net/tls: Add support for async encryption of records for performance
authorVakul Garg <vakul.garg@nxp.com>
Fri, 21 Sep 2018 04:16:13 +0000 (09:46 +0530)
committerDavid S. Miller <davem@davemloft.net>
Sat, 22 Sep 2018 02:17:34 +0000 (19:17 -0700)
In current implementation, tls records are encrypted & transmitted
serially. Till the time the previously submitted user data is encrypted,
the implementation waits and on finish starts transmitting the record.
This approach of encrypt-one record at a time is inefficient when
asynchronous crypto accelerators are used. For each record, there are
overheads of interrupts, driver softIRQ scheduling etc. Also the crypto
accelerator sits idle most of time while an encrypted record's pages are
handed over to tcp stack for transmission.

This patch enables encryption of multiple records in parallel when an
async capable crypto accelerator is present in system. This is achieved
by allowing the user space application to send more data using sendmsg()
even while previously issued data is being processed by crypto
accelerator. This requires returning the control back to user space
application after submitting encryption request to accelerator. This
also means that zero-copy mode of encryption cannot be used with async
accelerator as we must be done with user space application buffer before
returning from sendmsg().

There can be multiple records in flight to/from the accelerator. Each of
the record is represented by 'struct tls_rec'. This is used to store the
memory pages for the record.

After the records are encrypted, they are added in a linked list called
tx_ready_list which contains encrypted tls records sorted as per tls
sequence number. The records from tx_ready_list are transmitted using a
newly introduced function called tls_tx_records(). The tx_ready_list is
polled for any record ready to be transmitted in sendmsg(), sendpage()
after initiating encryption of new tls records. This achieves parallel
encryption and transmission of records when async accelerator is
present.

There could be situation when crypto accelerator completes encryption
later than polling of tx_ready_list by sendmsg()/sendpage(). Therefore
we need a deferred work context to be able to transmit records from
tx_ready_list. The deferred work context gets scheduled if applications
are not sending much data through the socket. If the applications issue
sendmsg()/sendpage() in quick succession, then the scheduling of
tx_work_handler gets cancelled as the tx_ready_list would be polled from
application's context itself. This saves scheduling overhead of deferred
work.

The patch also brings some side benefit. We are able to get rid of the
concept of CLOSED record. This is because the records once closed are
either encrypted and then placed into tx_ready_list or if encryption
fails, the socket error is set. This simplifies the kernel tls
sendpath. However since tls_device.c is still using macros, accessory
functions for CLOSED records have been retained.

Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tls.h
net/tls/tls_main.c
net/tls/tls_sw.c

index 9f3c4ea9ad6f1b04ccfa2fe0657c356e9cf54123..3aa73e2d8823581ba6cf6ad0d69ce6eb9184f71e 100644 (file)
@@ -41,7 +41,7 @@
 #include <linux/tcp.h>
 #include <net/tcp.h>
 #include <net/strparser.h>
-
+#include <crypto/aead.h>
 #include <uapi/linux/tls.h>
 
 
@@ -93,24 +93,47 @@ enum {
        TLS_NUM_CONFIG,
 };
 
-struct tls_sw_context_tx {
-       struct crypto_aead *aead_send;
-       struct crypto_wait async_wait;
-
-       char aad_space[TLS_AAD_SPACE_SIZE];
-
-       unsigned int sg_plaintext_size;
-       int sg_plaintext_num_elem;
+/* TLS records are maintained in 'struct tls_rec'. It stores the memory pages
+ * allocated or mapped for each TLS record. After encryption, the records are
+ * stores in a linked list.
+ */
+struct tls_rec {
+       struct list_head list;
+       int tx_flags;
        struct scatterlist sg_plaintext_data[MAX_SKB_FRAGS];
-
-       unsigned int sg_encrypted_size;
-       int sg_encrypted_num_elem;
        struct scatterlist sg_encrypted_data[MAX_SKB_FRAGS];
 
        /* AAD | sg_plaintext_data | sg_tag */
        struct scatterlist sg_aead_in[2];
        /* AAD | sg_encrypted_data (data contain overhead for hdr&iv&tag) */
        struct scatterlist sg_aead_out[2];
+
+       unsigned int sg_plaintext_size;
+       unsigned int sg_encrypted_size;
+       int sg_plaintext_num_elem;
+       int sg_encrypted_num_elem;
+
+       char aad_space[TLS_AAD_SPACE_SIZE];
+       struct aead_request aead_req;
+       u8 aead_req_ctx[];
+};
+
+struct tx_work {
+       struct delayed_work work;
+       struct sock *sk;
+};
+
+struct tls_sw_context_tx {
+       struct crypto_aead *aead_send;
+       struct crypto_wait async_wait;
+       struct tx_work tx_work;
+       struct tls_rec *open_rec;
+       struct list_head tx_ready_list;
+       atomic_t encrypt_pending;
+       int async_notify;
+
+#define BIT_TX_SCHEDULED       0
+       unsigned long tx_bitmask;
 };
 
 struct tls_sw_context_rx {
@@ -197,6 +220,8 @@ struct tls_context {
 
        struct scatterlist *partially_sent_record;
        u16 partially_sent_offset;
+       u64 tx_seq_number;      /* Next TLS seqnum to be transmitted */
+
        unsigned long flags;
        bool in_tcp_sendpages;
 
