net/tls: Add generic NIC offload infrastructure
authorIlya Lesokhin <ilyal@mellanox.com>
Mon, 30 Apr 2018 07:16:16 +0000 (10:16 +0300)
committerDavid S. Miller <davem@davemloft.net>
Tue, 1 May 2018 13:42:47 +0000 (09:42 -0400)
This patch adds a generic infrastructure to offload TLS crypto to a
network device. It enables the kernel TLS socket to skip encryption
and authentication operations on the transmit side of the data path.
Leaving those computationally expensive operations to the NIC.

The NIC offload infrastructure builds TLS records and pushes them to
the TCP layer just like the SW KTLS implementation and using the same
API.
TCP segmentation is mostly unaffected. Currently the only exception is
that we prevent mixed SKBs where only part of the payload requires
offload. In the future we are likely to add a similar restriction
following a change cipher spec record.

The notable differences between SW KTLS and NIC offloaded TLS
implementations are as follows:
1. The offloaded implementation builds "plaintext TLS record", those
records contain plaintext instead of ciphertext and place holder bytes
instead of authentication tags.
2. The offloaded implementation maintains a mapping from TCP sequence
number to TLS records. Thus given a TCP SKB sent from a NIC offloaded
TLS socket, we can use the tls NIC offload infrastructure to obtain
enough context to encrypt the payload of the SKB.
A TLS record is released when the last byte of the record is ack'ed,
this is done through the new icsk_clean_acked callback.

The infrastructure should be extendable to support various NIC offload
implementations.  However it is currently written with the
implementation below in mind:
The NIC assumes that packets from each offloaded stream are sent as
plaintext and in-order. It keeps track of the TLS records in the TCP
stream. When a packet marked for offload is transmitted, the NIC
encrypts the payload in-place and puts authentication tags in the
relevant place holders.

The responsibility for handling out-of-order packets (i.e. TCP
retransmission, qdisc drops) falls on the netdev driver.

The netdev driver keeps track of the expected TCP SN from the NIC's
perspective.  If the next packet to transmit matches the expected TCP
SN, the driver advances the expected TCP SN, and transmits the packet
with TLS offload indication.

If the next packet to transmit does not match the expected TCP SN. The
driver calls the TLS layer to obtain the TLS record that includes the
TCP of the packet for transmission. Using this TLS record, the driver
posts a work entry on the transmit queue to reconstruct the NIC TLS
state required for the offload of the out-of-order packet. It updates
the expected TCP SN accordingly and transmits the now in-order packet.
The same queue is used for packet transmission and TLS context
reconstruction to avoid the need for flushing the transmit queue before
issuing the context reconstruction request.

Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Aviad Yehezkel <aviadye@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tls.h
net/tls/Kconfig
net/tls/Makefile
net/tls/tls_device.c [new file with mode: 0644]
net/tls/tls_device_fallback.c [new file with mode: 0644]
net/tls/tls_main.c

index 95a8c60b36bee88fd07b04870c03019fb1d4a066..8c56809eb384a2de8c4de46bd6ce97c9b943354a 100644 (file)
@@ -116,6 +116,37 @@ struct tls_sw_context_rx {
        bool decrypted;
 };
 
+struct tls_record_info {
+       struct list_head list;
+       u32 end_seq;
+       int len;
+       int num_frags;
+       skb_frag_t frags[MAX_SKB_FRAGS];
+};
+
+struct tls_offload_context {
+       struct crypto_aead *aead_send;
+       spinlock_t lock;        /* protects records list */
+       struct list_head records_list;
+       struct tls_record_info *open_record;
+       struct tls_record_info *retransmit_hint;
+       u64 hint_record_sn;
+       u64 unacked_record_sn;
+
+       struct scatterlist sg_tx_data[MAX_SKB_FRAGS];
+       void (*sk_destruct)(struct sock *sk);
+       u8 driver_state[];
+       /* The TLS layer reserves room for driver specific state
+        * Currently the belief is that there is not enough
+        * driver specific state to justify another layer of indirection
+        */
+#define TLS_DRIVER_STATE_SIZE (max_t(size_t, 8, sizeof(void *)))
+};
+
+#define TLS_OFFLOAD_CONTEXT_SIZE                                               \
+       (ALIGN(sizeof(struct tls_offload_context), sizeof(void *)) +           \
+        TLS_DRIVER_STATE_SIZE)
+
 enum {
        TLS_PENDING_CLOSED_RECORD
 };
@@ -195,9 +226,28 @@ ssize_t tls_sw_splice_read(struct socket *sock, loff_t *ppos,
                           struct pipe_inode_info *pipe,
                           size_t len, unsigned int flags);
 
-void tls_sk_destruct(struct sock *sk, struct tls_context *ctx);
-void tls_icsk_clean_acked(struct sock *sk);
+int tls_set_device_offload(struct sock *sk, struct tls_context *ctx);
+int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
+int tls_device_sendpage(struct sock *sk, struct page *page,
+                       int offset, size_t size, int flags);
+void tls_device_sk_destruct(struct sock *sk);
+void tls_device_init(void);
+void tls_device_cleanup(void);
+
+struct tls_record_info *tls_get_record(struct tls_offload_context *context,
+                                      u32 seq, u64 *p_record_sn);
+
+static inline bool tls_record_is_start_marker(struct tls_record_info *rec)
+{
+       return rec->len == 0;
+}
+
+static inline u32 tls_record_start_seq(struct tls_record_info *rec)
+{
+       return rec->end_seq - rec->len;
+}
 
+void tls_sk_destruct(struct sock *sk, struct tls_context *ctx);
 int tls_push_sg(struct sock *sk, struct tls_context *ctx,
                struct scatterlist *sg, u16 first_offset,
                int flags);
@@ -234,6 +284,13 @@ static inline bool tls_is_pending_open_record(struct tls_context *tls_ctx)
        return tls_ctx->pending_open_record_frags;
 }
 
+static inline bool tls_is_sk_tx_device_offloaded(struct sock *sk)
+{
+       return sk_fullsock(sk) &&
+              /* matches smp_store_release in tls_set_device_offload */
+              smp_load_acquire(&sk->sk_destruct) == &tls_device_sk_destruct;
+}
+
 static inline void tls_err_abort(struct sock *sk, int err)
 {
        sk->sk_err = err;
@@ -329,4 +386,12 @@ int tls_proccess_cmsg(struct sock *sk, struct msghdr *msg,
 void tls_register_device(struct tls_device *device);
 void tls_unregister_device(struct tls_device *device);
 
+struct sk_buff *tls_validate_xmit_skb(struct sock *sk,
+                                     struct net_device *dev,
+                                     struct sk_buff *skb);
+
+int tls_sw_fallback_init(struct sock *sk,
+                        struct tls_offload_context *offload_ctx,
+                        struct tls_crypto_info *crypto_info);
+
 #endif /* _TLS_OFFLOAD_H */
index 89b8745a986f06c00bd35f7a4fc6fff47c25120a..73f05ece53d0c955df2bac6d043e1e354d8fb7dc 100644 (file)
@@ -14,3 +14,13 @@ config TLS
        encryption handling of the TLS protocol to be done in-kernel.
 
