254954776f705c0d0754e6d1e8f7b61dbeeac1e4
[dpdk.git] / 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_OUTPUT (1 << 11)
29 #define MLX5_REGEX_WQE_CTRL_OFFSET 12
30 #define MLX5_REGEX_WQE_METADATA_OFFSET 16
31 #define MLX5_REGEX_WQE_GATHER_OFFSET 32
32 #define MLX5_REGEX_WQE_SCATTER_OFFSET 48
33 #define MLX5_REGEX_METADATA_OFF 32
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 static inline uint32_t
42 cq_size_get(struct mlx5_regex_cq *cq)
43 {
44         return (1U << cq->log_nb_desc);
45 }
46
47 struct mlx5_regex_job {
48         uint64_t user_id;
49         volatile uint8_t *output;
50         volatile uint8_t *metadata;
51 } __rte_cached_aligned;
52
53 static inline void
54 set_data_seg(struct mlx5_wqe_data_seg *seg,
55              uint32_t length, uint32_t lkey,
56              uintptr_t address)
57 {
58         seg->byte_count = rte_cpu_to_be_32(length);
59         seg->lkey = rte_cpu_to_be_32(lkey);
60         seg->addr = rte_cpu_to_be_64(address);
61 }
62
63 static inline void
64 set_metadata_seg(struct mlx5_wqe_metadata_seg *seg,
65                  uint32_t mmo_control_31_0, uint32_t lkey,
66                  uintptr_t address)
67 {
68         seg->mmo_control_31_0 = htobe32(mmo_control_31_0);
69         seg->lkey = rte_cpu_to_be_32(lkey);
70         seg->addr = rte_cpu_to_be_64(address);
71 }
72
73 static inline void
74 set_regex_ctrl_seg(void *seg, uint8_t le, uint16_t subset_id0,
75                    uint16_t subset_id1, uint16_t subset_id2,
76                    uint16_t subset_id3, uint8_t ctrl)
77 {
78         MLX5_SET(regexp_mmo_control, seg, le, le);
79         MLX5_SET(regexp_mmo_control, seg, ctrl, ctrl);
80         MLX5_SET(regexp_mmo_control, seg, subset_id_0, subset_id0);
81         MLX5_SET(regexp_mmo_control, seg, subset_id_1, subset_id1);
82         MLX5_SET(regexp_mmo_control, seg, subset_id_2, subset_id2);
83         MLX5_SET(regexp_mmo_control, seg, subset_id_3, subset_id3);
84 }
85
86 static inline void
87 set_wqe_ctrl_seg(struct mlx5_wqe_ctrl_seg *seg, uint16_t pi, uint8_t opcode,
88                  uint8_t opmod, uint32_t qp_num, uint8_t fm_ce_se, uint8_t ds,
89                  uint8_t signature, uint32_t imm)
90 {
91         seg->opmod_idx_opcode = rte_cpu_to_be_32(((uint32_t)opmod << 24) |
92                                                  ((uint32_t)pi << 8) |
93                                                  opcode);
94         seg->qpn_ds = rte_cpu_to_be_32((qp_num << 8) | ds);
95         seg->fm_ce_se = fm_ce_se;
96         seg->signature = signature;
97         seg->imm = imm;
98 }
99
100 static inline void
101 prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
102          struct mlx5_regex_sq *sq, struct rte_regex_ops *op,
103          struct mlx5_regex_job *job)
104 {
105         size_t wqe_offset = (sq->pi & (sq_size_get(sq) - 1)) * MLX5_SEND_WQE_BB;
106         uint32_t lkey;
107
108         lkey = mlx5_mr_addr2mr_bh(priv->pd, 0,
109                                   &priv->mr_scache, &qp->mr_ctrl,
110                                   rte_pktmbuf_mtod(op->mbuf, uintptr_t),
111                                   !!(op->mbuf->ol_flags & EXT_ATTACHED_MBUF));
112         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
113         int ds = 4; /*  ctrl + meta + input + output */
114
115         set_wqe_ctrl_seg((struct mlx5_wqe_ctrl_seg *)wqe, sq->pi,
116                          MLX5_OPCODE_MMO, MLX5_OPC_MOD_MMO_REGEX, sq->obj->id,
117                          0, ds, 0, 0);
118         set_regex_ctrl_seg(wqe + 12, 0, op->group_id0, op->group_id1,
119                            op->group_id2,
120                            op->group_id3, 0);
121         struct mlx5_wqe_data_seg *input_seg =
122                 (struct mlx5_wqe_data_seg *)(wqe +
123                                              MLX5_REGEX_WQE_GATHER_OFFSET);
