drivers/regex/mlx5/mlx5_regex_fastpath.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2020 Mellanox Technologies, Ltd
   3  */
   4
   5 #include <unistd.h>
   6 #include <strings.h>
   7 #include <stdint.h>
   8 #include <sys/mman.h>
   9
  10 #include <rte_malloc.h>
  11 #include <rte_log.h>
  12 #include <rte_errno.h>
  13 #include <rte_bus_pci.h>
  14 #include <rte_pci.h>
  15 #include <rte_regexdev_driver.h>
  16 #include <rte_mbuf.h>
  17
  18 #include <infiniband/mlx5dv.h>
  19 #include <mlx5_glue.h>
  20 #include <mlx5_common.h>
  21 #include <mlx5_prm.h>
  22
  23 #include "mlx5_regex_utils.h"
  24 #include "mlx5_rxp.h"
  25 #include "mlx5_regex.h"
  26
  27 #define MLX5_REGEX_MAX_WQE_INDEX 0xffff
  28 #define MLX5_REGEX_METADATA_SIZE ((size_t)64)
  29 #define MLX5_REGEX_MAX_OUTPUT (((size_t)1) << 11)
  30 #define MLX5_REGEX_WQE_CTRL_OFFSET 12
  31 #define MLX5_REGEX_WQE_METADATA_OFFSET 16
  32 #define MLX5_REGEX_WQE_GATHER_OFFSET 32
  33 #define MLX5_REGEX_WQE_SCATTER_OFFSET 48
  34 #define MLX5_REGEX_METADATA_OFF 32
  35 #define MLX5_REGEX_UMR_WQE_SIZE 192
  36 /* The maximum KLMs can be added to one UMR indirect mkey. */
  37 #define MLX5_REGEX_MAX_KLM_NUM 128
  38 /* The KLM array size for one job. */
  39 #define MLX5_REGEX_KLMS_SIZE \
  40         ((MLX5_REGEX_MAX_KLM_NUM) * sizeof(struct mlx5_klm))
  41 /* In WQE set mode, the pi should be quarter of the MLX5_REGEX_MAX_WQE_INDEX. */
  42 #define MLX5_REGEX_UMR_QP_PI_IDX(pi, ops) \
  43         (((pi) + (ops)) & (MLX5_REGEX_MAX_WQE_INDEX >> 2))
  44
  45 static inline uint32_t
  46 qp_size_get(struct mlx5_regex_hw_qp *qp)
  47 {
  48         return (1U << qp->log_nb_desc);
  49 }
  50
  51 static inline uint32_t
  52 cq_size_get(struct mlx5_regex_cq *cq)
  53 {
  54         return (1U << cq->log_nb_desc);
  55 }
  56
  57 struct mlx5_regex_job {
  58         uint64_t user_id;
  59         volatile uint8_t *output;
  60         volatile uint8_t *metadata;
  61         struct mlx5_klm *imkey_array; /* Indirect mkey's KLM array. */
  62         struct mlx5_devx_obj *imkey; /* UMR WQE's indirect meky. */
  63 } __rte_cached_aligned;
  64
  65 static inline void
  66 set_data_seg(struct mlx5_wqe_data_seg *seg,
  67              uint32_t length, uint32_t lkey,
  68              uintptr_t address)
  69 {
  70         seg->byte_count = rte_cpu_to_be_32(length);
  71         seg->lkey = rte_cpu_to_be_32(lkey);
  72         seg->addr = rte_cpu_to_be_64(address);
  73 }
  74
  75 static inline void
  76 set_metadata_seg(struct mlx5_wqe_metadata_seg *seg,
  77                  uint32_t mmo_control_31_0, uint32_t lkey,
  78                  uintptr_t address)
  79 {
  80         seg->mmo_control_31_0 = htobe32(mmo_control_31_0);
  81         seg->lkey = rte_cpu_to_be_32(lkey);
  82         seg->addr = rte_cpu_to_be_64(address);
  83 }
  84
  85 static inline void
  86 set_regex_ctrl_seg(void *seg, uint8_t le, uint16_t subset_id0,
  87                    uint16_t subset_id1, uint16_t subset_id2,
  88                    uint16_t subset_id3, uint8_t ctrl)
  89 {
  90         MLX5_SET(regexp_mmo_control, seg, le, le);
  91         MLX5_SET(regexp_mmo_control, seg, ctrl, ctrl);
  92         MLX5_SET(regexp_mmo_control, seg, subset_id_0, subset_id0);
  93         MLX5_SET(regexp_mmo_control, seg, subset_id_1, subset_id1);
  94         MLX5_SET(regexp_mmo_control, seg, subset_id_2, subset_id2);
  95         MLX5_SET(regexp_mmo_control, seg, subset_id_3, subset_id3);
  96 }
  97
  98 static inline void
  99 set_wqe_ctrl_seg(struct mlx5_wqe_ctrl_seg *seg, uint16_t pi, uint8_t opcode,
 100                  uint8_t opmod, uint32_t qp_num, uint8_t fm_ce_se, uint8_t ds,
 101                  uint8_t signature, uint32_t imm)
 102 {
 103         seg->opmod_idx_opcode = rte_cpu_to_be_32(((uint32_t)opmod << 24) |
 104                                                  ((uint32_t)pi << 8) |
 105                                                  opcode);
 106         seg->qpn_ds = rte_cpu_to_be_32((qp_num << 8) | ds);
 107         seg->fm_ce_se = fm_ce_se;
 108         seg->signature = signature;
 109         seg->imm = imm;
 110 }
 111
 112 /**
 113  * Query LKey from a packet buffer for QP. If not found, add the mempool.
