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
qp_size_get(struct mlx5_regex_hw_qp * qp)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
cq_size_get(struct mlx5_regex_cq * cq)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
set_data_seg(struct mlx5_wqe_data_seg * seg,uint32_t length,uint32_t lkey,uintptr_t address)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
set_metadata_seg(struct mlx5_wqe_metadata_seg * seg,uint32_t mmo_control_31_0,uint32_t lkey,uintptr_t address)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
set_regex_ctrl_seg(void * seg,uint8_t le,uint16_t subset_id0,uint16_t subset_id1,uint16_t subset_id2,uint16_t subset_id3,uint8_t ctrl)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
set_wqe_ctrl_seg(struct mlx5_wqe_ctrl_seg * seg,uint16_t pi,uint8_t opcode,uint8_t opmod,uint32_t qp_num,uint8_t fm_ce_se,uint8_t ds,uint8_t signature,uint32_t imm)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 static inline void
__prep_one(struct mlx5_regex_priv * priv,struct mlx5_regex_hw_qp * qp_obj,struct rte_regex_ops * op,struct mlx5_regex_job * job,size_t pi,struct mlx5_klm * klm)113 __prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_hw_qp *qp_obj,
114 struct rte_regex_ops *op, struct mlx5_regex_job *job,
115 size_t pi, struct mlx5_klm *klm)
116 {
117 size_t wqe_offset = (pi & (qp_size_get(qp_obj) - 1)) *
118 (MLX5_SEND_WQE_BB << (priv->has_umr ? 2 : 0)) +
119 (priv->has_umr ? MLX5_REGEX_UMR_WQE_SIZE : 0);
120 uint16_t group0 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID0_VALID_F ?
121 op->group_id0 : 0;
122 uint16_t group1 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID1_VALID_F ?
123 op->group_id1 : 0;
124 uint16_t group2 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID2_VALID_F ?
125 op->group_id2 : 0;
126 uint16_t group3 = op->req_flags & RTE_REGEX_OPS_REQ_GROUP_ID3_VALID_F ?
127 op->group_id3 : 0;
128 uint8_t control = op->req_flags &
129 RTE_REGEX_OPS_REQ_MATCH_HIGH_PRIORITY_F ? 1 : 0;
130
131 /* For backward compatibility. */
132 if (!(op->req_flags & (RTE_REGEX_OPS_REQ_GROUP_ID0_VALID_F |
133 RTE_REGEX_OPS_REQ_GROUP_ID1_VALID_F |
134 RTE_REGEX_OPS_REQ_GROUP_ID2_VALID_F |
135 RTE_REGEX_OPS_REQ_GROUP_ID3_VALID_F)))
136 group0 = op->group_id0;
137 uint8_t *wqe = (uint8_t *)(uintptr_t)qp_obj->qp_obj.wqes + wqe_offset;
138 int ds = 4; /* ctrl + meta + input + output */
139
140 set_wqe_ctrl_seg((struct mlx5_wqe_ctrl_seg *)wqe,
141 (priv->has_umr ? (pi * 4 + 3) : pi),
142 MLX5_OPCODE_MMO, MLX5_OPC_MOD_MMO_REGEX,
143 qp_obj->qp_obj.qp->id, 0, ds, 0, 0);
144 set_regex_ctrl_seg(wqe + 12, 0, group0, group1, group2, group3,
145 control);
146 struct mlx5_wqe_data_seg *input_seg =
147 (struct mlx5_wqe_data_seg *)(wqe +
148 MLX5_REGEX_WQE_GATHER_OFFSET);
149 input_seg->byte_count = rte_cpu_to_be_32(klm->byte_count);
150 input_seg->addr = rte_cpu_to_be_64(klm->address);
151 input_seg->lkey = klm->mkey;
152 job->user_id = op->user_id;
153 }
154
155 static inline void
prep_one(struct mlx5_regex_priv * priv,struct mlx5_regex_qp * qp,struct mlx5_regex_hw_qp * qp_obj,struct rte_regex_ops * op,struct mlx5_regex_job * job)156 prep_one(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