124         input_seg->byte_count =
125                 rte_cpu_to_be_32(rte_pktmbuf_data_len(op->mbuf));
126         input_seg->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(op->mbuf,
127                                                             uintptr_t));
128         input_seg->lkey = lkey;
129         job->user_id = op->user_id;
130         sq->db_pi = sq->pi;
131         sq->pi = (sq->pi + 1) & MLX5_REGEX_MAX_WQE_INDEX;
132 }
133
134 static inline void
135 send_doorbell(struct mlx5dv_devx_uar *uar, struct mlx5_regex_sq *sq)
136 {
137         size_t wqe_offset = (sq->db_pi & (sq_size_get(sq) - 1)) *
138                 MLX5_SEND_WQE_BB;
139         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
140         ((struct mlx5_wqe_ctrl_seg *)wqe)->fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
141         uint64_t *doorbell_addr =
142                 (uint64_t *)((uint8_t *)uar->base_addr + 0x800);
143         rte_io_wmb();
144         sq->dbr[MLX5_SND_DBR] = rte_cpu_to_be_32((sq->db_pi + 1) &
145                                                  MLX5_REGEX_MAX_WQE_INDEX);
146         rte_wmb();
147         *doorbell_addr = *(volatile uint64_t *)wqe;
148         rte_wmb();
149 }
150
151 static inline int
152 can_send(struct mlx5_regex_sq *sq) {
153         return ((uint16_t)(sq->pi - sq->ci) < sq_size_get(sq));
154 }
155
156 static inline uint32_t
157 job_id_get(uint32_t qid, size_t sq_size, size_t index) {
158         return qid * sq_size + (index & (sq_size - 1));
159 }
160
161 uint16_t
162 mlx5_regexdev_enqueue(struct rte_regexdev *dev, uint16_t qp_id,
163                       struct rte_regex_ops **ops, uint16_t nb_ops)
164 {
165         struct mlx5_regex_priv *priv = dev->data->dev_private;
166         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
167         struct mlx5_regex_sq *sq;
168         size_t sqid, job_id, i = 0;
169
170         while ((sqid = ffs(queue->free_sqs))) {
171                 sqid--; /* ffs returns 1 for bit 0 */
172                 sq = &queue->sqs[sqid];
173                 while (can_send(sq)) {
174                         job_id = job_id_get(sqid, sq_size_get(sq), sq->pi);
175                         prep_one(priv, queue, sq, ops[i], &queue->jobs[job_id]);
176                         i++;
177                         if (unlikely(i == nb_ops)) {
178                                 send_doorbell(priv->uar, sq);
179                                 goto out;
180                         }
181                 }
182                 queue->free_sqs &= ~(1 << sqid);
183                 send_doorbell(priv->uar, sq);
184         }
185
186 out:
187         queue->pi += i;
188         return i;
189 }
190
191 #define MLX5_REGEX_RESP_SZ 8
192
193 static inline void
194 extract_result(struct rte_regex_ops *op, struct mlx5_regex_job *job)
195 {
196         size_t j, offset;
197         op->user_id = job->user_id;
198         op->nb_matches = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
199                                            MLX5_REGEX_METADATA_OFF,
200                                            match_count);
201         op->nb_actual_matches = MLX5_GET_VOLATILE(regexp_metadata,
202                                                   job->metadata +
203                                                   MLX5_REGEX_METADATA_OFF,
204                                                   detected_match_count);
205         for (j = 0; j < op->nb_matches; j++) {
206                 offset = MLX5_REGEX_RESP_SZ * j;
207                 op->matches[j].rule_id =
208                         MLX5_GET_VOLATILE(regexp_match_tuple,
209                                           (job->output + offset), rule_id);
210                 op->matches[j].start_offset =
211                         MLX5_GET_VOLATILE(regexp_match_tuple,