        If unsure, say N.
+
+config TLS_DEVICE
+       bool "Transport Layer Security HW offload"
+       depends on TLS
+       select SOCK_VALIDATE_XMIT
+       default n
+       help
+       Enable kernel support for HW offload of the TLS protocol.
+
+       If unsure, say N.
index a930fd1c4f7b88651d5686355cbea8ba55942941..4d6b728a67d0c122a41c1c0d5d5840598ba56472 100644 (file)
@@ -5,3 +5,5 @@
 obj-$(CONFIG_TLS) += tls.o
 
 tls-y := tls_main.o tls_sw.o
+
+tls-$(CONFIG_TLS_DEVICE) += tls_device.o tls_device_fallback.o
diff --git a/net/tls/tls_device.c b/net/tls/tls_device.c
new file mode 100644 (file)
index 0000000..ac45d62
--- /dev/null
@@ -0,0 +1,764 @@
+/* Copyright (c) 2018, Mellanox Technologies All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <crypto/aead.h>
+#include <linux/highmem.h>
+#include <linux/module.h>
+#include <linux/netdevice.h>
+#include <net/dst.h>
+#include <net/inet_connection_sock.h>
+#include <net/tcp.h>
+#include <net/tls.h>
+
+/* device_offload_lock is used to synchronize tls_dev_add
+ * against NETDEV_DOWN notifications.
+ */
+static DECLARE_RWSEM(device_offload_lock);
+
+static void tls_device_gc_task(struct work_struct *work);
+
+static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
+static LIST_HEAD(tls_device_gc_list);
+static LIST_HEAD(tls_device_list);
+static DEFINE_SPINLOCK(tls_device_lock);
+
+static void tls_device_free_ctx(struct tls_context *ctx)
+{
+       struct tls_offload_context *offload_ctx = tls_offload_ctx(ctx);
+
+       kfree(offload_ctx);
+       kfree(ctx);
+}
+
+static void tls_device_gc_task(struct work_struct *work)
+{
+       struct tls_context *ctx, *tmp;
+       unsigned long flags;
+       LIST_HEAD(gc_list);
+
+       spin_lock_irqsave(&tls_device_lock, flags);
+       list_splice_init(&tls_device_gc_list, &gc_list);
+       spin_unlock_irqrestore(&tls_device_lock, flags);
+
+       list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
+               struct net_device *netdev = ctx->netdev;
+
+               if (netdev) {
+                       netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
+                                                       TLS_OFFLOAD_CTX_DIR_TX);
+                       dev_put(netdev);
+               }
+
+               list_del(&ctx->list);
+               tls_device_free_ctx(ctx);
+       }
+}
+
+static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&tls_device_lock, flags);
+       list_move_tail(&ctx->list, &tls_device_gc_list);
+
+       /* schedule_work inside the spinlock
+        * to make sure tls_device_down waits for that work.
+        */
+       schedule_work(&tls_device_gc_work);
+
+       spin_unlock_irqrestore(&tls_device_lock, flags);
+}
+
+/* We assume that the socket is already connected */
+static struct net_device *get_netdev_for_sock(struct sock *sk)
+{
+       struct dst_entry *dst = sk_dst_get(sk);
+       struct net_device *netdev = NULL;
+
+       if (likely(dst)) {
+               netdev = dst->dev;
+               dev_hold(netdev);
+       }
+
+       dst_release(dst);
+
+       return netdev;
+}
+
+static void destroy_record(struct tls_record_info *record)
+{
+       int nr_frags = record->num_frags;
+       skb_frag_t *frag;
+
+       while (nr_frags-- > 0) {
+               frag = &record->frags[nr_frags];
+               __skb_frag_unref(frag);
+       }
+       kfree(record);
+}
+
+static void delete_all_records(struct tls_offload_context *offload_ctx)
+{
+       struct tls_record_info *info, *temp;
+
+       list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
+               list_del(&info->list);
+               destroy_record(info);
+       }
+
+       offload_ctx->retransmit_hint = NULL;
+}
+
+static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
+{
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_record_info *info, *temp;
+       struct tls_offload_context *ctx;
+       u64 deleted_records = 0;
+       unsigned long flags;
+
+       if (!tls_ctx)
+               return;
+
+       ctx = tls_offload_ctx(tls_ctx);
+
+       spin_lock_irqsave(&ctx->lock, flags);
+       info = ctx->retransmit_hint;
+       if (info && !before(acked_seq, info->end_seq)) {
+               ctx->retransmit_hint = NULL;
+               list_del(&info->list);
+               destroy_record(info);
+               deleted_records++;
+       }
+
+       list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
+               if (before(acked_seq, info->end_seq))
+                       break;
+               list_del(&info->list);
+
+               destroy_record(info);
+               deleted_records++;
+       }
+
+       ctx->unacked_record_sn += deleted_records;
+       spin_unlock_irqrestore(&ctx->lock, flags);
+}
+
+/* At this point, there should be no references on this
+ * socket and no in-flight SKBs associated with this
+ * socket, so it is safe to free all the resources.
+ */
+void tls_device_sk_destruct(struct sock *sk)
+{
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_offload_context *ctx = tls_offload_ctx(tls_ctx);
+
+       if (ctx->open_record)
+               destroy_record(ctx->open_record);
+
+       delete_all_records(ctx);
+       crypto_free_aead(ctx->aead_send);
+       ctx->sk_destruct(sk);
+       clean_acked_data_disable(inet_csk(sk));
+
+       if (refcount_dec_and_test(&tls_ctx->refcount))
+               tls_device_queue_ctx_destruction(tls_ctx);
+}
+EXPORT_SYMBOL(tls_device_sk_destruct);
+
+static void tls_append_frag(struct tls_record_info *record,
+                           struct page_frag *pfrag,
+                           int size)
+{
+       skb_frag_t *frag;
+
+       frag = &record->frags[record->num_frags - 1];
+       if (frag->page.p == pfrag->page &&
+           frag->page_offset + frag->size == pfrag->offset) {
+               frag->size += size;
+       } else {
+               ++frag;
+               frag->page.p = pfrag->page;
+               frag->page_offset = pfrag->offset;
+               frag->size = size;
+               ++record->num_frags;
+               get_page(pfrag->page);
+       }
+
+       pfrag->offset += size;
+       record->len += size;
+}
+
+static int tls_push_record(struct sock *sk,
+                          struct tls_context *ctx,
+                          struct tls_offload_context *offload_ctx,
+                          struct tls_record_info *record,
+                          struct page_frag *pfrag,
+                          int flags,
+                          unsigned char record_type)
+{
+       struct tcp_sock *tp = tcp_sk(sk);
+       struct page_frag dummy_tag_frag;
+       skb_frag_t *frag;
+       int i;
+
+       /* fill prepend */
+       frag = &record->frags[0];
+       tls_fill_prepend(ctx,
+                        skb_frag_address(frag),
+                        record->len - ctx->tx.prepend_size,
+                        record_type);
+
+       /* HW doesn't care about the data in the tag, because it fills it. */
+       dummy_tag_frag.page = skb_frag_page(frag);
+       dummy_tag_frag.offset = 0;
+
+       tls_append_frag(record, &dummy_tag_frag, ctx->tx.tag_size);
+       record->end_seq = tp->write_seq + record->len;
+       spin_lock_irq(&offload_ctx->lock);
+       list_add_tail(&record->list, &offload_ctx->records_list);
+       spin_unlock_irq(&offload_ctx->lock);
+       offload_ctx->open_record = NULL;
+       set_bit(TLS_PENDING_CLOSED_RECORD, &ctx->flags);
+       tls_advance_record_sn(sk, &ctx->tx);
+
+       for (i = 0; i < record->num_frags; i++) {
+               frag = &record->frags[i];
+               sg_unmark_end(&offload_ctx->sg_tx_data[i]);
+               sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
+                           frag->size, frag->page_offset);
+               sk_mem_charge(sk, frag->size);
+               get_page(skb_frag_page(frag));
+       }
+       sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
+
