drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_tx.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 // Copyright (c) 2019 Mellanox Technologies.
   3
   4 #include <linux/tls.h>
   5 #include "en.h"
   6 #include "en/txrx.h"
   7 #include "en_accel/ktls.h"
   8
   9 enum {
  10         MLX5E_STATIC_PARAMS_CONTEXT_TLS_1_2 = 0x2,
  11 };
  12
  13 enum {
  14         MLX5E_ENCRYPTION_STANDARD_TLS = 0x1,
  15 };
  16
  17 #define EXTRACT_INFO_FIELDS do { \
  18         salt    = info->salt;    \
  19         rec_seq = info->rec_seq; \
  20         salt_sz    = sizeof(info->salt);    \
  21         rec_seq_sz = sizeof(info->rec_seq); \
  22 } while (0)
  23
  24 static void
  25 fill_static_params_ctx(void *ctx, struct mlx5e_ktls_offload_context_tx *priv_tx)
  26 {
  27         struct tls_crypto_info *crypto_info = priv_tx->crypto_info;
  28         struct tls12_crypto_info_aes_gcm_128 *info;
  29         char *initial_rn, *gcm_iv;
  30         u16 salt_sz, rec_seq_sz;
  31         char *salt, *rec_seq;
  32         u8 tls_version;
  33
  34         if (WARN_ON(crypto_info->cipher_type != TLS_CIPHER_AES_GCM_128))
  35                 return;
  36
  37         info = (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
  38         EXTRACT_INFO_FIELDS;
  39
  40         gcm_iv      = MLX5_ADDR_OF(tls_static_params, ctx, gcm_iv);
  41         initial_rn  = MLX5_ADDR_OF(tls_static_params, ctx, initial_record_number);
  42
  43         memcpy(gcm_iv,      salt,    salt_sz);
  44         memcpy(initial_rn,  rec_seq, rec_seq_sz);
  45
  46         tls_version = MLX5E_STATIC_PARAMS_CONTEXT_TLS_1_2;
  47
  48         MLX5_SET(tls_static_params, ctx, tls_version, tls_version);
  49         MLX5_SET(tls_static_params, ctx, const_1, 1);
  50         MLX5_SET(tls_static_params, ctx, const_2, 2);
  51         MLX5_SET(tls_static_params, ctx, encryption_standard,
  52                  MLX5E_ENCRYPTION_STANDARD_TLS);
  53         MLX5_SET(tls_static_params, ctx, dek_index, priv_tx->key_id);
  54 }
  55
  56 static void
  57 build_static_params(struct mlx5e_umr_wqe *wqe, u16 pc, u32 sqn,
  58                     struct mlx5e_ktls_offload_context_tx *priv_tx,
  59                     bool fence)
  60 {
  61         struct mlx5_wqe_ctrl_seg     *cseg  = &wqe->ctrl;
  62         struct mlx5_wqe_umr_ctrl_seg *ucseg = &wqe->uctrl;
  63
  64 #define STATIC_PARAMS_DS_CNT \
  65         DIV_ROUND_UP(MLX5E_KTLS_STATIC_UMR_WQE_SZ, MLX5_SEND_WQE_DS)
  66
  67         cseg->opmod_idx_opcode = cpu_to_be32((pc << 8) | MLX5_OPCODE_UMR |
  68                                              (MLX5_OPC_MOD_TLS_TIS_STATIC_PARAMS << 24));
  69         cseg->qpn_ds           = cpu_to_be32((sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
  70                                              STATIC_PARAMS_DS_CNT);
  71         cseg->fm_ce_se         = fence ? MLX5_FENCE_MODE_INITIATOR_SMALL : 0;
  72         cseg->tisn             = cpu_to_be32(priv_tx->tisn << 8);
  73
  74         ucseg->flags = MLX5_UMR_INLINE;
  75         ucseg->bsf_octowords = cpu_to_be16(MLX5_ST_SZ_BYTES(tls_static_params) / 16);
  76
  77         fill_static_params_ctx(wqe->tls_static_params_ctx, priv_tx);
  78 }
  79
  80 static void
  81 fill_progress_params_ctx(void *ctx, struct mlx5e_ktls_offload_context_tx *priv_tx)