157 struct mlx5_regex_hw_qp *qp_obj, struct rte_regex_ops *op,
158 struct mlx5_regex_job *job)
159 {
160 struct mlx5_klm klm;
161
162 klm.byte_count = rte_pktmbuf_data_len(op->mbuf);
163 klm.mkey = mlx5_mr_mb2mr(&qp->mr_ctrl, op->mbuf);
164 klm.address = rte_pktmbuf_mtod(op->mbuf, uintptr_t);
165 __prep_one(priv, qp_obj, op, job, qp_obj->pi, &klm);
166 qp_obj->db_pi = qp_obj->pi;
167 qp_obj->pi = (qp_obj->pi + 1) & MLX5_REGEX_MAX_WQE_INDEX;
168 }
169
170 static inline void
send_doorbell(struct mlx5_regex_priv * priv,struct mlx5_regex_hw_qp * qp)171 send_doorbell(struct mlx5_regex_priv *priv, struct mlx5_regex_hw_qp *qp)
172 {
173 size_t wqe_offset = (qp->db_pi & (qp_size_get(qp) - 1)) *
174 (MLX5_SEND_WQE_BB << (priv->has_umr ? 2 : 0)) +
175 (priv->has_umr ? MLX5_REGEX_UMR_WQE_SIZE : 0);
176 uint8_t *wqe = (uint8_t *)(uintptr_t)qp->qp_obj.wqes + wqe_offset;
177 uint32_t actual_pi = (priv->has_umr ? (qp->db_pi * 4 + 3) : qp->db_pi) &
178 MLX5_REGEX_MAX_WQE_INDEX;
179
180 /* Or the fm_ce_se instead of set, avoid the fence be cleared. */
181 ((struct mlx5_wqe_ctrl_seg *)wqe)->fm_ce_se |= MLX5_WQE_CTRL_CQ_UPDATE;
182 mlx5_doorbell_ring(&priv->uar.bf_db, *(volatile uint64_t *)wqe,
183 actual_pi, &qp->qp_obj.db_rec[MLX5_SND_DBR],
184 !priv->uar.dbnc);
185 }
186
187 static inline int
get_free(struct mlx5_regex_hw_qp * qp,uint8_t has_umr)188 get_free(struct mlx5_regex_hw_qp *qp, uint8_t has_umr) {
189 return (qp_size_get(qp) - ((qp->pi - qp->ci) &
190 (has_umr ? (MLX5_REGEX_MAX_WQE_INDEX >> 2) :
191 MLX5_REGEX_MAX_WQE_INDEX)));
192 }
193
194 static inline uint32_t
job_id_get(uint32_t qid,size_t qp_size,size_t index)195 job_id_get(uint32_t qid, size_t qp_size, size_t index) {
196 return qid * qp_size + (index & (qp_size - 1));
197 }
198
199 #ifdef HAVE_MLX5_UMR_IMKEY
200 static inline int
mkey_klm_available(struct mlx5_klm * klm,uint32_t pos,uint32_t new)201 mkey_klm_available(struct mlx5_klm *klm, uint32_t pos, uint32_t new)
202 {
203 return (klm && ((pos + new) <= MLX5_REGEX_MAX_KLM_NUM));
204 }
205
206 static inline void
complete_umr_wqe(struct mlx5_regex_qp * qp,struct mlx5_regex_hw_qp * qp_obj,struct mlx5_regex_job * mkey_job,size_t umr_index,uint32_t klm_size,uint32_t total_len)207 complete_umr_wqe(struct mlx5_regex_qp *qp, struct mlx5_regex_hw_qp *qp_obj,
208 struct mlx5_regex_job *mkey_job,
209 size_t umr_index, uint32_t klm_size, uint32_t total_len)
210 {
211 size_t wqe_offset = (umr_index & (qp_size_get(qp_obj) - 1)) *
212 (MLX5_SEND_WQE_BB * 4);
213 struct mlx5_wqe_ctrl_seg *wqe = (struct mlx5_wqe_ctrl_seg *)((uint8_t *)
214 (uintptr_t)qp_obj->qp_obj.wqes + wqe_offset);
215 struct mlx5_wqe_umr_ctrl_seg *ucseg =
216 (struct mlx5_wqe_umr_ctrl_seg *)(wqe + 1);
217 struct mlx5_wqe_mkey_context_seg *mkc =
218 (struct mlx5_wqe_mkey_context_seg *)(ucseg + 1);
219 struct mlx5_klm *iklm = (struct mlx5_klm *)(mkc + 1);
220 uint16_t klm_align = RTE_ALIGN(klm_size, 4);
221
222 memset(wqe, 0, MLX5_REGEX_UMR_WQE_SIZE);
223 /* Set WQE control seg. Non-inline KLM UMR WQE size must be 9 WQE_DS. */