 114  *
 115  * @param priv
 116  *   Pointer to the priv object.
 117  * @param mr_ctrl
 118  *   Pointer to per-queue MR control structure.
 119  * @param mbuf
 120  *   Pointer to source mbuf, to search in.
 121  *
 122  * @return
 123  *   Searched LKey on success, UINT32_MAX on no match.
 124  */
 125 static inline uint32_t
 126 mlx5_regex_addr2mr(struct mlx5_regex_priv *priv, struct mlx5_mr_ctrl *mr_ctrl,
 127                    struct rte_mbuf *mbuf)
 128 {
 129         uintptr_t addr = rte_pktmbuf_mtod(mbuf, uintptr_t);
 130         uint32_t lkey;
 131
 132         /* Check generation bit to see if there's any change on existing MRs. */
 133         if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
 134                 mlx5_mr_flush_local_cache(mr_ctrl);
 135         /* Linear search on MR cache array. */
 136         lkey = mlx5_mr_lookup_lkey(mr_ctrl->cache, &mr_ctrl->mru,
 137                                    MLX5_MR_CACHE_N, addr);
 138         if (likely(lkey != UINT32_MAX))
 139                 return lkey;
 140         /* Take slower bottom-half on miss. */
 141         return mlx5_mr_addr2mr_bh(priv->pd, 0, &priv->mr_scache, mr_ctrl, addr,
 142                                   !!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
 143 }
 144
 145
 146 static inline void
 147 __prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_hw_qp *qp_obj,
 148            struct rte_regex_ops *op, struct mlx5_regex_job *job,
 149            size_t pi, struct mlx5_klm *klm)
 150 {
 151         size_t wqe_offset = (pi & (qp_size_get(qp_obj) - 1)) *
 152                             (MLX5_SEND_WQE_BB << (priv->has_umr ? 2 : 0)) +
 153                             (priv->has_umr ? MLX5_REGEX_UMR_WQE_SIZE : 0);
 154         uint16_t group0 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID0_VALID_F ?
 155                                 op->group_id0 : 0;
 156         uint16_t group1 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID1_VALID_F ?
 157                                 op->group_id1 : 0;
 158         uint16_t group2 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID2_VALID_F ?
 159                                 op->group_id2 : 0;
 160         uint16_t group3 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID3_VALID_F ?
 161                                 op->group_id3 : 0;
 162         uint8_t control = op->req_flags &
 163                                 RTE_REGEX_OPS_REQ_MATCH_HIGH_PRIORITY_F ? 1 : 0;
 164
 165         /* For backward compatibility. */
 166         if (!(op->req_flags & (RTE_REGEX_OPS_REQ_GROUP_ID0_VALID_F |
 167                                RTE_REGEX_OPS_REQ_GROUP_ID1_VALID_F |
 168                                RTE_REGEX_OPS_REQ_GROUP_ID2_VALID_F |
 169                                RTE_REGEX_OPS_REQ_GROUP_ID3_VALID_F)))
 170                 group0 = op->group_id0;
 171         uint8_t *wqe = (uint8_t *)(uintptr_t)qp_obj->qp_obj.wqes + wqe_offset;
 172         int ds = 4; /*  ctrl + meta + input + output */
 173
 174         set_wqe_ctrl_seg((struct mlx5_wqe_ctrl_seg *)wqe,
 175                          (priv->has_umr ? (pi * 4 + 3) : pi),
 176                          MLX5_OPCODE_MMO, MLX5_OPC_MOD_MMO_REGEX,
 177                          qp_obj->qp_obj.qp->id, 0, ds, 0, 0);
 178         set_regex_ctrl_seg(wqe + 12, 0, group0, group1, group2, group3,
 179                            control);
 180         struct mlx5_wqe_data_seg *input_seg =
 181                 (struct mlx5_wqe_data_seg *)(wqe +
 182                                              MLX5_REGEX_WQE_GATHER_OFFSET);
 183         input_seg->byte_count = rte_cpu_to_be_32(klm->byte_count);
 184         input_seg->addr = rte_cpu_to_be_64(klm->address);
 185         input_seg->lkey = klm->mkey;
 186         job->user_id = op->user_id;
 187 }
 188
 189 static inline void
 190 prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
 191          struct mlx5_regex_hw_qp *qp_obj, struct rte_regex_ops *op,
 192          struct mlx5_regex_job *job)