+       /* all ready, send */
+       return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
+}
+
+static int tls_create_new_record(struct tls_offload_context *offload_ctx,
+                                struct page_frag *pfrag,
+                                size_t prepend_size)
+{
+       struct tls_record_info *record;
+       skb_frag_t *frag;
+
+       record = kmalloc(sizeof(*record), GFP_KERNEL);
+       if (!record)
+               return -ENOMEM;
+
+       frag = &record->frags[0];
+       __skb_frag_set_page(frag, pfrag->page);
+       frag->page_offset = pfrag->offset;
+       skb_frag_size_set(frag, prepend_size);
+
+       get_page(pfrag->page);
+       pfrag->offset += prepend_size;
+
+       record->num_frags = 1;
+       record->len = prepend_size;
+       offload_ctx->open_record = record;
+       return 0;
+}
+
+static int tls_do_allocation(struct sock *sk,
+                            struct tls_offload_context *offload_ctx,
+                            struct page_frag *pfrag,
+                            size_t prepend_size)
+{
+       int ret;
+
+       if (!offload_ctx->open_record) {
+               if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
+                                                  sk->sk_allocation))) {
+                       sk->sk_prot->enter_memory_pressure(sk);
+                       sk_stream_moderate_sndbuf(sk);
+                       return -ENOMEM;
+               }
+
+               ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
+               if (ret)
+                       return ret;
+
+               if (pfrag->size > pfrag->offset)
+                       return 0;
+       }
+
+       if (!sk_page_frag_refill(sk, pfrag))
+               return -ENOMEM;
+
+       return 0;
+}
+
+static int tls_push_data(struct sock *sk,
+                        struct iov_iter *msg_iter,
+                        size_t size, int flags,
+                        unsigned char record_type)
+{
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_offload_context *ctx = tls_offload_ctx(tls_ctx);
+       int tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
+       int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
+       struct tls_record_info *record = ctx->open_record;
+       struct page_frag *pfrag;
+       size_t orig_size = size;
+       u32 max_open_record_len;
+       int copy, rc = 0;
+       bool done = false;
+       long timeo;
+
+       if (flags &
+           ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
+               return -ENOTSUPP;
+
+       if (sk->sk_err)
+               return -sk->sk_err;
+
+       timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
+       rc = tls_complete_pending_work(sk, tls_ctx, flags, &timeo);
+       if (rc < 0)
+               return rc;
+
+       pfrag = sk_page_frag(sk);
+
+       /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
+        * we need to leave room for an authentication tag.
+        */
+       max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
+                             tls_ctx->tx.prepend_size;
+       do {
+               rc = tls_do_allocation(sk, ctx, pfrag,
+                                      tls_ctx->tx.prepend_size);
+               if (rc) {
+                       rc = sk_stream_wait_memory(sk, &timeo);
+                       if (!rc)
+                               continue;
+
+                       record = ctx->open_record;
+                       if (!record)
+                               break;
+handle_error:
+                       if (record_type != TLS_RECORD_TYPE_DATA) {
+                               /* avoid sending partial
+                                * record with type !=
+                                * application_data
+                                */
+                               size = orig_size;
+                               destroy_record(record);
+                               ctx->open_record = NULL;
+                       } else if (record->len > tls_ctx->tx.prepend_size) {
+                               goto last_record;
+                       }
+
+                       break;
+               }
+
+               record = ctx->open_record;
+               copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
+               copy = min_t(size_t, copy, (max_open_record_len - record->len));
+
+               if (copy_from_iter_nocache(page_address(pfrag->page) +
+                                              pfrag->offset,
+                                          copy, msg_iter) != copy) {
+                       rc = -EFAULT;
+                       goto handle_error;
+               }
+               tls_append_frag(record, pfrag, copy);
+
+               size -= copy;
+               if (!size) {
+last_record:
+                       tls_push_record_flags = flags;
+                       if (more) {
+                               tls_ctx->pending_open_record_frags =
+                                               record->num_frags;
+                               break;
+                       }
+
+                       done = true;
+               }
+
+               if (done || record->len >= max_open_record_len ||
+                   (record->num_frags >= MAX_SKB_FRAGS - 1)) {
+                       rc = tls_push_record(sk,
+                                            tls_ctx,
+                                            ctx,
+                                            record,
+                                            pfrag,
+                                            tls_push_record_flags,
+                                            record_type);
+                       if (rc < 0)
+                               break;
+               }
+       } while (!done);
+
+       if (orig_size - size > 0)
+               rc = orig_size - size;
+
+       return rc;
+}
+
+int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
+{
+       unsigned char record_type = TLS_RECORD_TYPE_DATA;
+       int rc;
+
+       lock_sock(sk);
+
+       if (unlikely(msg->msg_controllen)) {
+               rc = tls_proccess_cmsg(sk, msg, &record_type);
+               if (rc)
+                       goto out;
+       }
+
+       rc = tls_push_data(sk, &msg->msg_iter, size,
+                          msg->msg_flags, record_type);
+
+out:
+       release_sock(sk);
+       return rc;
+}
+
+int tls_device_sendpage(struct sock *sk, struct page *page,
+                       int offset, size_t size, int flags)
+{
+       struct iov_iter msg_iter;
+       char *kaddr = kmap(page);
+       struct kvec iov;
+       int rc;
+
+       if (flags & MSG_SENDPAGE_NOTLAST)
+               flags |= MSG_MORE;
+
+       lock_sock(sk);
+
+       if (flags & MSG_OOB) {
+               rc = -ENOTSUPP;
+               goto out;
+       }
+
+       iov.iov_base = kaddr + offset;
+       iov.iov_len = size;
+       iov_iter_kvec(&msg_iter, WRITE | ITER_KVEC, &iov, 1, size);
+       rc = tls_push_data(sk, &msg_iter, size,
+                          flags, TLS_RECORD_TYPE_DATA);
+       kunmap(page);
+
+out:
+       release_sock(sk);
+       return rc;
+}
+
+struct tls_record_info *tls_get_record(struct tls_offload_context *context,
+                                      u32 seq, u64 *p_record_sn)
+{
+       u64 record_sn = context->hint_record_sn;
+       struct tls_record_info *info;
+
+       info = context->retransmit_hint;
+       if (!info ||
+           before(seq, info->end_seq - info->len)) {
+               /* if retransmit_hint is irrelevant start
+                * from the beggining of the list
+                */
+               info = list_first_entry(&context->records_list,
+                                       struct tls_record_info, list);
+               record_sn = context->unacked_record_sn;
+       }
+
+       list_for_each_entry_from(info, &context->records_list, list) {
+               if (before(seq, info->end_seq)) {