  82 {
  83         MLX5_SET(tls_progress_params, ctx, tisn, priv_tx->tisn);
  84         MLX5_SET(tls_progress_params, ctx, record_tracker_state,
  85                  MLX5E_TLS_PROGRESS_PARAMS_RECORD_TRACKER_STATE_START);
  86         MLX5_SET(tls_progress_params, ctx, auth_state,
  87                  MLX5E_TLS_PROGRESS_PARAMS_AUTH_STATE_NO_OFFLOAD);
  88 }
  89
  90 static void
  91 build_progress_params(struct mlx5e_tx_wqe *wqe, u16 pc, u32 sqn,
  92                       struct mlx5e_ktls_offload_context_tx *priv_tx,
  93                       bool fence)
  94 {
  95         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
  96
  97 #define PROGRESS_PARAMS_DS_CNT \
  98         DIV_ROUND_UP(MLX5E_KTLS_PROGRESS_WQE_SZ, MLX5_SEND_WQE_DS)
  99
 100         cseg->opmod_idx_opcode =
 101                 cpu_to_be32((pc << 8) | MLX5_OPCODE_SET_PSV |
 102                             (MLX5_OPC_MOD_TLS_TIS_PROGRESS_PARAMS << 24));
 103         cseg->qpn_ds           = cpu_to_be32((sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
 104                                              PROGRESS_PARAMS_DS_CNT);
 105         cseg->fm_ce_se         = fence ? MLX5_FENCE_MODE_INITIATOR_SMALL : 0;
 106
 107         fill_progress_params_ctx(wqe->tls_progress_params_ctx, priv_tx);
 108 }
 109
 110 static void tx_fill_wi(struct mlx5e_txqsq *sq,
 111                        u16 pi, u8 num_wqebbs,
 112                        skb_frag_t *resync_dump_frag,
 113                        u32 num_bytes)
 114 {
 115         struct mlx5e_tx_wqe_info *wi = &sq->db.wqe_info[pi];
 116
 117         wi->skb              = NULL;
 118         wi->num_wqebbs       = num_wqebbs;
 119         wi->resync_dump_frag = resync_dump_frag;
 120         wi->num_bytes        = num_bytes;
 121 }
 122
 123 void mlx5e_ktls_tx_offload_set_pending(struct mlx5e_ktls_offload_context_tx *priv_tx)
 124 {
 125         priv_tx->ctx_post_pending = true;
 126 }
 127
 128 static bool
 129 mlx5e_ktls_tx_offload_test_and_clear_pending(struct mlx5e_ktls_offload_context_tx *priv_tx)
 130 {
 131         bool ret = priv_tx->ctx_post_pending;
 132
 133         priv_tx->ctx_post_pending = false;
 134
 135         return ret;
 136 }
 137
 138 static void
 139 post_static_params(struct mlx5e_txqsq *sq,
 140                    struct mlx5e_ktls_offload_context_tx *priv_tx,
 141                    bool fence)
 142 {
 143         struct mlx5e_umr_wqe *umr_wqe;
 144         u16 pi;
 145
 146         umr_wqe = mlx5e_sq_fetch_wqe(sq, MLX5E_KTLS_STATIC_UMR_WQE_SZ, &pi);
 147         build_static_params(umr_wqe, sq->pc, sq->sqn, priv_tx, fence);
 148         tx_fill_wi(sq, pi, MLX5E_KTLS_STATIC_WQEBBS, NULL, 0);
 149         sq->pc += MLX5E_KTLS_STATIC_WQEBBS;
 150 }
 151
 152 static void
 153 post_progress_params(struct mlx5e_txqsq *sq,
 154                      struct mlx5e_ktls_offload_context_tx *priv_tx,
 155                      bool fence)
 156 {
 157         struct mlx5e_tx_wqe *wqe;
 158         u16 pi;
 159
 160         wqe = mlx5e_sq_fetch_wqe(sq, MLX5E_KTLS_PROGRESS_WQE_SZ, &pi);
 161         build_progress_params(wqe, sq->pc, sq->sqn, priv_tx, fence);
 162         tx_fill_wi(sq, pi, MLX5E_KTLS_PROGRESS_WQEBBS, NULL, 0);
 163         sq->pc += MLX5E_KTLS_PROGRESS_WQEBBS;
 164 }
 165
 166 static void