@@ -261,6 +286,7 @@ int tls_device_sendpage(struct sock *sk, struct page *page,
 void tls_device_sk_destruct(struct sock *sk);
 void tls_device_init(void);
 void tls_device_cleanup(void);
+int tls_tx_records(struct sock *sk, int flags);
 
 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
                                       u32 seq, u64 *p_record_sn);
@@ -279,6 +305,9 @@ 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);
+int tls_push_partial_record(struct sock *sk, struct tls_context *ctx,
+                           int flags);
+
 int tls_push_pending_closed_record(struct sock *sk, struct tls_context *ctx,
                                   int flags, long *timeo);
 
@@ -312,6 +341,23 @@ static inline bool tls_is_pending_open_record(struct tls_context *tls_ctx)
        return tls_ctx->pending_open_record_frags;
 }
 
+static inline bool is_tx_ready(struct tls_context *tls_ctx,
+                              struct tls_sw_context_tx *ctx)
+{
+       struct tls_rec *rec;
+       u64 seq;
+
+       rec = list_first_entry(&ctx->tx_ready_list, struct tls_rec, list);
+       if (!rec)
+               return false;
+
+       seq = be64_to_cpup((const __be64 *)&rec->aad_space);
+       if (seq == tls_ctx->tx_seq_number)
+               return true;
+       else
+               return false;
+}
+
 struct sk_buff *
 tls_validate_xmit_skb(struct sock *sk, struct net_device *dev,
                      struct sk_buff *skb);
index 523622dc74f8b969113b0435b39f5d0f3d070304..06094de7a3d92305a1b2af247018af1719dec075 100644 (file)
@@ -141,7 +141,6 @@ retry:
                size = sg->length;
        }
 
-       clear_bit(TLS_PENDING_CLOSED_RECORD, &ctx->flags);
        ctx->in_tcp_sendpages = false;
        ctx->sk_write_space(sk);
 
@@ -193,15 +192,12 @@ int tls_proccess_cmsg(struct sock *sk, struct msghdr *msg,
        return rc;
 }
 
-int tls_push_pending_closed_record(struct sock *sk, struct tls_context *ctx,
-                                  int flags, long *timeo)
+int tls_push_partial_record(struct sock *sk, struct tls_context *ctx,
+                           int flags)
 {
        struct scatterlist *sg;
        u16 offset;
 
-       if (!tls_is_partially_sent_record(ctx))
-               return ctx->push_pending_record(sk, flags);
-
        sg = ctx->partially_sent_record;
        offset = ctx->partially_sent_offset;
 
@@ -209,9 +205,23 @@ int tls_push_pending_closed_record(struct sock *sk, struct tls_context *ctx,
        return tls_push_sg(sk, ctx, sg, offset, flags);
 }
 
+int tls_push_pending_closed_record(struct sock *sk,
+                                  struct tls_context *tls_ctx,
+                                  int flags, long *timeo)
+{
+       struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+
+       if (tls_is_partially_sent_record(tls_ctx) ||
+           !list_empty(&ctx->tx_ready_list))
+               return tls_tx_records(sk, flags);
+       else
+               return tls_ctx->push_pending_record(sk, flags);
+}
+
 static void tls_write_space(struct sock *sk)
 {
        struct tls_context *ctx = tls_get_ctx(sk);
+       struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
 
        /* If in_tcp_sendpages call lower protocol write space handler
         * to ensure we wake up any waiting operations there. For example
@@ -222,20 +232,11 @@ static void tls_write_space(struct sock *sk)
                return;
        }
 
-       if (!sk->sk_write_pending && tls_is_pending_closed_record(ctx)) {
-               gfp_t sk_allocation = sk->sk_allocation;
-               int rc;
-               long timeo = 0;
-
-               sk->sk_allocation = GFP_ATOMIC;
-               rc = tls_push_pending_closed_record(sk, ctx,
-                                                   MSG_DONTWAIT |
-                                                   MSG_NOSIGNAL,
-                                                   &timeo);
-               sk->sk_allocation = sk_allocation;
-
-               if (rc < 0)
-                       return;
+       /* Schedule the transmission if tx list is ready */
+       if (is_tx_ready(ctx, tx_ctx) && !sk->sk_write_pending) {
+               /* Schedule the transmission */
+               if (!test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
+                       schedule_delayed_work(&tx_ctx->tx_work.work, 0);
        }
 
        ctx->sk_write_space(sk);
@@ -270,19 +271,6 @@ static void tls_sk_proto_close(struct sock *sk, long timeout)
        if (!tls_complete_pending_work(sk, ctx, 0, &timeo))
                tls_handle_open_record(sk, 0);
 
-       if (ctx->partially_sent_record) {
-               struct scatterlist *sg = ctx->partially_sent_record;
-
-               while (1) {
-                       put_page(sg_page(sg));
-                       sk_mem_uncharge(sk, sg->length);
-
-                       if (sg_is_last(sg))
-                               break;
-                       sg++;
-               }
-       }
-
        /* We need these for tls_sw_fallback handling of other packets */
        if (ctx->tx_conf == TLS_SW) {
                kfree(ctx->tx.rec_seq);
index 5ff51bac8b469903b70ec9304ef06cf60d265ed0..bcb24c498b8417638d35fc8b1e0e72cb5eb7ab26 100644 (file)
@@ -246,18 +246,19 @@ static void trim_both_sgl(struct sock *sk, int target_size)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec = ctx->open_rec;
 
