regex/mlx5: fix type of setup constants
[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 <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 UINT32_C(64)
29 #define MLX5_REGEX_MAX_OUTPUT RTE_BIT32(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
36 static inline uint32_t
37 sq_size_get(struct mlx5_regex_sq *sq)
38 {
39         return (1U << sq->log_nb_desc);
40 }
41
42 static inline uint32_t
43 cq_size_get(struct mlx5_regex_cq *cq)
44 {
45         return (1U << cq->log_nb_desc);
46 }
47
48 struct mlx5_regex_job {
49         uint64_t user_id;
50         volatile uint8_t *output;
51         volatile uint8_t *metadata;
52 } __rte_cached_aligned;
53
54 static inline void
55 set_data_seg(struct mlx5_wqe_data_seg *seg,
56              uint32_t length, uint32_t lkey,
57              uintptr_t address)
58 {
59         seg->byte_count = rte_cpu_to_be_32(length);
60         seg->lkey = rte_cpu_to_be_32(lkey);
61         seg->addr = rte_cpu_to_be_64(address);
62 }
63
64 static inline void
65 set_metadata_seg(struct mlx5_wqe_metadata_seg *seg,
66                  uint32_t mmo_control_31_0, uint32_t lkey,
67                  uintptr_t address)
68 {
69         seg->mmo_control_31_0 = htobe32(mmo_control_31_0);
70         seg->lkey = rte_cpu_to_be_32(lkey);
71         seg->addr = rte_cpu_to_be_64(address);
72 }
73
74 static inline void
75 set_regex_ctrl_seg(void *seg, uint8_t le, uint16_t subset_id0,
76                    uint16_t subset_id1, uint16_t subset_id2,
77                    uint16_t subset_id3, uint8_t ctrl)
78 {
79         MLX5_SET(regexp_mmo_control, seg, le, le);
80         MLX5_SET(regexp_mmo_control, seg, ctrl, ctrl);
81         MLX5_SET(regexp_mmo_control, seg, subset_id_0, subset_id0);
82         MLX5_SET(regexp_mmo_control, seg, subset_id_1, subset_id1);
83         MLX5_SET(regexp_mmo_control, seg, subset_id_2, subset_id2);
84         MLX5_SET(regexp_mmo_control, seg, subset_id_3, subset_id3);
85 }
86
87 static inline void
88 set_wqe_ctrl_seg(struct mlx5_wqe_ctrl_seg *seg, uint16_t pi, uint8_t opcode,
89                  uint8_t opmod, uint32_t qp_num, uint8_t fm_ce_se, uint8_t ds,
90                  uint8_t signature, uint32_t imm)
91 {
92         seg->opmod_idx_opcode = rte_cpu_to_be_32(((uint32_t)opmod << 24) |
93                                                  ((uint32_t)pi << 8) |
94                                                  opcode);
95         seg->qpn_ds = rte_cpu_to_be_32((qp_num << 8) | ds);
96         seg->fm_ce_se = fm_ce_se;
97         seg->signature = signature;
98         seg->imm = imm;
99 }
100
101 static inline void
102 prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
103          struct mlx5_regex_sq *sq, struct rte_regex_ops *op,
104          struct mlx5_regex_job *job)
105 {
106         size_t wqe_offset = (sq->pi & (sq_size_get(sq) - 1)) * MLX5_SEND_WQE_BB;
107         uint32_t lkey;
108
109         lkey = mlx5_mr_addr2mr_bh(priv->pd, 0,
110                                   &priv->mr_scache, &qp->mr_ctrl,
111                                   rte_pktmbuf_mtod(op->mbuf, uintptr_t),
112                                   !!(op->mbuf->ol_flags & EXT_ATTACHED_MBUF));
113         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
114         int ds = 4; /*  ctrl + meta + input + output */
115
116         set_wqe_ctrl_seg((struct mlx5_wqe_ctrl_seg *)wqe, sq->pi,
117                          MLX5_OPCODE_MMO, MLX5_OPC_MOD_MMO_REGEX, sq->obj->id,
118                          0, ds, 0, 0);
119         set_regex_ctrl_seg(wqe + 12, 0, op->group_id0, op->group_id1,
120                            op->group_id2,
121                            op->group_id3, 0);
122         struct mlx5_wqe_data_seg *input_seg =
123                 (struct mlx5_wqe_data_seg *)(wqe +
124                                              MLX5_REGEX_WQE_GATHER_OFFSET);
