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 <sys/mman.h>
   7
   8 #include <rte_malloc.h>
   9 #include <rte_log.h>
  10 #include <rte_errno.h>
  11 #include <rte_bus_pci.h>
  12 #include <rte_pci.h>
  13 #include <rte_regexdev_driver.h>
  14 #include <rte_mbuf.h>
  15
  16 #include <infiniband/mlx5dv.h>
  17 #include <mlx5_glue.h>
  18 #include <mlx5_common.h>
  19 #include <mlx5_prm.h>
  20 #include <strings.h>
  21
  22 #include "mlx5_regex_utils.h"
  23 #include "mlx5_rxp.h"
  24 #include "mlx5_regex.h"
  25
  26 #define MLX5_REGEX_MAX_WQE_INDEX 0xffff
  27 #define MLX5_REGEX_METADATA_SIZE 64
  28 #define MLX5_REGEX_MAX_INPUT (1 << 14)
  29 #define MLX5_REGEX_MAX_OUTPUT (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
  35 static inline uint32_t
  36 sq_size_get(struct mlx5_regex_sq *sq)
  37 {
  38         return (1U << sq->log_nb_desc);
  39 }
  40
  41 struct mlx5_regex_job {
  42         uint64_t user_id;
  43         uint8_t *input;
  44         volatile uint8_t *output;
  45         volatile uint8_t *metadata;
  46 } __rte_cached_aligned;
  47
  48 static inline void
  49 set_data_seg(struct mlx5_wqe_data_seg *seg,
  50              uint32_t length, uint32_t lkey,
  51              uintptr_t address)
  52 {
  53         seg->byte_count = rte_cpu_to_be_32(length);
  54         seg->lkey = rte_cpu_to_be_32(lkey);
  55         seg->addr = rte_cpu_to_be_64(address);
  56 }
  57
  58 static inline void
  59 set_metadata_seg(struct mlx5_wqe_metadata_seg *seg,
  60                  uint32_t mmo_control_31_0, uint32_t lkey,
  61                  uintptr_t address)
  62 {
  63         seg->mmo_control_31_0 = htobe32(mmo_control_31_0);
  64         seg->lkey = rte_cpu_to_be_32(lkey);
  65         seg->addr = rte_cpu_to_be_64(address);
  66 }
  67
  68 static inline void
  69 set_regex_ctrl_seg(void *seg, uint8_t le, uint16_t subset_id0,
  70                    uint16_t subset_id1, uint16_t subset_id2,
  71                    uint16_t subset_id3, uint8_t ctrl)
  72 {
  73         MLX5_SET(regexp_mmo_control, seg, le, le);
  74         MLX5_SET(regexp_mmo_control, seg, ctrl, ctrl);
  75         MLX5_SET(regexp_mmo_control, seg, subset_id_0, subset_id0);
  76         MLX5_SET(regexp_mmo_control, seg, subset_id_1, subset_id1);
  77         MLX5_SET(regexp_mmo_control, seg, subset_id_2, subset_id2);
  78         MLX5_SET(regexp_mmo_control, seg, subset_id_3, subset_id3);
  79 }
  80
  81 static inline void
  82 set_wqe_ctrl_seg(struct mlx5_wqe_ctrl_seg *seg, uint16_t pi, uint8_t opcode,
  83                  uint8_t opmod, uint32_t qp_num, uint8_t fm_ce_se, uint8_t ds,
  84                  uint8_t signature, uint32_t imm)
  85 {
  86         seg->opmod_idx_opcode = rte_cpu_to_be_32(((uint32_t)opmod << 24) |
  87                                                  ((uint32_t)pi << 8) |
  88                                                  opcode);
  89         seg->qpn_ds = rte_cpu_to_be_32((qp_num << 8) | ds);
  90         seg->fm_ce_se = fm_ce_se;
  91         seg->signature = signature;
  92         seg->imm = imm;
  93 }
  94
  95 static inline void
  96 prep_one(struct mlx5_regex_sq *sq, struct rte_regex_ops *op,
  97          struct mlx5_regex_job *job)
  98 {
  99         size_t wqe_offset = (sq->pi & (sq_size_get(sq) - 1)) * MLX5_SEND_WQE_BB;
 100         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
 101         int ds = 4; /*  ctrl + meta + input + output */
 102
 103         memcpy(job->input,
 104                 rte_pktmbuf_mtod(op->mbuf, void *),
 105                 rte_pktmbuf_data_len(op->mbuf));
 106         set_wqe_ctrl_seg((struct mlx5_wqe_ctrl_seg *)wqe, sq->pi,
 107                          MLX5_OPCODE_MMO, MLX5_OPC_MOD_MMO_REGEX, sq->obj->id,
 108                          0, ds, 0, 0);
 109         set_regex_ctrl_seg(wqe + 12, 0, op->group_id0, op->group_id1,
 110                            op->group_id2,
 111                            op->group_id3, 0);