-       trim_sg(sk, ctx->sg_plaintext_data,
-               &ctx->sg_plaintext_num_elem,
-               &ctx->sg_plaintext_size,
+       trim_sg(sk, rec->sg_plaintext_data,
+               &rec->sg_plaintext_num_elem,
+               &rec->sg_plaintext_size,
                target_size);
 
        if (target_size > 0)
                target_size += tls_ctx->tx.overhead_size;
 
-       trim_sg(sk, ctx->sg_encrypted_data,
-               &ctx->sg_encrypted_num_elem,
-               &ctx->sg_encrypted_size,
+       trim_sg(sk, rec->sg_encrypted_data,
+               &rec->sg_encrypted_num_elem,
+               &rec->sg_encrypted_size,
                target_size);
 }
 
@@ -265,15 +266,16 @@ static int alloc_encrypted_sg(struct sock *sk, int len)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec = ctx->open_rec;
        int rc = 0;
 
        rc = sk_alloc_sg(sk, len,
-                        ctx->sg_encrypted_data, 0,
-                        &ctx->sg_encrypted_num_elem,
-                        &ctx->sg_encrypted_size, 0);
+                        rec->sg_encrypted_data, 0,
+                        &rec->sg_encrypted_num_elem,
+                        &rec->sg_encrypted_size, 0);
 
        if (rc == -ENOSPC)
-               ctx->sg_encrypted_num_elem = ARRAY_SIZE(ctx->sg_encrypted_data);
+               rec->sg_encrypted_num_elem = ARRAY_SIZE(rec->sg_encrypted_data);
 
        return rc;
 }
@@ -282,14 +284,15 @@ static int alloc_plaintext_sg(struct sock *sk, int len)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec = ctx->open_rec;
        int rc = 0;
 
-       rc = sk_alloc_sg(sk, len, ctx->sg_plaintext_data, 0,
-                        &ctx->sg_plaintext_num_elem, &ctx->sg_plaintext_size,
+       rc = sk_alloc_sg(sk, len, rec->sg_plaintext_data, 0,
+                        &rec->sg_plaintext_num_elem, &rec->sg_plaintext_size,
                         tls_ctx->pending_open_record_frags);
 
        if (rc == -ENOSPC)
-               ctx->sg_plaintext_num_elem = ARRAY_SIZE(ctx->sg_plaintext_data);
+               rec->sg_plaintext_num_elem = ARRAY_SIZE(rec->sg_plaintext_data);
 
        return rc;
 }
@@ -311,37 +314,192 @@ static void tls_free_both_sg(struct sock *sk)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec = ctx->open_rec;
 
-       free_sg(sk, ctx->sg_encrypted_data, &ctx->sg_encrypted_num_elem,
-               &ctx->sg_encrypted_size);
+       /* Return if there is no open record */
+       if (!rec)
+               return;
+
+       free_sg(sk, rec->sg_encrypted_data,
+               &rec->sg_encrypted_num_elem,
+               &rec->sg_encrypted_size);
+
+       free_sg(sk, rec->sg_plaintext_data,
+               &rec->sg_plaintext_num_elem,
+               &rec->sg_plaintext_size);
+}
+
+static bool append_tx_ready_list(struct tls_context *tls_ctx,
+                                struct tls_sw_context_tx *ctx,
+                                struct tls_rec *enc_rec)
+{
+       u64 new_seq = be64_to_cpup((const __be64 *)&enc_rec->aad_space);
+       struct list_head *pos;
+
+       /* Need to insert encrypted record in tx_ready_list sorted
+        * as per sequence number. Traverse linked list from tail.
+        */
+       list_for_each_prev(pos, &ctx->tx_ready_list) {
+               struct tls_rec *rec = (struct tls_rec *)pos;
+               u64 seq = be64_to_cpup((const __be64 *)&rec->aad_space);
+
+               if (new_seq > seq)
+                       break;
+       }
+
+       list_add((struct list_head *)&enc_rec->list, pos);
+
+       return is_tx_ready(tls_ctx, ctx);
+}
+
+int tls_tx_records(struct sock *sk, int flags)
+{
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec, *tmp;
+       int tx_flags, rc = 0;
+
+       if (tls_is_partially_sent_record(tls_ctx)) {
+               rec = list_first_entry(&ctx->tx_ready_list,
+                                      struct tls_rec, list);
+
+               if (flags == -1)
+                       tx_flags = rec->tx_flags;
+               else
+                       tx_flags = flags;
+
+               rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
+               if (rc)
+                       goto tx_err;
+
+               /* Full record has been transmitted.
+                * Remove the head of tx_ready_list
+                */
+               tls_ctx->tx_seq_number++;
+               list_del(&rec->list);
+               kfree(rec);
+       }
+
+       /* Tx all ready records which have expected sequence number */
+       list_for_each_entry_safe(rec, tmp, &ctx->tx_ready_list, list) {
+               u64 seq = be64_to_cpup((const __be64 *)&rec->aad_space);
+
+               if (seq == tls_ctx->tx_seq_number) {
+                       if (flags == -1)
+                               tx_flags = rec->tx_flags;
+                       else
+                               tx_flags = flags;
+
+                       rc = tls_push_sg(sk, tls_ctx,
+                                        &rec->sg_encrypted_data[0],
+                                        0, tx_flags);
+                       if (rc)
+                               goto tx_err;
+
+                       tls_ctx->tx_seq_number++;
+                       list_del(&rec->list);
+                       kfree(rec);
+               } else {
+                       break;
+               }
+       }
+
+tx_err:
+       if (rc < 0 && rc != -EAGAIN)
+               tls_err_abort(sk, EBADMSG);
+
+       return rc;
+}
+
+static void tls_encrypt_done(struct crypto_async_request *req, int err)
+{
+       struct aead_request *aead_req = (struct aead_request *)req;
+       struct sock *sk = req->data;
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec;
+       bool ready = false;
+       int pending;
+
+       rec = container_of(aead_req, struct tls_rec, aead_req);
+
+       rec->sg_encrypted_data[0].offset -= tls_ctx->tx.prepend_size;
+       rec->sg_encrypted_data[0].length += tls_ctx->tx.prepend_size;
+
+       free_sg(sk, rec->sg_plaintext_data,
+               &rec->sg_plaintext_num_elem, &rec->sg_plaintext_size);
+
+       /* Free the record if error is previously set on socket */
+       if (err || sk->sk_err) {
+               free_sg(sk, rec->sg_encrypted_data,
+                       &rec->sg_encrypted_num_elem, &rec->sg_encrypted_size);
 