212                                           (job->output +  offset), start_ptr);
213                 op->matches[j].len =
214                         MLX5_GET_VOLATILE(regexp_match_tuple,
215                                           (job->output +  offset), length);
216         }
217 }
218
219 static inline volatile struct mlx5_cqe *
220 poll_one(struct mlx5_regex_cq *cq)
221 {
222         volatile struct mlx5_cqe *cqe;
223         size_t next_cqe_offset;
224
225         next_cqe_offset =  (cq->ci & (cq_size_get(cq) - 1));
226         cqe = (volatile struct mlx5_cqe *)(cq->cqe + next_cqe_offset);
227         rte_io_wmb();
228
229         int ret = check_cqe(cqe, cq_size_get(cq), cq->ci);
230
231         if (unlikely(ret == MLX5_CQE_STATUS_ERR)) {
232                 DRV_LOG(ERR, "Completion with error on qp 0x%x",  0);
233                 return NULL;
234         }
235
236         if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN))
237                 return NULL;
238
239         return cqe;
240 }
241
242
243 /**
244  * DPDK callback for dequeue.
245  *
246  * @param dev
247  *   Pointer to the regex dev structure.
248  * @param qp_id
249  *   The queue to enqueue the traffic to.
250  * @param ops
251  *   List of regex ops to dequeue.
252  * @param nb_ops
253  *   Number of ops in ops parameter.
254  *
255  * @return
256  *   Number of packets successfully dequeued (<= pkts_n).
257  */
258 uint16_t
259 mlx5_regexdev_dequeue(struct rte_regexdev *dev, uint16_t qp_id,
260                       struct rte_regex_ops **ops, uint16_t nb_ops)
261 {
262         struct mlx5_regex_priv *priv = dev->data->dev_private;
263         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
264         struct mlx5_regex_cq *cq = &queue->cq;
265         volatile struct mlx5_cqe *cqe;
266         size_t i = 0;
267
268         while ((cqe = poll_one(cq))) {
269                 uint16_t wq_counter
270                         = (rte_be_to_cpu_16(cqe->wqe_counter) + 1) &
271                           MLX5_REGEX_MAX_WQE_INDEX;
272                 size_t sqid = cqe->rsvd3[2];
273                 struct mlx5_regex_sq *sq = &queue->sqs[sqid];
274                 while (sq->ci != wq_counter) {
275                         if (unlikely(i == nb_ops)) {
276                                 /* Return without updating cq->ci */
277                                 goto out;
278                         }
279                         uint32_t job_id = job_id_get(sqid, sq_size_get(sq),
280                                                      sq->ci);
281                         extract_result(ops[i], &queue->jobs[job_id]);
282                         sq->ci = (sq->ci + 1) & MLX5_REGEX_MAX_WQE_INDEX;
283                         i++;
284                 }
285                 cq->ci = (cq->ci + 1) & 0xffffff;
286                 rte_wmb();
287                 cq->dbr[0] = rte_cpu_to_be_32(cq->ci);
288                 queue->free_sqs |= (1 << sqid);
289         }
290
291 out:
292         queue->ci += i;
293         return i;
294 }
295
296 static void
297 setup_sqs(struct mlx5_regex_qp *queue)
298 {
299         size_t sqid, entry;
300         uint32_t job_id;
301         for (sqid = 0; sqid < queue->nb_obj; sqid++) {
302                 struct mlx5_regex_sq *sq = &queue->sqs[sqid];
303                 uint8_t *wqe = (uint8_t *)sq->wqe;
304                 for (entry = 0 ; entry < sq_size_get(sq); entry++) {
305                         job_id = sqid * sq_size_get(sq) + entry;
306                         struct mlx5_regex_job *job = &queue->jobs[job_id];
307
308                         set_metadata_seg((struct mlx5_wqe_metadata_seg *)
309                                          (wqe + MLX5_REGEX_WQE_METADATA_OFFSET),
310                                          0, queue->metadata->lkey,