125         input_seg->byte_count =
126                 rte_cpu_to_be_32(rte_pktmbuf_data_len(op->mbuf));
127         input_seg->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(op->mbuf,
128                                                             uintptr_t));
129         input_seg->lkey = lkey;
130         job->user_id = op->user_id;
131         sq->db_pi = sq->pi;
132         sq->pi = (sq->pi + 1) & MLX5_REGEX_MAX_WQE_INDEX;
133 }
134
135 static inline void
136 send_doorbell(struct mlx5dv_devx_uar *uar, struct mlx5_regex_sq *sq)
137 {
138         size_t wqe_offset = (sq->db_pi & (sq_size_get(sq) - 1)) *
139                 MLX5_SEND_WQE_BB;
140         uint8_t *wqe = (uint8_t *)sq->wqe + wqe_offset;
141         ((struct mlx5_wqe_ctrl_seg *)wqe)->fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
142         uint64_t *doorbell_addr =
143                 (uint64_t *)((uint8_t *)uar->base_addr + 0x800);
144         rte_io_wmb();
145         sq->dbr[MLX5_SND_DBR] = rte_cpu_to_be_32((sq->db_pi + 1) &
146                                                  MLX5_REGEX_MAX_WQE_INDEX);
147         rte_wmb();
148         *doorbell_addr = *(volatile uint64_t *)wqe;
149         rte_wmb();
150 }
151
152 static inline int
153 can_send(struct mlx5_regex_sq *sq) {
154         return ((uint16_t)(sq->pi - sq->ci) < sq_size_get(sq));
155 }
156
157 static inline uint32_t
158 job_id_get(uint32_t qid, size_t sq_size, size_t index) {
159         return qid * sq_size + (index & (sq_size - 1));
160 }
161
162 uint16_t
163 mlx5_regexdev_enqueue(struct rte_regexdev *dev, uint16_t qp_id,
164                       struct rte_regex_ops **ops, uint16_t nb_ops)
165 {
166         struct mlx5_regex_priv *priv = dev->data->dev_private;
167         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
168         struct mlx5_regex_sq *sq;
169         size_t sqid, job_id, i = 0;
170
171         while ((sqid = ffs(queue->free_sqs))) {
172                 sqid--; /* ffs returns 1 for bit 0 */
173                 sq = &queue->sqs[sqid];
174                 while (can_send(sq)) {
175                         job_id = job_id_get(sqid, sq_size_get(sq), sq->pi);
176                         prep_one(priv, queue, sq, ops[i], &queue->jobs[job_id]);
177                         i++;
178                         if (unlikely(i == nb_ops)) {
179                                 send_doorbell(priv->uar, sq);
180                                 goto out;
181                         }
182                 }
183                 queue->free_sqs &= ~(1 << sqid);
184                 send_doorbell(priv->uar, sq);
185         }
186
187 out:
188         queue->pi += i;
189         return i;
190 }
191
192 #define MLX5_REGEX_RESP_SZ 8
193
194 static inline void
195 extract_result(struct rte_regex_ops *op, struct mlx5_regex_job *job)
196 {
197         size_t j, offset;
198         op->user_id = job->user_id;
199         op->nb_matches = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
200                                            MLX5_REGEX_METADATA_OFF,
201                                            match_count);
202         op->nb_actual_matches = MLX5_GET_VOLATILE(regexp_metadata,
203                                                   job->metadata +
204                                                   MLX5_REGEX_METADATA_OFF,
205                                                   detected_match_count);
206         for (j = 0; j < op->nb_matches; j++) {
207                 offset = MLX5_REGEX_RESP_SZ * j;
208                 op->matches[j].rule_id =
209                         MLX5_GET_VOLATILE(regexp_match_tuple,
210                                           (job->output + offset), rule_id);
211                 op->matches[j].start_offset =
212                         MLX5_GET_VOLATILE(regexp_match_tuple,