-       free_sg(sk, ctx->sg_plaintext_data, &ctx->sg_plaintext_num_elem,
-               &ctx->sg_plaintext_size);
+               kfree(rec);
+               rec = NULL;
+
+               /* If err is already set on socket, return the same code */
+               if (sk->sk_err) {
+                       ctx->async_wait.err = sk->sk_err;
+               } else {
+                       ctx->async_wait.err = err;
+                       tls_err_abort(sk, err);
+               }
+       }
+
+       /* Append the record in tx queue */
+       if (rec)
+               ready = append_tx_ready_list(tls_ctx, ctx, rec);
+
+       pending = atomic_dec_return(&ctx->encrypt_pending);
+
+       if (!pending && READ_ONCE(ctx->async_notify))
+               complete(&ctx->async_wait.completion);
+
+       if (!ready)
+               return;
+
+       /* Schedule the transmission */
+       if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
+               schedule_delayed_work(&ctx->tx_work.work, 1);
 }
 
-static int tls_do_encryption(struct tls_context *tls_ctx,
+static int tls_do_encryption(struct sock *sk,
+                            struct tls_context *tls_ctx,
                             struct tls_sw_context_tx *ctx,
                             struct aead_request *aead_req,
                             size_t data_len)
 {
+       struct tls_rec *rec = ctx->open_rec;
        int rc;
 
-       ctx->sg_encrypted_data[0].offset += tls_ctx->tx.prepend_size;
-       ctx->sg_encrypted_data[0].length -= tls_ctx->tx.prepend_size;
+       rec->sg_encrypted_data[0].offset += tls_ctx->tx.prepend_size;
+       rec->sg_encrypted_data[0].length -= tls_ctx->tx.prepend_size;
 
        aead_request_set_tfm(aead_req, ctx->aead_send);
        aead_request_set_ad(aead_req, TLS_AAD_SPACE_SIZE);
-       aead_request_set_crypt(aead_req, ctx->sg_aead_in, ctx->sg_aead_out,
+       aead_request_set_crypt(aead_req, rec->sg_aead_in,
+                              rec->sg_aead_out,
                               data_len, tls_ctx->tx.iv);
 
        aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
-                                 crypto_req_done, &ctx->async_wait);
+                                 tls_encrypt_done, sk);
+
+       atomic_inc(&ctx->encrypt_pending);
 
-       rc = crypto_wait_req(crypto_aead_encrypt(aead_req), &ctx->async_wait);
+       rc = crypto_aead_encrypt(aead_req);
+       if (!rc || rc != -EINPROGRESS) {
+               atomic_dec(&ctx->encrypt_pending);
+               rec->sg_encrypted_data[0].offset -= tls_ctx->tx.prepend_size;
+               rec->sg_encrypted_data[0].length += tls_ctx->tx.prepend_size;
+       }
 
-       ctx->sg_encrypted_data[0].offset -= tls_ctx->tx.prepend_size;
-       ctx->sg_encrypted_data[0].length += tls_ctx->tx.prepend_size;
+       /* Case of encryption failure */
+       if (rc && rc != -EINPROGRESS)
+               return rc;
 
+       /* Unhook the record from context if encryption is not failure */
+       ctx->open_rec = NULL;
+       tls_advance_record_sn(sk, &tls_ctx->tx);
        return rc;
 }
 
@@ -350,53 +508,49 @@ static int tls_push_record(struct sock *sk, int flags,
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec = ctx->open_rec;
        struct aead_request *req;
        int rc;
 
-       req = aead_request_alloc(ctx->aead_send, sk->sk_allocation);
-       if (!req)
-               return -ENOMEM;
+       if (!rec)
+               return 0;
 
-       sg_mark_end(ctx->sg_plaintext_data + ctx->sg_plaintext_num_elem - 1);
-       sg_mark_end(ctx->sg_encrypted_data + ctx->sg_encrypted_num_elem - 1);
+       rec->tx_flags = flags;
+       req = &rec->aead_req;
 