 112         struct mlx5_wqe_data_seg *input_seg =
 113                 (struct mlx5_wqe_data_seg *)(wqe +
 114                                              MLX5_REGEX_WQE_GATHER_OFFSET);
 115         input_seg->byte_count =
 116                 rte_cpu_to_be_32(rte_pktmbuf_data_len(op->mbuf));
 117         job->user_id = op->user_id;
 118         sq->db_pi = sq->pi;
 119         sq->pi = (sq->pi + 1) & MLX5_REGEX_MAX_WQE_INDEX;
 120 }
 121
 122 static inline void
 123 send_doorbell(struct mlx5dv_devx_uar *uar, struct mlx5_regex_sq *sq)
 124 {
 125         size_t wqe_offset = (sq->db_pi & (sq_size_get(sq) - 1)) *
 126                 MLX5_SEND_WQE_BB;
 127         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
 128         ((struct mlx5_wqe_ctrl_seg *)wqe)->fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
 129         uint64_t *doorbell_addr =
 130                 (uint64_t *)((uint8_t *)uar->base_addr + 0x800);
 131         rte_cio_wmb();
 132         sq->dbr[MLX5_SND_DBR] = rte_cpu_to_be_32((sq->db_pi + 1) &
 133                                                  MLX5_REGEX_MAX_WQE_INDEX);
 134         rte_wmb();
 135         *doorbell_addr = *(volatile uint64_t *)wqe;
 136         rte_wmb();
 137 }
 138
 139 static inline int
 140 can_send(struct mlx5_regex_sq *sq) {
 141         return unlikely(sq->ci > sq->pi) ?
 142                         MLX5_REGEX_MAX_WQE_INDEX + sq->pi - sq->ci <
 143                         sq_size_get(sq) :
 144                         sq->pi - sq->ci < sq_size_get(sq);
 145 }
 146
 147 static inline uint32_t
 148 job_id_get(uint32_t qid, size_t sq_size, size_t index) {
 149         return qid * sq_size + index % sq_size;
 150 }
 151
 152 uint16_t
 153 mlx5_regexdev_enqueue(struct rte_regexdev *dev, uint16_t qp_id,
 154                       struct rte_regex_ops **ops, uint16_t nb_ops)
 155 {
 156         struct mlx5_regex_priv *priv = dev->data->dev_private;
 157         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
 158         struct mlx5_regex_sq *sq;
 159         size_t sqid, job_id, i = 0;
 160
 161         while ((sqid = ffs(queue->free_sqs))) {
 162                 sqid--; /* ffs returns 1 for bit 0 */
 163                 sq = &queue->sqs[sqid];
 164                 while (can_send(sq)) {
 165                         job_id = job_id_get(sqid, sq_size_get(sq), sq->pi);
 166                         prep_one(sq, ops[i], &queue->jobs[job_id]);
 167                         i++;
 168                         if (unlikely(i == nb_ops)) {
 169                                 send_doorbell(priv->uar, sq);
 170                                 goto out;
 171                         }
 172                 }
 173                 queue->free_sqs &= ~(1 << sqid);
 174                 send_doorbell(priv->uar, sq);
 175         }
 176
 177 out:
 178         queue->pi += i;
 179         return i;
 180 }
 181
 182 static void
 183 setup_sqs(struct mlx5_regex_qp *queue)
 184 {
 185         size_t sqid, entry;
 186         uint32_t job_id;
 187         for (sqid = 0; sqid < queue->nb_obj; sqid++) {
 188                 struct mlx5_regex_sq *sq = &queue->sqs[sqid];
 189                 uint8_t *wqe = (uint8_t *)sq->wqe;
 190                 for (entry = 0 ; entry < sq_size_get(sq); entry++) {
 191                         job_id = sqid * sq_size_get(sq) + entry;
 192                         struct mlx5_regex_job *job = &queue->jobs[job_id];
 193
 194                         set_metadata_seg((struct mlx5_wqe_metadata_seg *)
 195                                          (wqe + MLX5_REGEX_WQE_METADATA_OFFSET),
 196                                          0, queue->metadata->lkey,
 197                                          (uintptr_t)job->metadata);
 198                         set_data_seg((struct mlx5_wqe_data_seg *)
 199                                      (wqe + MLX5_REGEX_WQE_GATHER_OFFSET),
 200                                      0, queue->inputs->lkey,
 201                                      (uintptr_t)job->input);