 193 {
 194         struct mlx5_klm klm;
 195
 196         klm.byte_count = rte_pktmbuf_data_len(op->mbuf);
 197         klm.mkey = mlx5_regex_addr2mr(priv, &qp->mr_ctrl, op->mbuf);
 198         klm.address = rte_pktmbuf_mtod(op->mbuf, uintptr_t);
 199         __prep_one(priv, qp_obj, op, job, qp_obj->pi, &klm);
 200         qp_obj->db_pi = qp_obj->pi;
 201         qp_obj->pi = (qp_obj->pi + 1) & MLX5_REGEX_MAX_WQE_INDEX;
 202 }
 203
 204 static inline void
 205 send_doorbell(struct mlx5_regex_priv *priv, struct mlx5_regex_hw_qp *qp_obj)
 206 {
 207         struct mlx5dv_devx_uar *uar = priv->uar;
 208         size_t wqe_offset = (qp_obj->db_pi & (qp_size_get(qp_obj) - 1)) *
 209                 (MLX5_SEND_WQE_BB << (priv->has_umr ? 2 : 0)) +
 210                 (priv->has_umr ? MLX5_REGEX_UMR_WQE_SIZE : 0);
 211         uint8_t *wqe = (uint8_t *)(uintptr_t)qp_obj->qp_obj.wqes + wqe_offset;
 212         /* Or the fm_ce_se instead of set, avoid the fence be cleared. */
 213         ((struct mlx5_wqe_ctrl_seg *)wqe)->fm_ce_se |= MLX5_WQE_CTRL_CQ_UPDATE;
 214         uint64_t *doorbell_addr =
 215                 (uint64_t *)((uint8_t *)uar->base_addr + 0x800);
 216         rte_io_wmb();
 217         qp_obj->qp_obj.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32((priv->has_umr ?
 218                                         (qp_obj->db_pi * 4 + 3) : qp_obj->db_pi)
 219                                         & MLX5_REGEX_MAX_WQE_INDEX);
 220         rte_wmb();
 221         *doorbell_addr = *(volatile uint64_t *)wqe;
 222         rte_wmb();
 223 }
 224
 225 static inline int
 226 get_free(struct mlx5_regex_hw_qp *qp, uint8_t has_umr) {
 227         return (qp_size_get(qp) - ((qp->pi - qp->ci) &
 228                         (has_umr ? (MLX5_REGEX_MAX_WQE_INDEX >> 2) :
 229                         MLX5_REGEX_MAX_WQE_INDEX)));
 230 }
 231
 232 static inline uint32_t
 233 job_id_get(uint32_t qid, size_t qp_size, size_t index) {
 234         return qid * qp_size + (index & (qp_size - 1));
 235 }
 236
 237 #ifdef HAVE_MLX5_UMR_IMKEY
 238 static inline int
 239 mkey_klm_available(struct mlx5_klm *klm, uint32_t pos, uint32_t new)
 240 {
 241         return (klm && ((pos + new) <= MLX5_REGEX_MAX_KLM_NUM));
 242 }
 243
 244 static inline void
 245 complete_umr_wqe(struct mlx5_regex_qp *qp, struct mlx5_regex_hw_qp *qp_obj,
 246                  struct mlx5_regex_job *mkey_job,
 247                  size_t umr_index, uint32_t klm_size, uint32_t total_len)
 248 {
 249         size_t wqe_offset = (umr_index & (qp_size_get(qp_obj) - 1)) *
 250                 (MLX5_SEND_WQE_BB * 4);
 251         struct mlx5_wqe_ctrl_seg *wqe = (struct mlx5_wqe_ctrl_seg *)((uint8_t *)
 252                                    (uintptr_t)qp_obj->qp_obj.wqes + wqe_offset);
 253         struct mlx5_wqe_umr_ctrl_seg *ucseg =
 254                                 (struct mlx5_wqe_umr_ctrl_seg *)(wqe + 1);
 255         struct mlx5_wqe_mkey_context_seg *mkc =
 256                                 (struct mlx5_wqe_mkey_context_seg *)(ucseg + 1);
 257         struct mlx5_klm *iklm = (struct mlx5_klm *)(mkc + 1);
 258         uint16_t klm_align = RTE_ALIGN(klm_size, 4);
 259
 260         memset(wqe, 0, MLX5_REGEX_UMR_WQE_SIZE);
 261         /* Set WQE control seg. Non-inline KLM UMR WQE size must be 9 WQE_DS. */
 262         set_wqe_ctrl_seg(wqe, (umr_index * 4), MLX5_OPCODE_UMR,
 263                          0, qp_obj->qp_obj.qp->id, 0, 9, 0,
 264                          rte_cpu_to_be_32(mkey_job->imkey->id));
 265         /* Set UMR WQE control seg. */
 266         ucseg->mkey_mask |= rte_cpu_to_be_64(MLX5_WQE_UMR_CTRL_MKEY_MASK_LEN |
 267                                 MLX5_WQE_UMR_CTRL_FLAG_TRNSLATION_OFFSET |
 268                                 MLX5_WQE_UMR_CTRL_MKEY_MASK_ACCESS_LOCAL_WRITE);
 269         ucseg->klm_octowords = rte_cpu_to_be_16(klm_align);
 270         /* Set mkey context seg. */
 271         mkc->len = rte_cpu_to_be_64(total_len);
 272         mkc->qpn_mkey = rte_cpu_to_be_32(0xffffff00 |