+                       if (!context->retransmit_hint ||
+                           after(info->end_seq,
+                                 context->retransmit_hint->end_seq)) {
+                               context->hint_record_sn = record_sn;
+                               context->retransmit_hint = info;
+                       }
+                       *p_record_sn = record_sn;
+                       return info;
+               }
+               record_sn++;
+       }
+
+       return NULL;
+}
+EXPORT_SYMBOL(tls_get_record);
+
+static int tls_device_push_pending_record(struct sock *sk, int flags)
+{
+       struct iov_iter msg_iter;
+
+       iov_iter_kvec(&msg_iter, WRITE | ITER_KVEC, NULL, 0, 0);
+       return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
+}
+
+int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
+{
+       u16 nonce_size, tag_size, iv_size, rec_seq_size;
+       struct tls_record_info *start_marker_record;
+       struct tls_offload_context *offload_ctx;
+       struct tls_crypto_info *crypto_info;
+       struct net_device *netdev;
+       char *iv, *rec_seq;
+       struct sk_buff *skb;
+       int rc = -EINVAL;
+       __be64 rcd_sn;
+
+       if (!ctx)
+               goto out;
+
+       if (ctx->priv_ctx_tx) {
+               rc = -EEXIST;
+               goto out;
+       }
+
+       start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
+       if (!start_marker_record) {
+               rc = -ENOMEM;
+               goto out;
+       }
+
+       offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE, GFP_KERNEL);
+       if (!offload_ctx) {
+               rc = -ENOMEM;
+               goto free_marker_record;
+       }
+
+       crypto_info = &ctx->crypto_send;
+       switch (crypto_info->cipher_type) {
+       case TLS_CIPHER_AES_GCM_128:
+               nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
+               tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
+               iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
+               iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
+               rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
+               rec_seq =
+                ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
+               break;
+       default:
+               rc = -EINVAL;
+               goto free_offload_ctx;
+       }
+
+       ctx->tx.prepend_size = TLS_HEADER_SIZE + nonce_size;
+       ctx->tx.tag_size = tag_size;
+       ctx->tx.overhead_size = ctx->tx.prepend_size + ctx->tx.tag_size;
+       ctx->tx.iv_size = iv_size;
+       ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
+                            GFP_KERNEL);
+       if (!ctx->tx.iv) {
+               rc = -ENOMEM;
+               goto free_offload_ctx;
+       }
+
+       memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
+
+       ctx->tx.rec_seq_size = rec_seq_size;
+       ctx->tx.rec_seq = kmalloc(rec_seq_size, GFP_KERNEL);
+       if (!ctx->tx.rec_seq) {
+               rc = -ENOMEM;
+               goto free_iv;
+       }
+       memcpy(ctx->tx.rec_seq, rec_seq, rec_seq_size);
+
+       rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
+       if (rc)
+               goto free_rec_seq;
+
+       /* start at rec_seq - 1 to account for the start marker record */
+       memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
+       offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
+
+       start_marker_record->end_seq = tcp_sk(sk)->write_seq;
+       start_marker_record->len = 0;
+       start_marker_record->num_frags = 0;
+
+       INIT_LIST_HEAD(&offload_ctx->records_list);
+       list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
+       spin_lock_init(&offload_ctx->lock);
+
+       clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
+       ctx->push_pending_record = tls_device_push_pending_record;
+       offload_ctx->sk_destruct = sk->sk_destruct;
+
+       /* TLS offload is greatly simplified if we don't send
+        * SKBs where only part of the payload needs to be encrypted.
+        * So mark the last skb in the write queue as end of record.
+        */
+       skb = tcp_write_queue_tail(sk);
+       if (skb)
+               TCP_SKB_CB(skb)->eor = 1;
+
+       refcount_set(&ctx->refcount, 1);
+
+       /* We support starting offload on multiple sockets
+        * concurrently, so we only need a read lock here.
+        * This lock must precede get_netdev_for_sock to prevent races between
+        * NETDEV_DOWN and setsockopt.
+        */
+       down_read(&device_offload_lock);
+       netdev = get_netdev_for_sock(sk);
+       if (!netdev) {
+               pr_err_ratelimited("%s: netdev not found\n", __func__);
+               rc = -EINVAL;
+               goto release_lock;
+       }
+
+       if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
+               rc = -ENOTSUPP;
+               goto release_netdev;
+       }
+
+       /* Avoid offloading if the device is down
+        * We don't want to offload new flows after
+        * the NETDEV_DOWN event
+        */
+       if (!(netdev->flags & IFF_UP)) {
+               rc = -EINVAL;
+               goto release_netdev;
+       }
+
+       ctx->priv_ctx_tx = offload_ctx;
+       rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
+                                            &ctx->crypto_send,
+                                            tcp_sk(sk)->write_seq);
+       if (rc)
+               goto release_netdev;
+
+       ctx->netdev = netdev;
+
+       spin_lock_irq(&tls_device_lock);
+       list_add_tail(&ctx->list, &tls_device_list);
+       spin_unlock_irq(&tls_device_lock);
+
+       sk->sk_validate_xmit_skb = tls_validate_xmit_skb;
+       /* following this assignment tls_is_sk_tx_device_offloaded
+        * will return true and the context might be accessed
+        * by the netdev's xmit function.
+        */
+       smp_store_release(&sk->sk_destruct,
+                         &tls_device_sk_destruct);
+       up_read(&device_offload_lock);
+       goto out;
+
+release_netdev:
+       dev_put(netdev);
+release_lock:
+       up_read(&device_offload_lock);
+       clean_acked_data_disable(inet_csk(sk));
+       crypto_free_aead(offload_ctx->aead_send);
+free_rec_seq:
+       kfree(ctx->tx.rec_seq);
+free_iv:
+       kfree(ctx->tx.iv);
+free_offload_ctx:
+       kfree(offload_ctx);
+       ctx->priv_ctx_tx = NULL;
+free_marker_record:
+       kfree(start_marker_record);
+out:
+       return rc;
+}
+
+static int tls_device_down(struct net_device *netdev)
+{
+       struct tls_context *ctx, *tmp;
+       unsigned long flags;
+       LIST_HEAD(list);
+
+       /* Request a write lock to block new offload attempts */
+       down_write(&device_offload_lock);
+
+       spin_lock_irqsave(&tls_device_lock, flags);
+       list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
+               if (ctx->netdev != netdev ||
+                   !refcount_inc_not_zero(&ctx->refcount))
+                       continue;
+
+               list_move(&ctx->list, &list);
+       }
+       spin_unlock_irqrestore(&tls_device_lock, flags);
+
+       list_for_each_entry_safe(ctx, tmp, &list, list) {
+               netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
+                                               TLS_OFFLOAD_CTX_DIR_TX);
+               ctx->netdev = NULL;
+               dev_put(netdev);
+               list_del_init(&ctx->list);
+
+               if (refcount_dec_and_test(&ctx->refcount))
+                       tls_device_free_ctx(ctx);
+       }
+
+       up_write(&device_offload_lock);
+
+       flush_work(&tls_device_gc_work);
+
+       return NOTIFY_DONE;
+}
+
+static int tls_dev_event(struct notifier_block *this, unsigned long event,
+                        void *ptr)
+{
+       struct net_device *dev = netdev_notifier_info_to_dev(ptr);
+
+       if (!(dev->features & NETIF_F_HW_TLS_TX))