213                                           (job->output +  offset), start_ptr);
214                 op->matches[j].len =
215                         MLX5_GET_VOLATILE(regexp_match_tuple,
216                                           (job->output +  offset), length);
217         }
218 }
219
220 static inline volatile struct mlx5_cqe *
221 poll_one(struct mlx5_regex_cq *cq)
222 {
223         volatile struct mlx5_cqe *cqe;
224         size_t next_cqe_offset;
225
226         next_cqe_offset =  (cq->ci & (cq_size_get(cq) - 1));
227         cqe = (volatile struct mlx5_cqe *)(cq->cqe + next_cqe_offset);
228         rte_io_wmb();
229
230         int ret = check_cqe(cqe, cq_size_get(cq), cq->ci);
231
232         if (unlikely(ret == MLX5_CQE_STATUS_ERR)) {
233                 DRV_LOG(ERR, "Completion with error on qp 0x%x",  0);
234                 return NULL;
235         }
236
237         if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN))
238                 return NULL;
239
240         return cqe;
241 }
242
243
244 /**
245  * DPDK callback for dequeue.
246  *
247  * @param dev
248  *   Pointer to the regex dev structure.
249  * @param qp_id
250  *   The queue to enqueue the traffic to.
251  * @param ops
252  *   List of regex ops to dequeue.
253  * @param nb_ops
254  *   Number of ops in ops parameter.
255  *
256  * @return
257  *   Number of packets successfully dequeued (<= pkts_n).
258  */
259 uint16_t
260 mlx5_regexdev_dequeue(struct rte_regexdev *dev, uint16_t qp_id,
261                       struct rte_regex_ops **ops, uint16_t nb_ops)
262 {
263         struct mlx5_regex_priv *priv = dev->data->dev_private;
264         struct mlx5_regex_qp *queue = &priv->qps[qp_id];
265         struct mlx5_regex_cq *cq = &queue->cq;
266         volatile struct mlx5_cqe *cqe;
267         size_t i = 0;
268
269         while ((cqe = poll_one(cq))) {
270                 uint16_t wq_counter
271                         = (rte_be_to_cpu_16(cqe->wqe_counter) + 1) &
272                           MLX5_REGEX_MAX_WQE_INDEX;
273                 size_t sqid = cqe->rsvd3[2];
274                 struct mlx5_regex_sq *sq = &queue->sqs[sqid];
275                 while (sq->ci != wq_counter) {
276                         if (unlikely(i == nb_ops)) {
277                                 /* Return without updating cq->ci */
278                                 goto out;
279                         }
280                         uint32_t job_id = job_id_get(sqid, sq_size_get(sq),
281                                                      sq->ci);
282                         extract_result(ops[i], &queue->jobs[job_id]);
283                         sq->ci = (sq->ci + 1) & MLX5_REGEX_MAX_WQE_INDEX;
284                         i++;
285                 }
286                 cq->ci = (cq->ci + 1) & 0xffffff;
287                 rte_wmb();
288                 cq->dbr[0] = rte_cpu_to_be_32(cq->ci);
289                 queue->free_sqs |= (1 << sqid);
290         }
291
292 out:
293         queue->ci += i;
294         return i;
295 }
296
297 static void
298 setup_sqs(struct mlx5_regex_qp *queue)
299 {
300         size_t sqid, entry;
301         uint32_t job_id;
302         for (sqid = 0; sqid < queue->nb_obj; sqid++) {
303                 struct mlx5_regex_sq *sq = &queue->sqs[sqid];
304                 uint8_t *wqe = (uint8_t *)sq->wqe;
305                 for (entry = 0 ; entry < sq_size_get(sq); entry++) {
306                         job_id = sqid * sq_size_get(sq) + entry;
307                         struct mlx5_regex_job *job = &queue->jobs[job_id];
308
309                         set_metadata_seg((struct mlx5_wqe_metadata_seg *)
310                                          (wqe + MLX5_REGEX_WQE_METADATA_OFFSET),
311                                          0, queue->metadata->lkey,