 273                                         (mkey_job->imkey->id & 0xff));
 274         /* Set UMR pointer to data seg. */
 275         iklm->address = rte_cpu_to_be_64
 276                                 ((uintptr_t)((char *)mkey_job->imkey_array));
 277         iklm->mkey = rte_cpu_to_be_32(qp->imkey_addr->lkey);
 278         iklm->byte_count = rte_cpu_to_be_32(klm_align);
 279         /* Clear the padding memory. */
 280         memset((uint8_t *)&mkey_job->imkey_array[klm_size], 0,
 281                sizeof(struct mlx5_klm) * (klm_align - klm_size));
 282
 283         /* Add the following RegEx WQE with fence. */
 284         wqe = (struct mlx5_wqe_ctrl_seg *)
 285                                 (((uint8_t *)wqe) + MLX5_REGEX_UMR_WQE_SIZE);
 286         wqe->fm_ce_se |= MLX5_WQE_CTRL_INITIATOR_SMALL_FENCE;
 287 }
 288
 289 static inline void
 290 prep_nop_regex_wqe_set(struct mlx5_regex_priv *priv,
 291                 struct mlx5_regex_hw_qp *qp, struct rte_regex_ops *op,
 292                 struct mlx5_regex_job *job, size_t pi, struct mlx5_klm *klm)
 293 {
 294         size_t wqe_offset = (pi & (qp_size_get(qp) - 1)) *
 295                             (MLX5_SEND_WQE_BB << 2);
 296         struct mlx5_wqe_ctrl_seg *wqe = (struct mlx5_wqe_ctrl_seg *)((uint8_t *)
 297                                    (uintptr_t)qp->qp_obj.wqes + wqe_offset);
 298
 299         /* Clear the WQE memory used as UMR WQE previously. */
 300         if ((rte_be_to_cpu_32(wqe->opmod_idx_opcode) & 0xff) != MLX5_OPCODE_NOP)
 301                 memset(wqe, 0, MLX5_REGEX_UMR_WQE_SIZE);
 302         /* UMR WQE size is 9 DS, align nop WQE to 3 WQEBBS(12 DS). */
 303         set_wqe_ctrl_seg(wqe, pi * 4, MLX5_OPCODE_NOP, 0, qp->qp_obj.qp->id,
 304                          0, 12, 0, 0);
 305         __prep_one(priv, qp, op, job, pi, klm);
 306 }
 307
 308 static inline void
 309 prep_regex_umr_wqe_set(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
 310                 struct mlx5_regex_hw_qp *qp_obj, struct rte_regex_ops **op,
 311                 size_t nb_ops)
 312 {
 313         struct mlx5_regex_job *job = NULL;
 314         size_t hw_qpid = qp_obj->qpn, mkey_job_id = 0;
 315         size_t left_ops = nb_ops;
 316         uint32_t klm_num = 0;
 317         uint32_t len = 0;
 318         struct mlx5_klm *mkey_klm = NULL;
 319         struct mlx5_klm klm;
 320
 321         while (left_ops--)
 322                 rte_prefetch0(op[left_ops]);
 323         left_ops = nb_ops;
 324         /*
 325          * Build the WQE set by reverse. In case the burst may consume
 326          * multiple mkeys, build the WQE set as normal will hard to
 327          * address the last mkey index, since we will only know the last
 328          * RegEx WQE's index when finishes building.
 329          */
 330         while (left_ops--) {
 331                 struct rte_mbuf *mbuf = op[left_ops]->mbuf;
 332                 size_t pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, left_ops);
 333
 334                 if (mbuf->nb_segs > 1) {
 335                         size_t scatter_size = 0;
 336
 337                         if (!mkey_klm_available(mkey_klm, klm_num,
 338                                                 mbuf->nb_segs)) {
 339                                 /*
 340                                  * The mkey's KLM is full, create the UMR
 341                                  * WQE in the next WQE set.
 342                                  */
 343                                 if (mkey_klm)
 344                                         complete_umr_wqe(qp, qp_obj,
 345                                                 &qp->jobs[mkey_job_id],
 346                                                 MLX5_REGEX_UMR_QP_PI_IDX(pi, 1),
 347                                                 klm_num, len);
 348                                 /*
 349                                  * Get the indircet mkey and KLM array index
 350                                  * from the last WQE set.