224 set_wqe_ctrl_seg(wqe, (umr_index * 4), MLX5_OPCODE_UMR,
225 0, qp_obj->qp_obj.qp->id, 0, 9, 0,
226 rte_cpu_to_be_32(mkey_job->imkey->id));
227 /* Set UMR WQE control seg. */
228 ucseg->mkey_mask |= rte_cpu_to_be_64(MLX5_WQE_UMR_CTRL_MKEY_MASK_LEN |
229 MLX5_WQE_UMR_CTRL_FLAG_TRNSLATION_OFFSET |
230 MLX5_WQE_UMR_CTRL_MKEY_MASK_ACCESS_LOCAL_WRITE);
231 ucseg->klm_octowords = rte_cpu_to_be_16(klm_align);
232 /* Set mkey context seg. */
233 mkc->len = rte_cpu_to_be_64(total_len);
234 mkc->qpn_mkey = rte_cpu_to_be_32(0xffffff00 |
235 (mkey_job->imkey->id & 0xff));
236 /* Set UMR pointer to data seg. */
237 iklm->address = rte_cpu_to_be_64
238 ((uintptr_t)((char *)mkey_job->imkey_array));
239 iklm->mkey = rte_cpu_to_be_32(qp->imkey_addr->lkey);
240 iklm->byte_count = rte_cpu_to_be_32(klm_align);
241 /* Clear the padding memory. */
242 memset((uint8_t *)&mkey_job->imkey_array[klm_size], 0,
243 sizeof(struct mlx5_klm) * (klm_align - klm_size));
244
245 /* Add the following RegEx WQE with fence. */
246 wqe = (struct mlx5_wqe_ctrl_seg *)
247 (((uint8_t *)wqe) + MLX5_REGEX_UMR_WQE_SIZE);
248 wqe->fm_ce_se |= MLX5_WQE_CTRL_INITIATOR_SMALL_FENCE;
249 }
250
251 static inline void
prep_nop_regex_wqe_set(struct mlx5_regex_priv * priv,struct mlx5_regex_hw_qp * qp,struct rte_regex_ops * op,struct mlx5_regex_job * job,size_t pi,struct mlx5_klm * klm)252 prep_nop_regex_wqe_set(struct mlx5_regex_priv *priv,
253 struct mlx5_regex_hw_qp *qp, struct rte_regex_ops *op,
254 struct mlx5_regex_job *job, size_t pi, struct mlx5_klm *klm)
255 {
256 size_t wqe_offset = (pi & (qp_size_get(qp) - 1)) *
257 (MLX5_SEND_WQE_BB << 2);
258 struct mlx5_wqe_ctrl_seg *wqe = (struct mlx5_wqe_ctrl_seg *)((uint8_t *)
259 (uintptr_t)qp->qp_obj.wqes + wqe_offset);
260
261 /* Clear the WQE memory used as UMR WQE previously. */
262 if ((rte_be_to_cpu_32(wqe->opmod_idx_opcode) & 0xff) != MLX5_OPCODE_NOP)
263 memset(wqe, 0, MLX5_REGEX_UMR_WQE_SIZE);
264 /* UMR WQE size is 9 DS, align nop WQE to 3 WQEBBS(12 DS). */
265 set_wqe_ctrl_seg(wqe, pi * 4, MLX5_OPCODE_NOP, 0, qp->qp_obj.qp->id,
266 0, 12, 0, 0);
267 __prep_one(priv, qp, op, job, pi, klm);
268 }
269
270 static inline void
prep_regex_umr_wqe_set(struct mlx5_regex_priv * priv,struct mlx5_regex_qp * qp,struct mlx5_regex_hw_qp * qp_obj,struct rte_regex_ops ** op,size_t nb_ops)271 prep_regex_umr_wqe_set(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp,
272 struct mlx5_regex_hw_qp *qp_obj, struct rte_regex_ops **op,
273 size_t nb_ops)
274 {
275 struct mlx5_regex_job *job = NULL;
276 size_t hw_qpid = qp_obj->qpn, mkey_job_id = 0;
277 size_t left_ops = nb_ops;
278 uint32_t klm_num = 0;
279 uint32_t len = 0;
280 struct mlx5_klm *mkey_klm = NULL;
281 struct mlx5_klm klm;
282 uintptr_t addr;
283
284 while (left_ops--)
285 rte_prefetch0(op[left_ops]);
286 left_ops = nb_ops;
287 /*
288 * Build the WQE set by reverse. In case the burst may consume
289 * multiple mkeys, build the WQE set as normal will hard to
290 * address the last mkey index, since we will only know the last
291 * RegEx WQE's index when finishes building.