+               return NOTIFY_DONE;
+
+       switch (event) {
+       case NETDEV_REGISTER:
+       case NETDEV_FEAT_CHANGE:
+               if  (dev->tlsdev_ops &&
+                    dev->tlsdev_ops->tls_dev_add &&
+                    dev->tlsdev_ops->tls_dev_del)
+                       return NOTIFY_DONE;
+               else
+                       return NOTIFY_BAD;
+       case NETDEV_DOWN:
+               return tls_device_down(dev);
+       }
+       return NOTIFY_DONE;
+}
+
+static struct notifier_block tls_dev_notifier = {
+       .notifier_call  = tls_dev_event,
+};
+
+void __init tls_device_init(void)
+{
+       register_netdevice_notifier(&tls_dev_notifier);
+}
+
+void __exit tls_device_cleanup(void)
+{
+       unregister_netdevice_notifier(&tls_dev_notifier);
+       flush_work(&tls_device_gc_work);
+}
diff --git a/net/tls/tls_device_fallback.c b/net/tls/tls_device_fallback.c
new file mode 100644 (file)
index 0000000..748914a
--- /dev/null
@@ -0,0 +1,450 @@
+/* Copyright (c) 2018, Mellanox Technologies All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <net/tls.h>
+#include <crypto/aead.h>
+#include <crypto/scatterwalk.h>
+#include <net/ip6_checksum.h>
+
+static void chain_to_walk(struct scatterlist *sg, struct scatter_walk *walk)
+{
+       struct scatterlist *src = walk->sg;
+       int diff = walk->offset - src->offset;
+
+       sg_set_page(sg, sg_page(src),
+                   src->length - diff, walk->offset);
+
+       scatterwalk_crypto_chain(sg, sg_next(src), 0, 2);
+}
+
+static int tls_enc_record(struct aead_request *aead_req,
+                         struct crypto_aead *aead, char *aad,
+                         char *iv, __be64 rcd_sn,
+                         struct scatter_walk *in,
+                         struct scatter_walk *out, int *in_len)
+{
+       unsigned char buf[TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE];
+       struct scatterlist sg_in[3];
+       struct scatterlist sg_out[3];
+       u16 len;
+       int rc;
+
+       len = min_t(int, *in_len, ARRAY_SIZE(buf));
+
+       scatterwalk_copychunks(buf, in, len, 0);
+       scatterwalk_copychunks(buf, out, len, 1);
+
+       *in_len -= len;
+       if (!*in_len)
+               return 0;
+
+       scatterwalk_pagedone(in, 0, 1);
+       scatterwalk_pagedone(out, 1, 1);
+
+       len = buf[4] | (buf[3] << 8);
+       len -= TLS_CIPHER_AES_GCM_128_IV_SIZE;
+
+       tls_make_aad(aad, len - TLS_CIPHER_AES_GCM_128_TAG_SIZE,
+                    (char *)&rcd_sn, sizeof(rcd_sn), buf[0]);
+
+       memcpy(iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, buf + TLS_HEADER_SIZE,
+              TLS_CIPHER_AES_GCM_128_IV_SIZE);
+
+       sg_init_table(sg_in, ARRAY_SIZE(sg_in));
+       sg_init_table(sg_out, ARRAY_SIZE(sg_out));
+       sg_set_buf(sg_in, aad, TLS_AAD_SPACE_SIZE);
+       sg_set_buf(sg_out, aad, TLS_AAD_SPACE_SIZE);
+       chain_to_walk(sg_in + 1, in);
+       chain_to_walk(sg_out + 1, out);
+
+       *in_len -= len;
+       if (*in_len < 0) {
+               *in_len += TLS_CIPHER_AES_GCM_128_TAG_SIZE;
+               /* the input buffer doesn't contain the entire record.
+                * trim len accordingly. The resulting authentication tag
+                * will contain garbage, but we don't care, so we won't
+                * include any of it in the output skb
+                * Note that we assume the output buffer length
+                * is larger then input buffer length + tag size
+                */
+               if (*in_len < 0)
+                       len += *in_len;
+
+               *in_len = 0;
+       }
+
+       if (*in_len) {
+               scatterwalk_copychunks(NULL, in, len, 2);
+               scatterwalk_pagedone(in, 0, 1);
+               scatterwalk_copychunks(NULL, out, len, 2);
+               scatterwalk_pagedone(out, 1, 1);
+       }
+
+       len -= TLS_CIPHER_AES_GCM_128_TAG_SIZE;
+       aead_request_set_crypt(aead_req, sg_in, sg_out, len, iv);
+
+       rc = crypto_aead_encrypt(aead_req);
+
+       return rc;
+}
+
+static void tls_init_aead_request(struct aead_request *aead_req,
+                                 struct crypto_aead *aead)
+{
+       aead_request_set_tfm(aead_req, aead);
+       aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
+}
+
+static struct aead_request *tls_alloc_aead_request(struct crypto_aead *aead,
+                                                  gfp_t flags)
+{
+       unsigned int req_size = sizeof(struct aead_request) +
+               crypto_aead_reqsize(aead);
+       struct aead_request *aead_req;
+
+       aead_req = kzalloc(req_size, flags);
+       if (aead_req)
+               tls_init_aead_request(aead_req, aead);
+       return aead_req;
+}
+
+static int tls_enc_records(struct aead_request *aead_req,
+                          struct crypto_aead *aead, struct scatterlist *sg_in,
+                          struct scatterlist *sg_out, char *aad, char *iv,
+                          u64 rcd_sn, int len)
+{
+       struct scatter_walk out, in;
+       int rc;
+
+       scatterwalk_start(&in, sg_in);
+       scatterwalk_start(&out, sg_out);
+
+       do {
+               rc = tls_enc_record(aead_req, aead, aad, iv,
+                                   cpu_to_be64(rcd_sn), &in, &out, &len);
+               rcd_sn++;
+
+       } while (rc == 0 && len);
+
+       scatterwalk_done(&in, 0, 0);
+       scatterwalk_done(&out, 1, 0);
+
+       return rc;
+}
+
+/* Can't use icsk->icsk_af_ops->send_check here because the ip addresses
+ * might have been changed by NAT.
+ */
+static void update_chksum(struct sk_buff *skb, int headln)
+{
+       struct tcphdr *th = tcp_hdr(skb);
+       int datalen = skb->len - headln;
+       const struct ipv6hdr *ipv6h;
+       const struct iphdr *iph;
+
+       /* We only changed the payload so if we are using partial we don't
+        * need to update anything.
+        */
+       if (likely(skb->ip_summed == CHECKSUM_PARTIAL))
+               return;
+
+       skb->ip_summed = CHECKSUM_PARTIAL;
+       skb->csum_start = skb_transport_header(skb) - skb->head;
+       skb->csum_offset = offsetof(struct tcphdr, check);
+
+       if (skb->sk->sk_family == AF_INET6) {
+               ipv6h = ipv6_hdr(skb);
+               th->check = ~csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
+                                            datalen, IPPROTO_TCP, 0);
+       } else {
+               iph = ip_hdr(skb);
+               th->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr, datalen,
+                                              IPPROTO_TCP, 0);
+       }
+}
+
+static void complete_skb(struct sk_buff *nskb, struct sk_buff *skb, int headln)
+{
+       skb_copy_header(nskb, skb);
+
+       skb_put(nskb, skb->len);
+       memcpy(nskb->data, skb->data, headln);
+       update_chksum(nskb, headln);
+
+       nskb->destructor = skb->destructor;
+       nskb->sk = skb->sk;
+       skb->destructor = NULL;
+       skb->sk = NULL;
+       refcount_add(nskb->truesize - skb->truesize,
+                    &nskb->sk->sk_wmem_alloc);
+}
+
+/* This function may be called after the user socket is already
+ * closed so make sure we don't use anything freed during
+ * tls_sk_proto_close here
+ */
+
+static int fill_sg_in(struct scatterlist *sg_in,
+                     struct sk_buff *skb,
+                     struct tls_offload_context *ctx,
+                     u64 *rcd_sn,
+                     s32 *sync_size,
+                     int *resync_sgs)
+{
+       int tcp_payload_offset = skb_transport_offset(skb) + tcp_hdrlen(skb);
+       int payload_len = skb->len - tcp_payload_offset;
+       u32 tcp_seq = ntohl(tcp_hdr(skb)->seq);
+       struct tls_record_info *record;
+       unsigned long flags;
+       int remaining;
+       int i;
+
+       spin_lock_irqsave(&ctx->lock, flags);
+       record = tls_get_record(ctx, tcp_seq, rcd_sn);
+       if (!record) {