 167 mlx5e_ktls_tx_post_param_wqes(struct mlx5e_txqsq *sq,
 168                               struct mlx5e_ktls_offload_context_tx *priv_tx,
 169                               bool skip_static_post, bool fence_first_post)
 170 {
 171         bool progress_fence = skip_static_post || !fence_first_post;
 172
 173         if (!skip_static_post)
 174                 post_static_params(sq, priv_tx, fence_first_post);
 175
 176         post_progress_params(sq, priv_tx, progress_fence);
 177 }
 178
 179 struct tx_sync_info {
 180         u64 rcd_sn;
 181         s32 sync_len;
 182         int nr_frags;
 183         skb_frag_t *frags[MAX_SKB_FRAGS];
 184 };
 185
 186 static bool tx_sync_info_get(struct mlx5e_ktls_offload_context_tx *priv_tx,
 187                              u32 tcp_seq, struct tx_sync_info *info)
 188 {
 189         struct tls_offload_context_tx *tx_ctx = priv_tx->tx_ctx;
 190         struct tls_record_info *record;
 191         int remaining, i = 0;
 192         unsigned long flags;
 193         bool ret = true;
 194
 195         spin_lock_irqsave(&tx_ctx->lock, flags);
 196         record = tls_get_record(tx_ctx, tcp_seq, &info->rcd_sn);
 197
 198         if (unlikely(!record)) {
 199                 ret = false;
 200                 goto out;
 201         }
 202
 203         if (unlikely(tcp_seq < tls_record_start_seq(record))) {
 204                 if (!tls_record_is_start_marker(record))
 205                         ret = false;
 206                 goto out;
 207         }
 208
 209         info->sync_len = tcp_seq - tls_record_start_seq(record);
 210         remaining = info->sync_len;
 211         while (remaining > 0) {
 212                 skb_frag_t *frag = &record->frags[i];
 213
 214                 __skb_frag_ref(frag);
 215                 remaining -= skb_frag_size(frag);
 216                 info->frags[i++] = frag;
 217         }
 218         /* reduce the part which will be sent with the original SKB */
 219         if (remaining < 0)
 220                 skb_frag_size_add(info->frags[i - 1], remaining);
 221         info->nr_frags = i;
 222 out:
 223         spin_unlock_irqrestore(&tx_ctx->lock, flags);
 224         return ret;
 225 }
 226
 227 static void
 228 tx_post_resync_params(struct mlx5e_txqsq *sq,
 229                       struct mlx5e_ktls_offload_context_tx *priv_tx,
 230                       u64 rcd_sn)
 231 {
 232         struct tls_crypto_info *crypto_info = priv_tx->crypto_info;
 233         struct tls12_crypto_info_aes_gcm_128 *info;
 234         __be64 rn_be = cpu_to_be64(rcd_sn);
 235         bool skip_static_post;
 236         u16 rec_seq_sz;
 237         char *rec_seq;
 238
 239         if (WARN_ON(crypto_info->cipher_type != TLS_CIPHER_AES_GCM_128))
 240                 return;
 241
 242         info = (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
 243         rec_seq = info->rec_seq;
 244         rec_seq_sz = sizeof(info->rec_seq);
 245
 246         skip_static_post = !memcmp(rec_seq, &rn_be, rec_seq_sz);
 247         if (!skip_static_post)
 248                 memcpy(rec_seq, &rn_be, rec_seq_sz);
 249
 250         mlx5e_ktls_tx_post_param_wqes(sq, priv_tx, skip_static_post, true);
 251 }
 252
 253 struct mlx5e_dump_wqe {
 254         struct mlx5_wqe_ctrl_seg ctrl;
 255         struct mlx5_wqe_data_seg data;
 256 };