 202                         set_data_seg((struct mlx5_wqe_data_seg *)
 203                                      (wqe + MLX5_REGEX_WQE_SCATTER_OFFSET),
 204                                      MLX5_REGEX_MAX_OUTPUT,
 205                                      queue->outputs->lkey,
 206                                      (uintptr_t)job->output);
 207                         wqe += 64;
 208                 }
 209                 queue->free_sqs |= 1 << sqid;
 210         }
 211 }
 212
 213 static int
 214 setup_buffers(struct mlx5_regex_qp *qp, struct ibv_pd *pd)
 215 {
 216         uint32_t i;
 217         int err;
 218
 219         void *ptr = rte_calloc(__func__, qp->nb_desc,
 220                                MLX5_REGEX_METADATA_SIZE,
 221                                MLX5_REGEX_METADATA_SIZE);
 222         if (!ptr)
 223                 return -ENOMEM;
 224
 225         qp->metadata = mlx5_glue->reg_mr(pd, ptr,
 226                                          MLX5_REGEX_METADATA_SIZE*qp->nb_desc,
 227                                          IBV_ACCESS_LOCAL_WRITE);
 228         if (!qp->metadata) {
 229                 rte_free(ptr);
 230                 return -EINVAL;
 231         }
 232         ptr = rte_calloc(__func__, qp->nb_desc,
 233                          MLX5_REGEX_MAX_INPUT,
 234                          MLX5_REGEX_MAX_INPUT);
 235
 236         if (!ptr) {
 237                 err = -ENOMEM;
 238                 goto err_input;
 239         }
 240         qp->inputs = mlx5_glue->reg_mr(pd, ptr,
 241                                        MLX5_REGEX_MAX_INPUT*qp->nb_desc,
 242                                        IBV_ACCESS_LOCAL_WRITE);
 243         if (!qp->inputs) {
 244                 rte_free(ptr);
 245                 err = -EINVAL;
 246                 goto err_input;
 247         }
 248
 249         ptr = rte_calloc(__func__, qp->nb_desc,
 250                          MLX5_REGEX_MAX_OUTPUT,
 251                          MLX5_REGEX_MAX_OUTPUT);
 252         if (!ptr) {
 253                 err = -ENOMEM;
 254                 goto err_output;
 255         }
 256         qp->outputs = mlx5_glue->reg_mr(pd, ptr,
 257                                         MLX5_REGEX_MAX_OUTPUT * qp->nb_desc,
 258                                         IBV_ACCESS_LOCAL_WRITE);
 259         if (!qp->outputs) {
 260                 rte_free(ptr);
 261                 err = -EINVAL;
 262                 goto err_output;
 263         }
 264
 265         /* distribute buffers to jobs */
 266         for (i = 0; i < qp->nb_desc; i++) {
 267                 qp->jobs[i].input =
 268                         (uint8_t *)qp->inputs->addr +
 269                         (i % qp->nb_desc) * MLX5_REGEX_MAX_INPUT;
 270                 qp->jobs[i].output =
 271                         (uint8_t *)qp->outputs->addr +
 272                         (i % qp->nb_desc) * MLX5_REGEX_MAX_OUTPUT;
 273                 qp->jobs[i].metadata =
 274                         (uint8_t *)qp->metadata->addr +
 275                         (i % qp->nb_desc) * MLX5_REGEX_METADATA_SIZE;
 276         }
 277         return 0;
 278
 279 err_output:
 280         ptr = qp->inputs->addr;
 281         rte_free(ptr);
 282         mlx5_glue->dereg_mr(qp->inputs);
 283 err_input:
 284         ptr = qp->metadata->addr;
 285         rte_free(ptr);
 286         mlx5_glue->dereg_mr(qp->metadata);
 287         return err;
 288 }
 289
 290 int
 291 mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
 292 {
 293         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
 294         int err;
 295
 296         qp->jobs = rte_calloc(__func__, qp->nb_desc, sizeof(*qp->jobs),
 297                               sizeof(*qp->jobs));
 298         if (!qp->jobs)
 299                 return -ENOMEM;
 300         err = setup_buffers(qp, priv->pd);
 301         if (err)
 302                 return err;
 303         setup_sqs(qp);
 304         return 0;
 305 }