+               spin_unlock_irqrestore(&ctx->lock, flags);
+               WARN(1, "Record not found for seq %u\n", tcp_seq);
+               return -EINVAL;
+       }
+
+       *sync_size = tcp_seq - tls_record_start_seq(record);
+       if (*sync_size < 0) {
+               int is_start_marker = tls_record_is_start_marker(record);
+
+               spin_unlock_irqrestore(&ctx->lock, flags);
+               /* This should only occur if the relevant record was
+                * already acked. In that case it should be ok
+                * to drop the packet and avoid retransmission.
+                *
+                * There is a corner case where the packet contains
+                * both an acked and a non-acked record.
+                * We currently don't handle that case and rely
+                * on TCP to retranmit a packet that doesn't contain
+                * already acked payload.
+                */
+               if (!is_start_marker)
+                       *sync_size = 0;
+               return -EINVAL;
+       }
+
+       remaining = *sync_size;
+       for (i = 0; remaining > 0; i++) {
+               skb_frag_t *frag = &record->frags[i];
+
+               __skb_frag_ref(frag);
+               sg_set_page(sg_in + i, skb_frag_page(frag),
+                           skb_frag_size(frag), frag->page_offset);
+
+               remaining -= skb_frag_size(frag);
+
+               if (remaining < 0)
+                       sg_in[i].length += remaining;
+       }
+       *resync_sgs = i;
+
+       spin_unlock_irqrestore(&ctx->lock, flags);
+       if (skb_to_sgvec(skb, &sg_in[i], tcp_payload_offset, payload_len) < 0)
+               return -EINVAL;
+
+       return 0;
+}
+
+static void fill_sg_out(struct scatterlist sg_out[3], void *buf,
+                       struct tls_context *tls_ctx,
+                       struct sk_buff *nskb,
+                       int tcp_payload_offset,
+                       int payload_len,
+                       int sync_size,
+                       void *dummy_buf)
+{
+       sg_set_buf(&sg_out[0], dummy_buf, sync_size);
+       sg_set_buf(&sg_out[1], nskb->data + tcp_payload_offset, payload_len);
+       /* Add room for authentication tag produced by crypto */
+       dummy_buf += sync_size;
+       sg_set_buf(&sg_out[2], dummy_buf, TLS_CIPHER_AES_GCM_128_TAG_SIZE);
+}
+
+static struct sk_buff *tls_enc_skb(struct tls_context *tls_ctx,
+                                  struct scatterlist sg_out[3],
+                                  struct scatterlist *sg_in,
+                                  struct sk_buff *skb,
+                                  s32 sync_size, u64 rcd_sn)
+{
+       int tcp_payload_offset = skb_transport_offset(skb) + tcp_hdrlen(skb);
+       struct tls_offload_context *ctx = tls_offload_ctx(tls_ctx);
+       int payload_len = skb->len - tcp_payload_offset;
+       void *buf, *iv, *aad, *dummy_buf;
+       struct aead_request *aead_req;
+       struct sk_buff *nskb = NULL;
+       int buf_len;
+
+       aead_req = tls_alloc_aead_request(ctx->aead_send, GFP_ATOMIC);
+       if (!aead_req)
+               return NULL;
+
+       buf_len = TLS_CIPHER_AES_GCM_128_SALT_SIZE +
+                 TLS_CIPHER_AES_GCM_128_IV_SIZE +
+                 TLS_AAD_SPACE_SIZE +
+                 sync_size +
+                 TLS_CIPHER_AES_GCM_128_TAG_SIZE;
+       buf = kmalloc(buf_len, GFP_ATOMIC);
+       if (!buf)
+               goto free_req;
+
+       iv = buf;
+       memcpy(iv, tls_ctx->crypto_send_aes_gcm_128.salt,
+              TLS_CIPHER_AES_GCM_128_SALT_SIZE);
+       aad = buf + TLS_CIPHER_AES_GCM_128_SALT_SIZE +
+             TLS_CIPHER_AES_GCM_128_IV_SIZE;
+       dummy_buf = aad + TLS_AAD_SPACE_SIZE;
+
+       nskb = alloc_skb(skb_headroom(skb) + skb->len, GFP_ATOMIC);
+       if (!nskb)
+               goto free_buf;
+
+       skb_reserve(nskb, skb_headroom(skb));
+
+       fill_sg_out(sg_out, buf, tls_ctx, nskb, tcp_payload_offset,
+                   payload_len, sync_size, dummy_buf);
+
+       if (tls_enc_records(aead_req, ctx->aead_send, sg_in, sg_out, aad, iv,
+                           rcd_sn, sync_size + payload_len) < 0)
+               goto free_nskb;
+
+       complete_skb(nskb, skb, tcp_payload_offset);
+
+       /* validate_xmit_skb_list assumes that if the skb wasn't segmented
+        * nskb->prev will point to the skb itself
+        */
+       nskb->prev = nskb;
+
+free_buf:
+       kfree(buf);
+free_req:
+       kfree(aead_req);
+       return nskb;
+free_nskb:
+       kfree_skb(nskb);
+       nskb = NULL;
+       goto free_buf;
+}
+
+static struct sk_buff *tls_sw_fallback(struct sock *sk, struct sk_buff *skb)
+{
+       int tcp_payload_offset = skb_transport_offset(skb) + tcp_hdrlen(skb);
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_offload_context *ctx = tls_offload_ctx(tls_ctx);
+       int payload_len = skb->len - tcp_payload_offset;
+       struct scatterlist *sg_in, sg_out[3];
+       struct sk_buff *nskb = NULL;
+       int sg_in_max_elements;
+       int resync_sgs = 0;
+       s32 sync_size = 0;
+       u64 rcd_sn;
+
+       /* worst case is:
+        * MAX_SKB_FRAGS in tls_record_info
+        * MAX_SKB_FRAGS + 1 in SKB head and frags.
+        */
+       sg_in_max_elements = 2 * MAX_SKB_FRAGS + 1;
+
+       if (!payload_len)
+               return skb;
+
+       sg_in = kmalloc_array(sg_in_max_elements, sizeof(*sg_in), GFP_ATOMIC);
+       if (!sg_in)
+               goto free_orig;
+
+       sg_init_table(sg_in, sg_in_max_elements);
+       sg_init_table(sg_out, ARRAY_SIZE(sg_out));
+
+       if (fill_sg_in(sg_in, skb, ctx, &rcd_sn, &sync_size, &resync_sgs)) {
+               /* bypass packets before kernel TLS socket option was set */
+               if (sync_size < 0 && payload_len <= -sync_size)
+                       nskb = skb_get(skb);
+               goto put_sg;
+       }
+
+       nskb = tls_enc_skb(tls_ctx, sg_out, sg_in, skb, sync_size, rcd_sn);
+
+put_sg:
+       while (resync_sgs)
+               put_page(sg_page(&sg_in[--resync_sgs]));
+       kfree(sg_in);
+free_orig:
+       kfree_skb(skb);
+       return nskb;
+}
+
+struct sk_buff *tls_validate_xmit_skb(struct sock *sk,
+                                     struct net_device *dev,
+                                     struct sk_buff *skb)
+{
+       if (dev == tls_get_ctx(sk)->netdev)
+               return skb;
+
+       return tls_sw_fallback(sk, skb);
+}
+
+int tls_sw_fallback_init(struct sock *sk,
+                        struct tls_offload_context *offload_ctx,
+                        struct tls_crypto_info *crypto_info)
+{
+       const u8 *key;
+       int rc;
+
+       offload_ctx->aead_send =
+           crypto_alloc_aead("gcm(aes)", 0, CRYPTO_ALG_ASYNC);
+       if (IS_ERR(offload_ctx->aead_send)) {
+               rc = PTR_ERR(offload_ctx->aead_send);
+               pr_err_ratelimited("crypto_alloc_aead failed rc=%d\n", rc);
+               offload_ctx->aead_send = NULL;
+               goto err_out;
+       }
+
+       key = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->key;
+
+       rc = crypto_aead_setkey(offload_ctx->aead_send, key,
+                               TLS_CIPHER_AES_GCM_128_KEY_SIZE);
+       if (rc)
+               goto free_aead;
+
+       rc = crypto_aead_setauthsize(offload_ctx->aead_send,
+                                    TLS_CIPHER_AES_GCM_128_TAG_SIZE);
+       if (rc)
+               goto free_aead;
+
+       return 0;
+free_aead:
+       crypto_free_aead(offload_ctx->aead_send);
+err_out:
+       return rc;
+}
index 545bf34ed599d80de96dceab7053508460a75420..3aafb871a0a82a83d5f3c721a7aa84a8ba0d9810 100644 (file)
@@ -54,6 +54,9 @@ enum {
 enum {
        TLS_BASE,
        TLS_SW,
+#ifdef CONFIG_TLS_DEVICE
+       TLS_HW,
+#endif
        TLS_HW_RECORD,
        TLS_NUM_CONFIG,
 };
@@ -280,6 +283,15 @@ static void tls_sk_proto_close(struct sock *sk, long timeout)
                tls_sw_free_resources_rx(sk);
        }
 