 257
 258 static int
 259 tx_post_resync_dump(struct mlx5e_txqsq *sq, struct sk_buff *skb,
 260                     skb_frag_t *frag, u32 tisn, bool first)
 261 {
 262         struct mlx5_wqe_ctrl_seg *cseg;
 263         struct mlx5_wqe_data_seg *dseg;
 264         struct mlx5e_dump_wqe *wqe;
 265         dma_addr_t dma_addr = 0;
 266         u8  num_wqebbs;
 267         u16 ds_cnt;
 268         int fsz;
 269         u16 pi;
 270
 271         wqe = mlx5e_sq_fetch_wqe(sq, sizeof(*wqe), &pi);
 272
 273         ds_cnt = sizeof(*wqe) / MLX5_SEND_WQE_DS;
 274         num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
 275
 276         cseg = &wqe->ctrl;
 277         dseg = &wqe->data;
 278
 279         cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8)  | MLX5_OPCODE_DUMP);
 280         cseg->qpn_ds           = cpu_to_be32((sq->sqn << 8) | ds_cnt);
 281         cseg->tisn             = cpu_to_be32(tisn << 8);
 282         cseg->fm_ce_se         = first ? MLX5_FENCE_MODE_INITIATOR_SMALL : 0;
 283
 284         fsz = skb_frag_size(frag);
 285         dma_addr = skb_frag_dma_map(sq->pdev, frag, 0, fsz,
 286                                     DMA_TO_DEVICE);
 287         if (unlikely(dma_mapping_error(sq->pdev, dma_addr)))
 288                 return -ENOMEM;
 289
 290         dseg->addr       = cpu_to_be64(dma_addr);
 291         dseg->lkey       = sq->mkey_be;
 292         dseg->byte_count = cpu_to_be32(fsz);
 293         mlx5e_dma_push(sq, dma_addr, fsz, MLX5E_DMA_MAP_PAGE);
 294
 295         tx_fill_wi(sq, pi, num_wqebbs, frag, fsz);
 296         sq->pc += num_wqebbs;
 297
 298         WARN(num_wqebbs > MLX5E_KTLS_MAX_DUMP_WQEBBS,
 299              "unexpected DUMP num_wqebbs, %d > %d",
 300              num_wqebbs, MLX5E_KTLS_MAX_DUMP_WQEBBS);
 301
 302         return 0;
 303 }
 304
 305 void mlx5e_ktls_tx_handle_resync_dump_comp(struct mlx5e_txqsq *sq,
 306                                            struct mlx5e_tx_wqe_info *wi,
 307                                            struct mlx5e_sq_dma *dma)
 308 {
 309         struct mlx5e_sq_stats *stats = sq->stats;
 310
 311         mlx5e_tx_dma_unmap(sq->pdev, dma);
 312         __skb_frag_unref(wi->resync_dump_frag);
 313         stats->tls_dump_packets++;
 314         stats->tls_dump_bytes += wi->num_bytes;
 315 }
 316
 317 static void tx_post_fence_nop(struct mlx5e_txqsq *sq)
 318 {
 319         struct mlx5_wq_cyc *wq = &sq->wq;
 320         u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
 321
 322         tx_fill_wi(sq, pi, 1, NULL, 0);
 323
 324         mlx5e_post_nop_fence(wq, sq->sqn, &sq->pc);
 325 }
 326
 327 static struct sk_buff *
 328 mlx5e_ktls_tx_handle_ooo(struct mlx5e_ktls_offload_context_tx *priv_tx,
 329                          struct mlx5e_txqsq *sq,
 330                          struct sk_buff *skb,
 331                          u32 seq)
 332 {
 333         struct mlx5e_sq_stats *stats = sq->stats;
 334         struct mlx5_wq_cyc *wq = &sq->wq;
 335         struct tx_sync_info info = {};
 336         u16 contig_wqebbs_room, pi;
 337         u8 num_wqebbs;
 338         int i;
 339
 340         if (!tx_sync_info_get(priv_tx, seq, &info)) {
 341                 /* We might get here if a retransmission reaches the driver
 342                  * after the relevant record is acked.