-       tls_make_aad(ctx->aad_space, ctx->sg_plaintext_size,
+       sg_mark_end(rec->sg_plaintext_data + rec->sg_plaintext_num_elem - 1);
+       sg_mark_end(rec->sg_encrypted_data + rec->sg_encrypted_num_elem - 1);
+
+       tls_make_aad(rec->aad_space, rec->sg_plaintext_size,
                     tls_ctx->tx.rec_seq, tls_ctx->tx.rec_seq_size,
                     record_type);
 
        tls_fill_prepend(tls_ctx,
-                        page_address(sg_page(&ctx->sg_encrypted_data[0])) +
-                        ctx->sg_encrypted_data[0].offset,
-                        ctx->sg_plaintext_size, record_type);
+                        page_address(sg_page(&rec->sg_encrypted_data[0])) +
+                        rec->sg_encrypted_data[0].offset,
+                        rec->sg_plaintext_size, record_type);
 
        tls_ctx->pending_open_record_frags = 0;
-       set_bit(TLS_PENDING_CLOSED_RECORD, &tls_ctx->flags);
-
-       rc = tls_do_encryption(tls_ctx, ctx, req, ctx->sg_plaintext_size);
-       if (rc < 0) {
-               /* If we are called from write_space and
-                * we fail, we need to set this SOCK_NOSPACE
-                * to trigger another write_space in the future.
-                */
-               set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
-               goto out_req;
-       }
 
-       free_sg(sk, ctx->sg_plaintext_data, &ctx->sg_plaintext_num_elem,
-               &ctx->sg_plaintext_size);
+       rc = tls_do_encryption(sk, tls_ctx, ctx, req, rec->sg_plaintext_size);
+       if (rc == -EINPROGRESS)
+               return -EINPROGRESS;
 
-       ctx->sg_encrypted_num_elem = 0;
-       ctx->sg_encrypted_size = 0;
+       free_sg(sk, rec->sg_plaintext_data, &rec->sg_plaintext_num_elem,
+               &rec->sg_plaintext_size);
 
-       /* Only pass through MSG_DONTWAIT and MSG_NOSIGNAL flags */
-       rc = tls_push_sg(sk, tls_ctx, ctx->sg_encrypted_data, 0, flags);
-       if (rc < 0 && rc != -EAGAIN)
+       if (rc < 0) {
                tls_err_abort(sk, EBADMSG);
+               return rc;
+       }
 
-       tls_advance_record_sn(sk, &tls_ctx->tx);
-out_req:
-       aead_request_free(req);
-       return rc;
+       /* Put the record in tx_ready_list and start tx if permitted.
+        * This happens only when encryption is not asynchronous.
+        */
+       if (append_tx_ready_list(tls_ctx, ctx, rec))
+               return tls_tx_records(sk, flags);
+
+       return 0;
 }
 
 static int tls_sw_push_pending_record(struct sock *sk, int flags)
@@ -473,11 +627,12 @@ static int memcopy_from_iter(struct sock *sk, struct iov_iter *from,
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
-       struct scatterlist *sg = ctx->sg_plaintext_data;
+       struct tls_rec *rec = ctx->open_rec;
+       struct scatterlist *sg = rec->sg_plaintext_data;
        int copy, i, rc = 0;
 
        for (i = tls_ctx->pending_open_record_frags;
-            i < ctx->sg_plaintext_num_elem; ++i) {
+            i < rec->sg_plaintext_num_elem; ++i) {
                copy = sg[i].length;
                if (copy_from_iter(
                                page_address(sg_page(&sg[i])) + sg[i].offset,
@@ -497,34 +652,85 @@ out:
        return rc;
 }
 
-int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
+struct tls_rec *get_rec(struct sock *sk)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
-       int ret;
-       int required_size;
+       struct tls_rec *rec;
+       int mem_size;
+
+       /* Return if we already have an open record */
+       if (ctx->open_rec)
+               return ctx->open_rec;
+
+       mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
+
+       rec = kzalloc(mem_size, sk->sk_allocation);
+       if (!rec)
+               return NULL;
+
+       sg_init_table(&rec->sg_plaintext_data[0],
+                     ARRAY_SIZE(rec->sg_plaintext_data));
+       sg_init_table(&rec->sg_encrypted_data[0],
+                     ARRAY_SIZE(rec->sg_encrypted_data));
+
+       sg_init_table(rec->sg_aead_in, 2);
+       sg_set_buf(&rec->sg_aead_in[0], rec->aad_space,
+                  sizeof(rec->aad_space));
+       sg_unmark_end(&rec->sg_aead_in[1]);
+       sg_chain(rec->sg_aead_in, 2, rec->sg_plaintext_data);
+
+       sg_init_table(rec->sg_aead_out, 2);
+       sg_set_buf(&rec->sg_aead_out[0], rec->aad_space,
+                  sizeof(rec->aad_space));
+       sg_unmark_end(&rec->sg_aead_out[1]);
+       sg_chain(rec->sg_aead_out, 2, rec->sg_encrypted_data);
+
+       ctx->open_rec = rec;
+
+       return rec;
+}
+
+int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
+{
        long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct crypto_tfm *tfm = crypto_aead_tfm(ctx->aead_send);
+       bool async_capable = tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC;
+       unsigned char record_type = TLS_RECORD_TYPE_DATA;
+       bool is_kvec = msg->msg_iter.type & ITER_KVEC;
        bool eor = !(msg->msg_flags & MSG_MORE);
        size_t try_to_copy, copied = 0;
-       unsigned char record_type = TLS_RECORD_TYPE_DATA;
-       int record_room;
+       struct tls_rec *rec;
+       int required_size;
+       int num_async = 0;
        bool full_record;
+       int record_room;
+       int num_zc = 0;
        int orig_size;
-       bool is_kvec = msg->msg_iter.type & ITER_KVEC;
+       int ret;
 
        if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
                return -ENOTSUPP;
 
        lock_sock(sk);
 