311                                          (uintptr_t)job->metadata);
312                         set_data_seg((struct mlx5_wqe_data_seg *)
313                                      (wqe + MLX5_REGEX_WQE_SCATTER_OFFSET),
314                                      MLX5_REGEX_MAX_OUTPUT,
315                                      queue->outputs->lkey,
316                                      (uintptr_t)job->output);
317                         wqe += 64;
318                 }
319                 queue->free_sqs |= 1 << sqid;
320         }
321 }
322
323 static int
324 setup_buffers(struct mlx5_regex_qp *qp, struct ibv_pd *pd)
325 {
326         uint32_t i;
327         int err;
328
329         void *ptr = rte_calloc(__func__, qp->nb_desc,
330                                MLX5_REGEX_METADATA_SIZE,
331                                MLX5_REGEX_METADATA_SIZE);
332         if (!ptr)
333                 return -ENOMEM;
334
335         qp->metadata = mlx5_glue->reg_mr(pd, ptr,
336                                          MLX5_REGEX_METADATA_SIZE*qp->nb_desc,
337                                          IBV_ACCESS_LOCAL_WRITE);
338         if (!qp->metadata) {
339                 DRV_LOG(ERR, "Failed to register metadata");
340                 rte_free(ptr);
341                 return -EINVAL;
342         }
343
344         ptr = rte_calloc(__func__, qp->nb_desc,
345                          MLX5_REGEX_MAX_OUTPUT,
346                          MLX5_REGEX_MAX_OUTPUT);
347         if (!ptr) {
348                 err = -ENOMEM;
349                 goto err_output;
350         }
351         qp->outputs = mlx5_glue->reg_mr(pd, ptr,
352                                         MLX5_REGEX_MAX_OUTPUT * qp->nb_desc,
353                                         IBV_ACCESS_LOCAL_WRITE);
354         if (!qp->outputs) {
355                 rte_free(ptr);
356                 DRV_LOG(ERR, "Failed to register output");
357                 err = -EINVAL;
358                 goto err_output;
359         }
360
361         /* distribute buffers to jobs */
362         for (i = 0; i < qp->nb_desc; i++) {
363                 qp->jobs[i].output =
364                         (uint8_t *)qp->outputs->addr +
365                         (i % qp->nb_desc) * MLX5_REGEX_MAX_OUTPUT;
366                 qp->jobs[i].metadata =
367                         (uint8_t *)qp->metadata->addr +
368                         (i % qp->nb_desc) * MLX5_REGEX_METADATA_SIZE;
369         }
370         return 0;
371
372 err_output:
373         ptr = qp->metadata->addr;
374         rte_free(ptr);
375         mlx5_glue->dereg_mr(qp->metadata);
376         return err;
377 }
378
379 int
380 mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
381 {
382         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
383         int err;
384
385         qp->jobs = rte_calloc(__func__, qp->nb_desc, sizeof(*qp->jobs), 64);
386         if (!qp->jobs)
387                 return -ENOMEM;
388         err = setup_buffers(qp, priv->pd);
389         if (err) {
390                 rte_free(qp->jobs);
391                 return err;
392         }
393         setup_sqs(qp);
394         return 0;
395 }
396
397 static void
398 free_buffers(struct mlx5_regex_qp *qp)
399 {
400         if (qp->metadata) {
401                 mlx5_glue->dereg_mr(qp->metadata);
402                 rte_free(qp->metadata->addr);
403         }
404         if (qp->outputs) {
405                 mlx5_glue->dereg_mr(qp->outputs);
406                 rte_free(qp->outputs->addr);
407         }
408 }
409
410 void
411 mlx5_regexdev_teardown_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
412 {
413         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
414
415         if (qp) {
416                 free_buffers(qp);
417                 if (qp->jobs)
418                         rte_free(qp->jobs);
419         }
420 }