 343                  * It should be safe to drop the packet in this case
 344                  */
 345                 stats->tls_drop_no_sync_data++;
 346                 goto err_out;
 347         }
 348
 349         if (unlikely(info.sync_len < 0)) {
 350                 u32 payload;
 351                 int headln;
 352
 353                 headln = skb_transport_offset(skb) + tcp_hdrlen(skb);
 354                 payload = skb->len - headln;
 355                 if (likely(payload <= -info.sync_len))
 356                         return skb;
 357
 358                 stats->tls_drop_bypass_req++;
 359                 goto err_out;
 360         }
 361
 362         stats->tls_ooo++;
 363
 364         num_wqebbs = MLX5E_KTLS_STATIC_WQEBBS + MLX5E_KTLS_PROGRESS_WQEBBS +
 365                 (info.nr_frags ? info.nr_frags * MLX5E_KTLS_MAX_DUMP_WQEBBS : 1);
 366         pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
 367         contig_wqebbs_room = mlx5_wq_cyc_get_contig_wqebbs(wq, pi);
 368         if (unlikely(contig_wqebbs_room < num_wqebbs))
 369                 mlx5e_fill_sq_frag_edge(sq, wq, pi, contig_wqebbs_room);
 370
 371         tx_post_resync_params(sq, priv_tx, info.rcd_sn);
 372
 373         for (i = 0; i < info.nr_frags; i++)
 374                 if (tx_post_resync_dump(sq, skb, info.frags[i],
 375                                         priv_tx->tisn, !i))
 376                         goto err_out;
 377
 378         /* If no dump WQE was sent, we need to have a fence NOP WQE before the
 379          * actual data xmit.
 380          */
 381         if (!info.nr_frags)
 382                 tx_post_fence_nop(sq);
 383
 384         return skb;
 385
 386 err_out:
 387         dev_kfree_skb_any(skb);
 388         return NULL;
 389 }
 390
 391 struct sk_buff *mlx5e_ktls_handle_tx_skb(struct net_device *netdev,
 392                                          struct mlx5e_txqsq *sq,
 393                                          struct sk_buff *skb,
 394                                          struct mlx5e_tx_wqe **wqe, u16 *pi)
 395 {
 396         struct mlx5e_ktls_offload_context_tx *priv_tx;
 397         struct mlx5e_sq_stats *stats = sq->stats;
 398         struct mlx5_wqe_ctrl_seg *cseg;
 399         struct tls_context *tls_ctx;
 400         int datalen;
 401         u32 seq;
 402
 403         if (!skb->sk || !tls_is_sk_tx_device_offloaded(skb->sk))
 404                 goto out;
 405
 406         datalen = skb->len - (skb_transport_offset(skb) + tcp_hdrlen(skb));
 407         if (!datalen)
 408                 goto out;
 409
 410         tls_ctx = tls_get_ctx(skb->sk);
 411         if (WARN_ON_ONCE(tls_ctx->netdev != netdev))
 412                 goto err_out;
 413
 414         priv_tx = mlx5e_get_ktls_tx_priv_ctx(tls_ctx);
 415
 416         if (unlikely(mlx5e_ktls_tx_offload_test_and_clear_pending(priv_tx))) {
 417                 mlx5e_ktls_tx_post_param_wqes(sq, priv_tx, false, false);
 418                 *wqe = mlx5e_sq_fetch_wqe(sq, sizeof(**wqe), pi);
 419                 stats->tls_ctx++;
 420         }
 421
 422         seq = ntohl(tcp_hdr(skb)->seq);
 423         if (unlikely(priv_tx->expected_seq != seq)) {
 424                 skb = mlx5e_ktls_tx_handle_ooo(priv_tx, sq, skb, seq);
 425                 if (unlikely(!skb))
 426                         goto out;
 427                 *wqe = mlx5e_sq_fetch_wqe(sq, sizeof(**wqe), pi);
 428         }
 429
 430         priv_tx->expected_seq = seq + datalen;
 431
 432         cseg = &(*wqe)->ctrl;
 433         cseg->tisn = cpu_to_be32(priv_tx->tisn << 8);
 434
 435         stats->tls_encrypted_packets += skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
 436         stats->tls_encrypted_bytes   += datalen;
 437
 438 out:
 439         return skb;
 440
 441 err_out:
 442         dev_kfree_skb_any(skb);
 443         return NULL;
 444 }