 351                                  */
 352                                 mkey_job_id = job_id_get(hw_qpid,
 353                                                 qp_size_get(qp_obj), pi);
 354                                 mkey_klm = qp->jobs[mkey_job_id].imkey_array;
 355                                 klm_num = 0;
 356                                 len = 0;
 357                         }
 358                         /* Build RegEx WQE's data segment KLM. */
 359                         klm.address = len;
 360                         klm.mkey = rte_cpu_to_be_32
 361                                         (qp->jobs[mkey_job_id].imkey->id);
 362                         while (mbuf) {
 363                                 /* Build indirect mkey seg's KLM. */
 364                                 mkey_klm->mkey = mlx5_regex_addr2mr
 365                                                 (priv, &qp->mr_ctrl, mbuf);
 366                                 mkey_klm->address = rte_cpu_to_be_64
 367                                         (rte_pktmbuf_mtod(mbuf, uintptr_t));
 368                                 mkey_klm->byte_count = rte_cpu_to_be_32
 369                                                 (rte_pktmbuf_data_len(mbuf));
 370                                 /*
 371                                  * Save the mbuf's total size for RegEx data
 372                                  * segment.
 373                                  */
 374                                 scatter_size += rte_pktmbuf_data_len(mbuf);
 375                                 mkey_klm++;
 376                                 klm_num++;
 377                                 mbuf = mbuf->next;
 378                         }
 379                         len += scatter_size;
 380                         klm.byte_count = scatter_size;
 381                 } else {
 382                         /* The single mubf case. Build the KLM directly. */
 383                         klm.mkey = mlx5_regex_addr2mr(priv, &qp->mr_ctrl, mbuf);
 384                         klm.address = rte_pktmbuf_mtod(mbuf, uintptr_t);
 385                         klm.byte_count = rte_pktmbuf_data_len(mbuf);
 386                 }
 387                 job = &qp->jobs[job_id_get(hw_qpid, qp_size_get(qp_obj), pi)];
 388                 /*
 389                  * Build the nop + RegEx WQE set by default. The fist nop WQE
 390                  * will be updated later as UMR WQE if scattered mubf exist.
 391                  */
 392                 prep_nop_regex_wqe_set(priv, qp_obj, op[left_ops], job, pi,
 393                                         &klm);
 394         }
 395         /*
 396          * Scattered mbuf have been added to the KLM array. Complete the build
 397          * of UMR WQE, update the first nop WQE as UMR WQE.
 398          */
 399         if (mkey_klm)
 400                 complete_umr_wqe(qp, qp_obj, &qp->jobs[mkey_job_id], qp_obj->pi,
 401                                  klm_num, len);
 402         qp_obj->db_pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, nb_ops - 1);
 403         qp_obj->pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, nb_ops);
 404 }
 405
 406 uint16_t
 407 mlx5_regexdev_enqueue_gga(struct rte_regexdev *dev, uint16_t qp_id,
 408                           struct rte_regex_ops **ops, uint16_t nb_ops)
 409 {
 410         struct mlx5_regex_priv *priv = dev->data->dev_private;
 411         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
 412         struct mlx5_regex_hw_qp *qp_obj;
 413         size_t hw_qpid, nb_left = nb_ops, nb_desc;
 414
 415         while ((hw_qpid = ffs(queue->free_qps))) {
 416                 hw_qpid--; /* ffs returns 1 for bit 0 */
 417                 qp_obj = &queue->qps[hw_qpid];
 418                 nb_desc = get_free(qp_obj, priv->has_umr);
 419                 if (nb_desc) {
 420                         /* The ops be handled can't exceed nb_ops. */
 421                         if (nb_desc > nb_left)
 422                                 nb_desc = nb_left;
 423                         else
 424                                 queue->free_qps &= ~(1 << hw_qpid);
 425                         prep_regex_umr_wqe_set(priv, queue, qp_obj, ops,
 426                                 nb_desc);
 427                         send_doorbell(priv, qp_obj);
 428                         nb_left -= nb_desc;
 429                 }
 430                 if (!nb_left)
 431                         break;
 432                 ops += nb_desc;
 433         }
 434         nb_ops -= nb_left;
 435         queue->pi += nb_ops;
 436         return nb_ops;
 437 }
 438 #endif
 439
 440 uint16_t
 441 mlx5_regexdev_enqueue(struct rte_regexdev *dev, uint16_t qp_id,
 442                       struct rte_regex_ops **ops, uint16_t nb_ops)
 443 {
 444         struct mlx5_regex_priv *priv = dev->data->dev_private;
 445         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