+#ifdef CONFIG_TLS_DEVICE
+       if (ctx->tx_conf != TLS_HW) {
+#else
+       {
+#endif
+               kfree(ctx);
+               ctx = NULL;
+       }
+
 skip_tx_cleanup:
        release_sock(sk);
        sk_proto_close(sk, timeout);
@@ -442,8 +454,16 @@ static int do_tls_setsockopt_conf(struct sock *sk, char __user *optval,
        }
 
        if (tx) {
-               rc = tls_set_sw_offload(sk, ctx, 1);
-               conf = TLS_SW;
+#ifdef CONFIG_TLS_DEVICE
+               rc = tls_set_device_offload(sk, ctx);
+               conf = TLS_HW;
+               if (rc) {
+#else
+               {
+#endif
+                       rc = tls_set_sw_offload(sk, ctx, 1);
+                       conf = TLS_SW;
+               }
        } else {
                rc = tls_set_sw_offload(sk, ctx, 0);
                conf = TLS_SW;
@@ -596,6 +616,16 @@ static void build_protos(struct proto prot[TLS_NUM_CONFIG][TLS_NUM_CONFIG],
        prot[TLS_SW][TLS_SW].recvmsg    = tls_sw_recvmsg;
        prot[TLS_SW][TLS_SW].close      = tls_sk_proto_close;
 