292 */
293 while (left_ops--) {
294 struct rte_mbuf *mbuf = op[left_ops]->mbuf;
295 size_t pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, left_ops);
296
297 if (mbuf->nb_segs > 1) {
298 size_t scatter_size = 0;
299
300 if (!mkey_klm_available(mkey_klm, klm_num,
301 mbuf->nb_segs)) {
302 /*
303 * The mkey's KLM is full, create the UMR
304 * WQE in the next WQE set.
305 */
306 if (mkey_klm)
307 complete_umr_wqe(qp, qp_obj,
308 &qp->jobs[mkey_job_id],
309 MLX5_REGEX_UMR_QP_PI_IDX(pi, 1),
310 klm_num, len);
311 /*
312 * Get the indircet mkey and KLM array index
313 * from the last WQE set.
314 */
315 mkey_job_id = job_id_get(hw_qpid,
316 qp_size_get(qp_obj), pi);
317 mkey_klm = qp->jobs[mkey_job_id].imkey_array;
318 klm_num = 0;
319 len = 0;
320 }
321 /* Build RegEx WQE's data segment KLM. */
322 klm.address = len;
323 klm.mkey = rte_cpu_to_be_32
324 (qp->jobs[mkey_job_id].imkey->id);
325 while (mbuf) {
326 addr = rte_pktmbuf_mtod(mbuf, uintptr_t);
327 /* Build indirect mkey seg's KLM. */
328 mkey_klm->mkey = mlx5_mr_mb2mr(&qp->mr_ctrl,
329 mbuf);
330 mkey_klm->address = rte_cpu_to_be_64(addr);
331 mkey_klm->byte_count = rte_cpu_to_be_32
332 (rte_pktmbuf_data_len(mbuf));
333 /*
334 * Save the mbuf's total size for RegEx data
335 * segment.
336 */
337 scatter_size += rte_pktmbuf_data_len(mbuf);
338 mkey_klm++;
339 klm_num++;
340 mbuf = mbuf->next;
341 }
342 len += scatter_size;
343 klm.byte_count = scatter_size;
344 } else {
345 /* The single mubf case. Build the KLM directly. */
346 klm.mkey = mlx5_mr_mb2mr(&qp->mr_ctrl, mbuf);
347 klm.address = rte_pktmbuf_mtod(mbuf, uintptr_t);
348 klm.byte_count = rte_pktmbuf_data_len(mbuf);
349 }
350 job = &qp->jobs[job_id_get(hw_qpid, qp_size_get(qp_obj), pi)];
351 /*
352 * Build the nop + RegEx WQE set by default. The fist nop WQE
353 * will be updated later as UMR WQE if scattered mubf exist.
354 */
355 prep_nop_regex_wqe_set(priv, qp_obj, op[left_ops], job, pi,
356 &klm);
357 }
358 /*
359 * Scattered mbuf have been added to the KLM array. Complete the build
360 * of UMR WQE, update the first nop WQE as UMR WQE.
