action * ibv_create_flow_action_esp(struct ibv_context *ctx,
struct ibv_flow_action_esp *esp); int ibv_modify_flow_action_esp(struct ibv_flow_action *action,
struct ibv_flow_action_esp *esp); int ibv_destroy_flow_action(struct ibv_flow_action *action);
An IPSEC ESP flow steering action allows a flow steering rule to decrypt or encrypt a packet after matching. Each action contains the necessary information for this operation in the params argument.
After the crypto operation the packet will continue to be processed by flow steering rules until it reaches a final action of discard or delivery.
After the action is created, then it should be associated with a struct ibv_flow_attr using struct ibv_flow_spec_action_handle flow specification. Each action can be associated with multiple flows, and ibv_modify_flow_action_esp will alter all associated flows simultaneously.
struct ibv_flow_action_esp {
struct ibv_flow_action_esp_attr *esp_attr;
/* See Key Material */
uint16_t keymat_proto;
uint16_t keymat_len;
void *keymat_ptr;
/* See Replay Protection */
uint16_t replay_proto;
uint16_t replay_len;
void *replay_ptr;
struct ibv_flow_action_esp_encap *esp_encap;
uint32_t comp_mask;
uint32_t esn;
};
The 32 bits of esn will be used to compute the full 64 bit ESN required for the AAD construction.
When in IB_UVERBS_FLOW_ACTION_ESP_FLAGS_INLINE_CRYPTO mode, the implementation will automatically track rollover of the lower 32 bits of the ESN. However, an update of the window is required once every 2^31 sequences.
When in IB_UVERBS_FLOW_ACTION_ESP_FLAGS_FULL_OFFLOAD mode this value is automatically incremended and it is also used for anti-replay checks.
struct ibv_flow_action_esp_attr {
uint32_t spi;
uint32_t seq;
uint32_t tfc_pad;
uint32_t flags;
uint64_t hard_limit_pkts;
};
The implementation may require that encrypt is only used with an egress flow steering rule, and that decrypt is only used with an ingress flow steering rule.
When esp_attr flag IB_UVERBS_FLOW_ACTION_ESP_FLAGS_FULL_OFFLOAD is set the ESP header and trailer are added and removed automatically during the cipher operation. In this case the esn and spi are used to populate and check the ESP header, and any information from the keymat (eg a IV) is placed in the headers and otherwise handled automatically.
For decrypt the hardware will perform anti-replay.
Decryption failure will cause the packet to be dropped.
This action must be combined with the flow steering that identifies the packets protected by the SA defined in this action.
The following members of the esp_attr are used only in full offload mode:
When esp_attr flag IB_UVERBS_FLOW_ACTION_ESP_FLAGS_INLINE_CRYPTO is set the user must providate packets with additional headers.
For encrypt the packet must contain a fully populated IPSEC packet except the data payload is left un-encrypted and there is no IPsec trailer. If the IV must be unpredictable, then a flag should indicate the transofrmation such as IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ.
IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ means that the IV is incremented sequentually. If the IV algorithm is supported by HW, then it could provide support for LSO offload with ESP inline crypto.
Finally, the IV used to encrypt the packet replaces the IV field provided, the payload is encrypted and authenticated, a trailer with padding is added and the ICV is added as well.
For decrypt the packet is authenticated and decrypted in-place, resulting in a decrypted IPSEC packet with no trailer. The result of decryption and authentication can be retrieved from an extended CQ via the ibv_wc_read_XXX(3) function.
This mode must be combined with the flow steering including IBV_FLOW_SPEC_IPV4 and IBV_FLOW_SPEC_ESP to match the outer packet headers to ensure that the action is only applied to IPSEC packets with the correct identifiers.
For inline crypto, we have some special requirements to maintain a stateless ESN while maintaining the same parameters as software. The system supports offloading a portion of the IPSEC flow, enabling a single flow to be split between multiple NICs.
We require a “modify” command once every 2^31 packets. This modify command allows the implementation in HW to be stateless, as follows:
ESN 1 ESN 2 ESN 3 |-------------*-------------|-------------*-------------|-------------* ^ ^ ^ ^ ^ ^
^ - marks where command invoked to update the SA ESN state machine.
| - marks the start of the ESN scope (0-2^32-1). At this point move SA ESN “new_window” bit to zero and increment ESN.
* - marks the middle of the ESN scope (2^31). At this point move SA ESN “new_window” bit to one.
For decryption the implementation uses the following state machine to determine ESN:
if (!overlap) {
use esn // regardless of packet.seq
} else { // new_window
if (packet.seq >= 2^31)
use esn
else // packet.seq < 2^31
use esn+1
}
This mechanism is controlled by the esp_attr flag:
Setting this flag indicates that the bottom of the replay window is between 2^31 - 2^32.
The AES GCM crypto algorithm as defined by RFC4106. This struct is to be provided in keymat_ptr when keymat_proto is set to IBV_ACTION_ESP_KEYMAT_AES_GCM.
struct ibv_flow_action_esp_aes_keymat_aes_gcm {
uint64_t iv;
uint32_t iv_algo; /* Use enum ib_uverbs_flow_action_esp_aes_gcm_keymat_iv_algo */
uint32_t salt;
uint32_t icv_len;
uint32_t key_len;
uint32_t aes_key[256 / 32];
};
For a given key, the IV MUST NOT be reused.
The only supported value is IB_UVERBS_FLOW_ACTION_IV_ALGO_SEQ to generate sequantial IVs.
A shifting bitmap is used to identify which packets have already been transmitted. Each bit in the bitmap represents a packet, it is set if a packet with this ESP sequence number has been received and it passed authentication. If a packet with the same sequence is received, then the bit is already set, causing replay protection to drop the packet. The bitmap represents a window of size sequence numbers. If a newer sequence number is received, then the bitmap will shift to represent this as in RFC6479. The replay window cannot shift more than 2^31 sequence numbers forward.
This struct is to be provided in replay_ptr when reply_proto is set to IBV_FLOW_ACTION_ESP_REPLAY_BMP. In this mode reply_ptr and reply_len should point to a struct ibv_flow_action_esp_replay_bmp containing: size : The size of the bitmap.
An esp_encap specification is required when eps_attr flags IB_UVERBS_FLOW_ACTION_ESP_FLAGS_TUNNEL is set. It is used to provide the fields for the encapsulation header that is added/removed to/from packets. Tunnel and Transport mode are defined as in RFC4301. UDP encapsulation of ESP can be specified by providing the appropriate UDP header.
This setting is only used in IB_UVERBS_FLOW_ACTION_ESP_FLAGS_FULL_OFFLOAD mode.
struct ibv_flow_action_esp_encap {
void *val; /* pointer to struct ibv_flow_xxxx_filter */
struct ibv_flow_action_esp_encap *next_ptr;
uint16_t len; /* Len of mask and pointer (separately) */
uint16_t type; /* Use flow_spec enum */
};
Each link in the list specifies a network header in the same manner as the flow steering API. The header should be selected from a supported header in `enum ibv_flow_spec_type'.
Upon success ibv_create_flow_action_esp will return a new struct ibv_flow_action object, on error NULL will be returned and errno will be set.
Upon success ibv_modify_action_esp will return 0. On error the value of errno will be returned. If ibv_modify_flow_action fails, it is guaranteed that the last action still holds. If it succeeds, there is a point in the future where the old action is applied on all packets until this point and the new one is applied on all packets from this point and on.
ibv_create_flow(3), ibv_destroy_action(3), RFC 4106