-       ret = tls_complete_pending_work(sk, tls_ctx, msg->msg_flags, &timeo);
-       if (ret)
-               goto send_end;
+       /* Wait till there is any pending write on socket */
+       if (unlikely(sk->sk_write_pending)) {
+               ret = wait_on_pending_writer(sk, &timeo);
+               if (unlikely(ret))
+                       goto send_end;
+       }
 
        if (unlikely(msg->msg_controllen)) {
                ret = tls_proccess_cmsg(sk, msg, &record_type);
-               if (ret)
-                       goto send_end;
+               if (ret) {
+                       if (ret == -EINPROGRESS)
+                               num_async++;
+                       else if (ret != -EAGAIN)
+                               goto send_end;
+               }
        }
 
        while (msg_data_left(msg)) {
@@ -533,20 +739,27 @@ int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
                        goto send_end;
                }
 
-               orig_size = ctx->sg_plaintext_size;
+               rec = get_rec(sk);
+               if (!rec) {
+                       ret = -ENOMEM;
+                       goto send_end;
+               }
+
+               orig_size = rec->sg_plaintext_size;
                full_record = false;
                try_to_copy = msg_data_left(msg);
-               record_room = TLS_MAX_PAYLOAD_SIZE - ctx->sg_plaintext_size;
+               record_room = TLS_MAX_PAYLOAD_SIZE - rec->sg_plaintext_size;
                if (try_to_copy >= record_room) {
                        try_to_copy = record_room;
                        full_record = true;
                }
 
-               required_size = ctx->sg_plaintext_size + try_to_copy +
+               required_size = rec->sg_plaintext_size + try_to_copy +
                                tls_ctx->tx.overhead_size;
 
                if (!sk_stream_memory_free(sk))
                        goto wait_for_sndbuf;
+
 alloc_encrypted:
                ret = alloc_encrypted_sg(sk, required_size);
                if (ret) {
@@ -557,33 +770,39 @@ alloc_encrypted:
                         * actually allocated. The difference is due
                         * to max sg elements limit
                         */
-                       try_to_copy -= required_size - ctx->sg_encrypted_size;
+                       try_to_copy -= required_size - rec->sg_encrypted_size;
                        full_record = true;
                }
-               if (!is_kvec && (full_record || eor)) {
+
+               if (!is_kvec && (full_record || eor) && !async_capable) {
                        ret = zerocopy_from_iter(sk, &msg->msg_iter,
-                               try_to_copy, &ctx->sg_plaintext_num_elem,
-                               &ctx->sg_plaintext_size,
-                               ctx->sg_plaintext_data,
-                               ARRAY_SIZE(ctx->sg_plaintext_data),
+                               try_to_copy, &rec->sg_plaintext_num_elem,
+                               &rec->sg_plaintext_size,
+                               rec->sg_plaintext_data,
+                               ARRAY_SIZE(rec->sg_plaintext_data),
                                true);
                        if (ret)
                                goto fallback_to_reg_send;
 
+                       num_zc++;
                        copied += try_to_copy;
                        ret = tls_push_record(sk, msg->msg_flags, record_type);
-                       if (ret)
-                               goto send_end;
+                       if (ret) {
+                               if (ret == -EINPROGRESS)
+                                       num_async++;
+                               else if (ret != -EAGAIN)
+                                       goto send_end;
+                       }
                        continue;
 
 fallback_to_reg_send:
-                       trim_sg(sk, ctx->sg_plaintext_data,
-                               &ctx->sg_plaintext_num_elem,
-                               &ctx->sg_plaintext_size,
+                       trim_sg(sk, rec->sg_plaintext_data,
+                               &rec->sg_plaintext_num_elem,
+                               &rec->sg_plaintext_size,
                                orig_size);
                }
 
-               required_size = ctx->sg_plaintext_size + try_to_copy;
+               required_size = rec->sg_plaintext_size + try_to_copy;
 alloc_plaintext:
                ret = alloc_plaintext_sg(sk, required_size);
                if (ret) {
@@ -594,13 +813,13 @@ alloc_plaintext:
                         * actually allocated. The difference is due
                         * to max sg elements limit
                         */
-                       try_to_copy -= required_size - ctx->sg_plaintext_size;
+                       try_to_copy -= required_size - rec->sg_plaintext_size;
                        full_record = true;
 
-                       trim_sg(sk, ctx->sg_encrypted_data,
-                               &ctx->sg_encrypted_num_elem,
-                               &ctx->sg_encrypted_size,
-                               ctx->sg_plaintext_size +
+                       trim_sg(sk, rec->sg_encrypted_data,
+                               &rec->sg_encrypted_num_elem,
+                               &rec->sg_encrypted_size,
+                               rec->sg_plaintext_size +
                                tls_ctx->tx.overhead_size);
                }
 
@@ -610,13 +829,12 @@ alloc_plaintext:
 
                copied += try_to_copy;
                if (full_record || eor) {
-push_record:
                        ret = tls_push_record(sk, msg->msg_flags, record_type);
                        if (ret) {
-                               if (ret == -ENOMEM)
-                                       goto wait_for_memory;
-
-                               goto send_end;
+                               if (ret == -EINPROGRESS)
+                                       num_async++;
+                               else if (ret != -EAGAIN)
+                                       goto send_end;
                        }
                }
 