 446         struct mlx5_regex_hw_qp *qp_obj;
 447         size_t hw_qpid, job_id, i = 0;
 448
 449         while ((hw_qpid = ffs(queue->free_qps))) {
 450                 hw_qpid--; /* ffs returns 1 for bit 0 */
 451                 qp_obj = &queue->qps[hw_qpid];
 452                 while (get_free(qp_obj, priv->has_umr)) {
 453                         job_id = job_id_get(hw_qpid, qp_size_get(qp_obj),
 454                                 qp_obj->pi);
 455                         prep_one(priv, queue, qp_obj, ops[i],
 456                                 &queue->jobs[job_id]);
 457                         i++;
 458                         if (unlikely(i == nb_ops)) {
 459                                 send_doorbell(priv, qp_obj);
 460                                 goto out;
 461                         }
 462                 }
 463                 queue->free_qps &= ~(1 << hw_qpid);
 464                 send_doorbell(priv, qp_obj);
 465         }
 466
 467 out:
 468         queue->pi += i;
 469         return i;
 470 }
 471
 472 #define MLX5_REGEX_RESP_SZ 8
 473
 474 static inline void
 475 extract_result(struct rte_regex_ops *op, struct mlx5_regex_job *job)
 476 {
 477         size_t j;
 478         size_t offset;
 479         uint16_t status;
 480
 481         op->user_id = job->user_id;
 482         op->nb_matches = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
 483                                            MLX5_REGEX_METADATA_OFF,
 484                                            match_count);
 485         op->nb_actual_matches = MLX5_GET_VOLATILE(regexp_metadata,
 486                                                   job->metadata +
 487                                                   MLX5_REGEX_METADATA_OFF,
 488                                                   detected_match_count);
 489         for (j = 0; j < op->nb_matches; j++) {
 490                 offset = MLX5_REGEX_RESP_SZ * j;
 491                 op->matches[j].rule_id =
 492                         MLX5_GET_VOLATILE(regexp_match_tuple,
 493                                           (job->output + offset), rule_id);
 494                 op->matches[j].start_offset =
 495                         MLX5_GET_VOLATILE(regexp_match_tuple,
 496                                           (job->output +  offset), start_ptr);
 497                 op->matches[j].len =
 498                         MLX5_GET_VOLATILE(regexp_match_tuple,
 499                                           (job->output +  offset), length);
 500         }
 501         status = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
 502                                    MLX5_REGEX_METADATA_OFF,
 503                                    status);
 504         op->rsp_flags = 0;
 505         if (status & MLX5_RXP_RESP_STATUS_PMI_SOJ)
 506                 op->rsp_flags |= RTE_REGEX_OPS_RSP_PMI_SOJ_F;
 507         if (status & MLX5_RXP_RESP_STATUS_PMI_EOJ)
 508                 op->rsp_flags |= RTE_REGEX_OPS_RSP_PMI_EOJ_F;
 509         if (status & MLX5_RXP_RESP_STATUS_MAX_LATENCY)
 510                 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_SCAN_TIMEOUT_F;
 511         if (status & MLX5_RXP_RESP_STATUS_MAX_MATCH)
 512                 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_MATCH_F;
 513         if (status & MLX5_RXP_RESP_STATUS_MAX_PREFIX)
 514                 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_PREFIX_F;
 515         if (status & MLX5_RXP_RESP_STATUS_MAX_PRI_THREADS)
 516                 op->rsp_flags |= RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
 517         if (status & MLX5_RXP_RESP_STATUS_MAX_SEC_THREADS)
 518                 op->rsp_flags |= RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
 519 }
 520
 521 static inline volatile struct mlx5_cqe *
 522 poll_one(struct mlx5_regex_cq *cq)
 523 {
 524         volatile struct mlx5_cqe *cqe;
 525         size_t next_cqe_offset;
 526
 527         next_cqe_offset =  (cq->ci & (cq_size_get(cq) - 1));
 528         cqe = (volatile struct mlx5_cqe *)(cq->cq_obj.cqes + next_cqe_offset);
 529         rte_io_wmb();
 530
 531         int ret = check_cqe(cqe, cq_size_get(cq), cq->ci);
 532
 533         if (unlikely(ret == MLX5_CQE_STATUS_ERR)) {
 534                 DRV_LOG(ERR, "Completion with error on qp 0x%x",  0);
 535                 return NULL;
 536         }
 537
 538         if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN))
 539                 return NULL;
 540
 541         return cqe;
 542 }
 543
 544
 545 /**
 546  * DPDK callback for dequeue.
 547  *
 548  * @param dev
 549  *   Pointer to the regex dev structure.
 550  * @param qp_id
 551  *   The queue to enqueue the traffic to.
 552  * @param ops
 553  *   List of regex ops to dequeue.
 554  * @param nb_ops
 555  *   Number of ops in ops parameter.
 556  *
 557  * @return
 558  *   Number of packets successfully dequeued (<= pkts_n).