+#ifdef CONFIG_TLS_DEVICE
+       prot[TLS_HW][TLS_BASE] = prot[TLS_BASE][TLS_BASE];
+       prot[TLS_HW][TLS_BASE].sendmsg          = tls_device_sendmsg;
+       prot[TLS_HW][TLS_BASE].sendpage         = tls_device_sendpage;
+
+       prot[TLS_HW][TLS_SW] = prot[TLS_BASE][TLS_SW];
+       prot[TLS_HW][TLS_SW].sendmsg            = tls_device_sendmsg;
+       prot[TLS_HW][TLS_SW].sendpage           = tls_device_sendpage;
+#endif
+
        prot[TLS_HW_RECORD][TLS_HW_RECORD] = *base;
        prot[TLS_HW_RECORD][TLS_HW_RECORD].hash         = tls_hw_hash;
        prot[TLS_HW_RECORD][TLS_HW_RECORD].unhash       = tls_hw_unhash;
@@ -630,7 +660,7 @@ static int tls_init(struct sock *sk)
        ctx->getsockopt = sk->sk_prot->getsockopt;
        ctx->sk_proto_close = sk->sk_prot->close;
 
-       /* Build IPv6 TLS whenever the address of tcpv6_prot changes */
+       /* Build IPv6 TLS whenever the address of tcpv6 _prot changes */
        if (ip_ver == TLSV6 &&
            unlikely(sk->sk_prot != smp_load_acquire(&saved_tcpv6_prot))) {
                mutex_lock(&tcpv6_prot_mutex);
@@ -680,6 +710,9 @@ static int __init tls_register(void)
        tls_sw_proto_ops.poll = tls_sw_poll;
        tls_sw_proto_ops.splice_read = tls_sw_splice_read;
 
+#ifdef CONFIG_TLS_DEVICE
+       tls_device_init();
+#endif
        tcp_register_ulp(&tcp_tls_ulp_ops);
 
        return 0;
@@ -688,6 +721,9 @@ static int __init tls_register(void)
 static void __exit tls_unregister(void)
 {
        tcp_unregister_ulp(&tcp_tls_ulp_ops);
+#ifdef CONFIG_TLS_DEVICE
+       tls_device_cleanup();
+#endif
 }
 
 module_init(tls_register);