@@ -632,15 +850,37 @@ trim_sgl:
                        goto send_end;
                }
 
-               if (tls_is_pending_closed_record(tls_ctx))
-                       goto push_record;
-
-               if (ctx->sg_encrypted_size < required_size)
+               if (rec->sg_encrypted_size < required_size)
                        goto alloc_encrypted;
 
                goto alloc_plaintext;
        }
 
+       if (!num_async) {
+               goto send_end;
+       } else if (num_zc) {
+               /* Wait for pending encryptions to get completed */
+               smp_store_mb(ctx->async_notify, true);
+
+               if (atomic_read(&ctx->encrypt_pending))
+                       crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
+               else
+                       reinit_completion(&ctx->async_wait.completion);
+
+               WRITE_ONCE(ctx->async_notify, false);
+
+               if (ctx->async_wait.err) {
+                       ret = ctx->async_wait.err;
+                       copied = 0;
+               }
+       }
+
+       /* Transmit if any encryptions have completed */
+       if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
+               cancel_delayed_work(&ctx->tx_work.work);
+               tls_tx_records(sk, msg->msg_flags);
+       }
+
 send_end:
        ret = sk_stream_error(sk, msg->msg_flags, ret);
 
@@ -651,16 +891,18 @@ send_end:
 int tls_sw_sendpage(struct sock *sk, struct page *page,
                    int offset, size_t size, int flags)
 {
+       long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
-       int ret;
-       long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
-       bool eor;
-       size_t orig_size = size;
        unsigned char record_type = TLS_RECORD_TYPE_DATA;
+       size_t orig_size = size;
        struct scatterlist *sg;
+       struct tls_rec *rec;
+       int num_async = 0;
        bool full_record;
        int record_room;
+       bool eor;
+       int ret;
 
        if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
                      MSG_SENDPAGE_NOTLAST))
@@ -673,9 +915,12 @@ int tls_sw_sendpage(struct sock *sk, struct page *page,
 
        sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 
-       ret = tls_complete_pending_work(sk, tls_ctx, flags, &timeo);
-       if (ret)
-               goto sendpage_end;
+       /* Wait till there is any pending write on socket */
+       if (unlikely(sk->sk_write_pending)) {
+               ret = wait_on_pending_writer(sk, &timeo);
+               if (unlikely(ret))
+                       goto sendpage_end;
+       }
 
        /* Call the sk_stream functions to manage the sndbuf mem. */
        while (size > 0) {
@@ -686,14 +931,20 @@ int tls_sw_sendpage(struct sock *sk, struct page *page,
                        goto sendpage_end;
                }
 
+               rec = get_rec(sk);
+               if (!rec) {
+                       ret = -ENOMEM;
+                       goto sendpage_end;
+               }
+
                full_record = false;
-               record_room = TLS_MAX_PAYLOAD_SIZE - ctx->sg_plaintext_size;
+               record_room = TLS_MAX_PAYLOAD_SIZE - rec->sg_plaintext_size;
                copy = size;
                if (copy >= record_room) {
                        copy = record_room;
                        full_record = true;
                }
-               required_size = ctx->sg_plaintext_size + copy +
+               required_size = rec->sg_plaintext_size + copy +
                              tls_ctx->tx.overhead_size;
 
                if (!sk_stream_memory_free(sk))
@@ -708,33 +959,32 @@ alloc_payload:
                         * actually allocated. The difference is due
                         * to max sg elements limit
                         */
-                       copy -= required_size - ctx->sg_plaintext_size;
+                       copy -= required_size - rec->sg_plaintext_size;
                        full_record = true;
                }
 
                get_page(page);
-               sg = ctx->sg_plaintext_data + ctx->sg_plaintext_num_elem;
+               sg = rec->sg_plaintext_data + rec->sg_plaintext_num_elem;
                sg_set_page(sg, page, copy, offset);
                sg_unmark_end(sg);
 
-               ctx->sg_plaintext_num_elem++;
+               rec->sg_plaintext_num_elem++;
 
                sk_mem_charge(sk, copy);
                offset += copy;
                size -= copy;
-               ctx->sg_plaintext_size += copy;
-               tls_ctx->pending_open_record_frags = ctx->sg_plaintext_num_elem;
+               rec->sg_plaintext_size += copy;
+               tls_ctx->pending_open_record_frags = rec->sg_plaintext_num_elem;
 
                if (full_record || eor ||
-                   ctx->sg_plaintext_num_elem ==
-                   ARRAY_SIZE(ctx->sg_plaintext_data)) {
-push_record:
+                   rec->sg_plaintext_num_elem ==
+                   ARRAY_SIZE(rec->sg_plaintext_data)) {
                        ret = tls_push_record(sk, flags, record_type);
                        if (ret) {
-                               if (ret == -ENOMEM)
-                                       goto wait_for_memory;
-
-                               goto sendpage_end;
+                               if (ret == -EINPROGRESS)
+                                       num_async++;
+                               else if (ret != -EAGAIN)
+                                       goto sendpage_end;
                        }
                }
                continue;
@@ -743,16 +993,20 @@ wait_for_sndbuf:
 wait_for_memory:
                ret = sk_stream_wait_memory(sk, &timeo);
                if (ret) {
-                       trim_both_sgl(sk, ctx->sg_plaintext_size);
+                       trim_both_sgl(sk, rec->sg_plaintext_size);
                        goto sendpage_end;
                }
 