 559  */
 560 uint16_t
 561 mlx5_regexdev_dequeue(struct rte_regexdev *dev, uint16_t qp_id,
 562                       struct rte_regex_ops **ops, uint16_t nb_ops)
 563 {
 564         struct mlx5_regex_priv *priv = dev->data->dev_private;
 565         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
 566         struct mlx5_regex_cq *cq = &queue->cq;
 567         volatile struct mlx5_cqe *cqe;
 568         size_t i = 0;
 569
 570         while ((cqe = poll_one(cq))) {
 571                 uint16_t wq_counter
 572                         = (rte_be_to_cpu_16(cqe->wqe_counter) + 1) &
 573                           MLX5_REGEX_MAX_WQE_INDEX;
 574                 size_t hw_qpid = cqe->rsvd3[2];
 575                 struct mlx5_regex_hw_qp *qp_obj = &queue->qps[hw_qpid];
 576
 577                 /* UMR mode WQE counter move as WQE set(4 WQEBBS).*/
 578                 if (priv->has_umr)
 579                         wq_counter >>= 2;
 580                 while (qp_obj->ci != wq_counter) {
 581                         if (unlikely(i == nb_ops)) {
 582                                 /* Return without updating cq->ci */
 583                                 goto out;
 584                         }
 585                         uint32_t job_id = job_id_get(hw_qpid,
 586                                         qp_size_get(qp_obj), qp_obj->ci);
 587                         extract_result(ops[i], &queue->jobs[job_id]);
 588                         qp_obj->ci = (qp_obj->ci + 1) & (priv->has_umr ?
 589                                  (MLX5_REGEX_MAX_WQE_INDEX >> 2) :
 590                                   MLX5_REGEX_MAX_WQE_INDEX);
 591                         i++;
 592                 }
 593                 cq->ci = (cq->ci + 1) & 0xffffff;
 594                 rte_wmb();
 595                 cq->cq_obj.db_rec[0] = rte_cpu_to_be_32(cq->ci);
 596                 queue->free_qps |= (1 << hw_qpid);
 597         }
 598
 599 out:
 600         queue->ci += i;
 601         return i;
 602 }
 603
 604 static void
 605 setup_qps(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *queue)
 606 {
 607         size_t hw_qpid, entry;
 608         uint32_t job_id;
 609         for (hw_qpid = 0; hw_qpid < queue->nb_obj; hw_qpid++) {
 610                 struct mlx5_regex_hw_qp *qp_obj = &queue->qps[hw_qpid];
 611                 uint8_t *wqe = (uint8_t *)(uintptr_t)qp_obj->qp_obj.wqes;
 612                 for (entry = 0 ; entry < qp_size_get(qp_obj); entry++) {
 613                         job_id = hw_qpid * qp_size_get(qp_obj) + entry;
 614                         struct mlx5_regex_job *job = &queue->jobs[job_id];
 615
 616                         /* Fill UMR WQE with NOP in advanced. */
 617                         if (priv->has_umr) {
 618                                 set_wqe_ctrl_seg
 619                                         ((struct mlx5_wqe_ctrl_seg *)wqe,
 620                                          entry * 2, MLX5_OPCODE_NOP, 0,
 621                                          qp_obj->qp_obj.qp->id, 0, 12, 0, 0);
 622                                 wqe += MLX5_REGEX_UMR_WQE_SIZE;
 623                         }
 624                         set_metadata_seg((struct mlx5_wqe_metadata_seg *)
 625                                          (wqe + MLX5_REGEX_WQE_METADATA_OFFSET),
 626                                          0, queue->metadata->lkey,
 627                                          (uintptr_t)job->metadata);
 628                         set_data_seg((struct mlx5_wqe_data_seg *)
 629                                      (wqe + MLX5_REGEX_WQE_SCATTER_OFFSET),
 630                                      MLX5_REGEX_MAX_OUTPUT,
 631                                      queue->outputs->lkey,
 632                                      (uintptr_t)job->output);
 633                         wqe += 64;
 634                 }
 635                 queue->free_qps |= 1 << hw_qpid;
 636         }
 637 }
 638
 639 static int
 640 setup_buffers(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp)
 641 {
 642         struct ibv_pd *pd = priv->pd;
 643         uint32_t i;
 644         int err;
 645
 646         void *ptr = rte_calloc(__func__, qp->nb_desc,
 647                                MLX5_REGEX_METADATA_SIZE,
 648                                MLX5_REGEX_METADATA_SIZE);
 649         if (!ptr)
 650                 return -ENOMEM;
 651
 652         qp->metadata = mlx5_glue->reg_mr(pd, ptr,
 653                                          MLX5_REGEX_METADATA_SIZE * qp->nb_desc,
 654                                          IBV_ACCESS_LOCAL_WRITE);
 655         if (!qp->metadata) {
 656                 DRV_LOG(ERR, "Failed to register metadata");
 657                 rte_free(ptr);
 658                 return -EINVAL;
 659         }
 660
 661         ptr = rte_calloc(__func__, qp->nb_desc,
 662                          MLX5_REGEX_MAX_OUTPUT,
 663                          MLX5_REGEX_MAX_OUTPUT);
 664         if (!ptr) {
 665                 err = -ENOMEM;
 666                 goto err_output;
 667         }
 668         qp->outputs = mlx5_glue->reg_mr(pd, ptr,
 669                                         MLX5_REGEX_MAX_OUTPUT * qp->nb_desc,
 670                                         IBV_ACCESS_LOCAL_WRITE);