312                                          (uintptr_t)job->metadata);
313                         set_data_seg((struct mlx5_wqe_data_seg *)
314                                      (wqe + MLX5_REGEX_WQE_SCATTER_OFFSET),
315                                      MLX5_REGEX_MAX_OUTPUT,
316                                      queue->outputs->lkey,
317                                      (uintptr_t)job->output);
318                         wqe += 64;
319                 }
320                 queue->free_sqs |= 1 << sqid;
321         }
322 }
323
324 static int
325 setup_buffers(struct mlx5_regex_qp *qp, struct ibv_pd *pd)
326 {
327         uint32_t i;
328         int err;
329
330         void *ptr = rte_calloc(__func__, qp->nb_desc,
331                                MLX5_REGEX_METADATA_SIZE,
332                                MLX5_REGEX_METADATA_SIZE);
333         if (!ptr)
334                 return -ENOMEM;
335
336         qp->metadata = mlx5_glue->reg_mr(pd, ptr,
337                                          MLX5_REGEX_METADATA_SIZE*qp->nb_desc,
338                                          IBV_ACCESS_LOCAL_WRITE);
339         if (!qp->metadata) {
340                 DRV_LOG(ERR, "Failed to register metadata");
341                 rte_free(ptr);
342                 return -EINVAL;
343         }
344
345         ptr = rte_calloc(__func__, qp->nb_desc,
346                          MLX5_REGEX_MAX_OUTPUT,
347                          MLX5_REGEX_MAX_OUTPUT);
348         if (!ptr) {
349                 err = -ENOMEM;
350                 goto err_output;
351         }
352         qp->outputs = mlx5_glue->reg_mr(pd, ptr,
353                                         MLX5_REGEX_MAX_OUTPUT * qp->nb_desc,
354                                         IBV_ACCESS_LOCAL_WRITE);
355         if (!qp->outputs) {
356                 rte_free(ptr);
357                 DRV_LOG(ERR, "Failed to register output");
358                 err = -EINVAL;
359                 goto err_output;
360         }
361
362         /* distribute buffers to jobs */
363         for (i = 0; i < qp->nb_desc; i++) {
364                 qp->jobs[i].output =
365                         (uint8_t *)qp->outputs->addr +
366                         (i % qp->nb_desc) * MLX5_REGEX_MAX_OUTPUT;
367                 qp->jobs[i].metadata =
368                         (uint8_t *)qp->metadata->addr +
369                         (i % qp->nb_desc) * MLX5_REGEX_METADATA_SIZE;
370         }
371         return 0;
372
373 err_output:
374         ptr = qp->metadata->addr;
375         rte_free(ptr);
376         mlx5_glue->dereg_mr(qp->metadata);
377         return err;
378 }
379
380 int
381 mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
382 {
383         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
384         int err;
385
386         qp->jobs = rte_calloc(__func__, qp->nb_desc, sizeof(*qp->jobs), 64);
387         if (!qp->jobs)
388                 return -ENOMEM;
389         err = setup_buffers(qp, priv->pd);
390         if (err) {
391                 rte_free(qp->jobs);
392                 return err;
393         }
394         setup_sqs(qp);
395         return 0;
396 }
397
398 static void
399 free_buffers(struct mlx5_regex_qp *qp)
400 {
401         if (qp->metadata) {
402                 mlx5_glue->dereg_mr(qp->metadata);
403                 rte_free(qp->metadata->addr);
404         }
405         if (qp->outputs) {
406                 mlx5_glue->dereg_mr(qp->outputs);
407                 rte_free(qp->outputs->addr);
408         }
409 }
410
411 void
412 mlx5_regexdev_teardown_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
413 {
414         struct mlx5_regex_qp *qp = &priv->qps[qp_id];
415
416         if (qp) {
417                 free_buffers(qp);
418                 if (qp->jobs)
419                         rte_free(qp->jobs);
420         }
421 }