-               if (tls_is_pending_closed_record(tls_ctx))
-                       goto push_record;
-
                goto alloc_payload;
        }
 
+       if (num_async) {
+               /* Transmit if any encryptions have completed */
+               if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
+                       cancel_delayed_work(&ctx->tx_work.work);
+                       tls_tx_records(sk, flags);
+               }
+       }
 sendpage_end:
        if (orig_size > size)
                ret = orig_size - size;
@@ -1300,6 +1554,49 @@ void tls_sw_free_resources_tx(struct sock *sk)
 {
        struct tls_context *tls_ctx = tls_get_ctx(sk);
        struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+       struct tls_rec *rec, *tmp;
+
+       /* Wait for any pending async encryptions to complete */
+       smp_store_mb(ctx->async_notify, true);
+       if (atomic_read(&ctx->encrypt_pending))
+               crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
+
+       cancel_delayed_work_sync(&ctx->tx_work.work);
+
+       /* Tx whatever records we can transmit and abandon the rest */
+       tls_tx_records(sk, -1);
+
+       /* Free up un-sent records in tx_ready_list. First, free
+        * the partially sent record if any at head of tx_list.
+        */
+       if (tls_ctx->partially_sent_record) {
+               struct scatterlist *sg = tls_ctx->partially_sent_record;
+
+               while (1) {
+                       put_page(sg_page(sg));
+                       sk_mem_uncharge(sk, sg->length);
+
+                       if (sg_is_last(sg))
+                               break;
+                       sg++;
+               }
+
+               tls_ctx->partially_sent_record = NULL;
+
+               rec = list_first_entry(&ctx->tx_ready_list,
+                                      struct tls_rec, list);
+               list_del(&rec->list);
+               kfree(rec);
+       }
+
+       list_for_each_entry_safe(rec, tmp, &ctx->tx_ready_list, list) {
+               free_sg(sk, rec->sg_encrypted_data,
+                       &rec->sg_encrypted_num_elem,
+                       &rec->sg_encrypted_size);
+
+               list_del(&rec->list);
+               kfree(rec);
+       }
 
        crypto_free_aead(ctx->aead_send);
        tls_free_both_sg(sk);
@@ -1336,6 +1633,24 @@ void tls_sw_free_resources_rx(struct sock *sk)
        kfree(ctx);
 }
 
+/* The work handler to transmitt the encrypted records in tx_ready_list */
+static void tx_work_handler(struct work_struct *work)
+{
+       struct delayed_work *delayed_work = to_delayed_work(work);
+       struct tx_work *tx_work = container_of(delayed_work,
+                                              struct tx_work, work);
+       struct sock *sk = tx_work->sk;
+       struct tls_context *tls_ctx = tls_get_ctx(sk);
+       struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
+
+       if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
+               return;
+
+       lock_sock(sk);
+       tls_tx_records(sk, -1);
+       release_sock(sk);
+}
+
 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
 {
        struct tls_crypto_info *crypto_info;
@@ -1385,6 +1700,9 @@ int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
                crypto_info = &ctx->crypto_send.info;
                cctx = &ctx->tx;
                aead = &sw_ctx_tx->aead_send;
+               INIT_LIST_HEAD(&sw_ctx_tx->tx_ready_list);
+               INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
+               sw_ctx_tx->tx_work.sk = sk;
        } else {
                crypto_init_wait(&sw_ctx_rx->async_wait);
                crypto_info = &ctx->crypto_recv.info;
@@ -1435,26 +1753,6 @@ int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
                goto free_iv;
        }
 
-       if (sw_ctx_tx) {
-               sg_init_table(sw_ctx_tx->sg_encrypted_data,
-                             ARRAY_SIZE(sw_ctx_tx->sg_encrypted_data));
-               sg_init_table(sw_ctx_tx->sg_plaintext_data,
-                             ARRAY_SIZE(sw_ctx_tx->sg_plaintext_data));
-
-               sg_init_table(sw_ctx_tx->sg_aead_in, 2);
-               sg_set_buf(&sw_ctx_tx->sg_aead_in[0], sw_ctx_tx->aad_space,
-                          sizeof(sw_ctx_tx->aad_space));
-               sg_unmark_end(&sw_ctx_tx->sg_aead_in[1]);
-               sg_chain(sw_ctx_tx->sg_aead_in, 2,
-                        sw_ctx_tx->sg_plaintext_data);
-               sg_init_table(sw_ctx_tx->sg_aead_out, 2);
-               sg_set_buf(&sw_ctx_tx->sg_aead_out[0], sw_ctx_tx->aad_space,
-                          sizeof(sw_ctx_tx->aad_space));
-               sg_unmark_end(&sw_ctx_tx->sg_aead_out[1]);
-               sg_chain(sw_ctx_tx->sg_aead_out, 2,
-                        sw_ctx_tx->sg_encrypted_data);
-       }
-
        if (!*aead) {
                *aead = crypto_alloc_aead("gcm(aes)", 0, 0);
                if (IS_ERR(*aead)) {
@@ -1491,6 +1789,8 @@ int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
                sw_ctx_rx->sk_poll = sk->sk_socket->ops->poll;
 
                strp_check_rcv(&sw_ctx_rx->strp);
+       } else {
+               ctx->tx_seq_number = be64_to_cpup((const __be64 *)rec_seq);
        }
 
        goto out;