 671         if (!qp->outputs) {
 672                 rte_free(ptr);
 673                 DRV_LOG(ERR, "Failed to register output");
 674                 err = -EINVAL;
 675                 goto err_output;
 676         }
 677
 678         if (priv->has_umr) {
 679                 ptr = rte_calloc(__func__, qp->nb_desc, MLX5_REGEX_KLMS_SIZE,
 680                                  MLX5_REGEX_KLMS_SIZE);
 681                 if (!ptr) {
 682                         err = -ENOMEM;
 683                         goto err_imkey;
 684                 }
 685                 qp->imkey_addr = mlx5_glue->reg_mr(pd, ptr,
 686                                         MLX5_REGEX_KLMS_SIZE * qp->nb_desc,
 687                                         IBV_ACCESS_LOCAL_WRITE);
 688                 if (!qp->imkey_addr) {
 689                         rte_free(ptr);
 690                         DRV_LOG(ERR, "Failed to register output");
 691                         err = -EINVAL;
 692                         goto err_imkey;
 693                 }
 694         }
 695
 696         /* distribute buffers to jobs */
 697         for (i = 0; i < qp->nb_desc; i++) {
 698                 qp->jobs[i].output =
 699                         (uint8_t *)qp->outputs->addr +
 700                         (i % qp->nb_desc) * MLX5_REGEX_MAX_OUTPUT;
 701                 qp->jobs[i].metadata =
 702                         (uint8_t *)qp->metadata->addr +
 703                         (i % qp->nb_desc) * MLX5_REGEX_METADATA_SIZE;
 704                 if (qp->imkey_addr)
 705                         qp->jobs[i].imkey_array = (struct mlx5_klm *)
 706                                 qp->imkey_addr->addr +
 707                                 (i % qp->nb_desc) * MLX5_REGEX_MAX_KLM_NUM;
 708         }
 709
 710         return 0;
 711
 712 err_imkey:
 713         ptr = qp->outputs->addr;
 714         rte_free(ptr);
 715         mlx5_glue->dereg_mr(qp->outputs);
 716 err_output:
 717         ptr = qp->metadata->addr;
 718         rte_free(ptr);
 719         mlx5_glue->dereg_mr(qp->metadata);
 720         return err;
 721 }
 722
 723 int
 724 mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
 725 {
 726         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
 727         struct mlx5_klm klm = { 0 };
 728         struct mlx5_devx_mkey_attr attr = {
 729                 .klm_array = &klm,
 730                 .klm_num = 1,
 731                 .umr_en = 1,
 732         };
 733         uint32_t i;
 734         int err = 0;
 735
 736         qp->jobs = rte_calloc(__func__, qp->nb_desc, sizeof(*qp->jobs), 64);
 737         if (!qp->jobs)
 738                 return -ENOMEM;
 739         err = setup_buffers(priv, qp);
 740         if (err) {
 741                 rte_free(qp->jobs);
 742                 return err;
 743         }
 744
 745         setup_qps(priv, qp);
 746
 747         if (priv->has_umr) {
 748 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 749                 if (regex_get_pdn(priv->pd, &attr.pd)) {
 750                         err = -rte_errno;
 751                         DRV_LOG(ERR, "Failed to get pdn.");
 752                         mlx5_regexdev_teardown_fastpath(priv, qp_id);
 753                         return err;
 754                 }
 755 #endif
 756                 for (i = 0; i < qp->nb_desc; i++) {
 757                         attr.klm_num = MLX5_REGEX_MAX_KLM_NUM;
 758                         attr.klm_array = qp->jobs[i].imkey_array;
 759                         qp->jobs[i].imkey = mlx5_devx_cmd_mkey_create(priv->ctx,
 760                                                                       &attr);
 761                         if (!qp->jobs[i].imkey) {
 762                                 err = -rte_errno;
 763                                 DRV_LOG(ERR, "Failed to allocate imkey.");
 764                                 mlx5_regexdev_teardown_fastpath(priv, qp_id);
 765                         }
 766                 }
 767         }
 768         return err;
 769 }
 770
 771 static void
 772 free_buffers(struct mlx5_regex_qp *qp)
 773 {
 774         if (qp->imkey_addr) {
 775                 mlx5_glue->dereg_mr(qp->imkey_addr);
 776                 rte_free(qp->imkey_addr->addr);
 777         }
 778         if (qp->metadata) {
 779                 mlx5_glue->dereg_mr(qp->metadata);
 780                 rte_free(qp->metadata->addr);
 781         }
 782         if (qp->outputs) {
 783                 mlx5_glue->dereg_mr(qp->outputs);
 784                 rte_free(qp->outputs->addr);
 785         }
 786 }
 787
 788 void
 789 mlx5_regexdev_teardown_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
 790 {
 791         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
 792         uint32_t i;
 793
 794         if (qp) {
 795                 for (i = 0; i < qp->nb_desc; i++) {
 796                         if (qp->jobs[i].imkey)
 797                                 claim_zero(mlx5_devx_cmd_destroy
 798                                                         (qp->jobs[i].imkey));
 799                 }
 800                 free_buffers(qp);
 801                 if (qp->jobs)
 802                         rte_free(qp->jobs);
 803         }
 804 }