361 */
362 if (mkey_klm)
363 complete_umr_wqe(qp, qp_obj, &qp->jobs[mkey_job_id], qp_obj->pi,
364 klm_num, len);
365 qp_obj->db_pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, nb_ops - 1);
366 qp_obj->pi = MLX5_REGEX_UMR_QP_PI_IDX(qp_obj->pi, nb_ops);
367 }
368
369 uint16_t
mlx5_regexdev_enqueue_gga(struct rte_regexdev * dev,uint16_t qp_id,struct rte_regex_ops ** ops,uint16_t nb_ops)370 mlx5_regexdev_enqueue_gga(struct rte_regexdev *dev, uint16_t qp_id,
371 struct rte_regex_ops **ops, uint16_t nb_ops)
372 {
373 struct mlx5_regex_priv *priv = dev->data->dev_private;
374 struct mlx5_regex_qp *queue = &priv->qps[qp_id];
375 struct mlx5_regex_hw_qp *qp_obj;
376 size_t hw_qpid, nb_left = nb_ops, nb_desc;
377
378 while ((hw_qpid = ffs(queue->free_qps))) {
379 hw_qpid--; /* ffs returns 1 for bit 0 */
380 qp_obj = &queue->qps[hw_qpid];
381 nb_desc = get_free(qp_obj, priv->has_umr);
382 if (nb_desc) {
383 /* The ops be handled can't exceed nb_ops. */
384 if (nb_desc > nb_left)
385 nb_desc = nb_left;
386 else
387 queue->free_qps &= ~(1 << hw_qpid);
388 prep_regex_umr_wqe_set(priv, queue, qp_obj, ops,
389 nb_desc);
390 send_doorbell(priv, qp_obj);
391 nb_left -= nb_desc;
392 }
393 if (!nb_left)
394 break;
395 ops += nb_desc;
396 }
397 nb_ops -= nb_left;
398 queue->pi += nb_ops;
399 return nb_ops;
400 }
401 #endif
402
403 uint16_t
mlx5_regexdev_enqueue(struct rte_regexdev * dev,uint16_t qp_id,struct rte_regex_ops ** ops,uint16_t nb_ops)404 mlx5_regexdev_enqueue(struct rte_regexdev *dev, uint16_t qp_id,
405 struct rte_regex_ops **ops, uint16_t nb_ops)
406 {
407 struct mlx5_regex_priv *priv = dev->data->dev_private;
408 struct mlx5_regex_qp *queue = &priv->qps[qp_id];
409 struct mlx5_regex_hw_qp *qp_obj;
410 size_t hw_qpid, job_id, i = 0;
411
412 while ((hw_qpid = ffs(queue->free_qps))) {
413 hw_qpid--; /* ffs returns 1 for bit 0 */
414 qp_obj = &queue->qps[hw_qpid];
415 while (get_free(qp_obj, priv->has_umr)) {
416 job_id = job_id_get(hw_qpid, qp_size_get(qp_obj),
417 qp_obj->pi);
418 prep_one(priv, queue, qp_obj, ops[i],
419 &queue->jobs[job_id]);
420 i++;
421 if (unlikely(i == nb_ops)) {
422 send_doorbell(priv, qp_obj);
423 goto out;
424 }
425 }
426 queue->free_qps &= ~(1 << hw_qpid);
427 send_doorbell(priv, qp_obj);
428 }
429
430 out:
431 queue->pi += i;
432 return i;
433 }
434
435 #define MLX5_REGEX_RESP_SZ 8
436
437 static inline void
extract_result(struct rte_regex_ops * op,struct mlx5_regex_job * job)438 extract_result(struct rte_regex_ops *op, struct mlx5_regex_job *job)
439 {
440 size_t j;
441 size_t offset;
442 uint16_t status;
443
444 op->user_id = job->user_id;
445 op->nb_matches = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
446 MLX5_REGEX_METADATA_OFF,
447 match_count);
448 op->nb_actual_matches = MLX5_GET_VOLATILE(regexp_metadata,
449 job->metadata +
450 MLX5_REGEX_METADATA_OFF,
451 detected_match_count);
452 for (j = 0; j < op->nb_matches; j++) {
453 offset = MLX5_REGEX_RESP_SZ * j;
454 op->matches[j].rule_id =
455 MLX5_GET_VOLATILE(regexp_match_tuple,
456 (job->output + offset), rule_id);
457 op->matches[j].start_offset =
458 MLX5_GET_VOLATILE(regexp_match_tuple,
459 (job->output + offset), start_ptr);
460 op->matches[j].len =
461 MLX5_GET_VOLATILE(regexp_match_tuple,
462 (job->output + offset), length);
463 }
464 status = MLX5_GET_VOLATILE(regexp_metadata, job->metadata +
465 MLX5_REGEX_METADATA_OFF,
466 status);
467 op->rsp_flags = 0;
468 if (status & MLX5_RXP_RESP_STATUS_PMI_SOJ)
469 op->rsp_flags |= RTE_REGEX_OPS_RSP_PMI_SOJ_F;
470 if (status & MLX5_RXP_RESP_STATUS_PMI_EOJ)
471 op->rsp_flags |= RTE_REGEX_OPS_RSP_PMI_EOJ_F;
472 if (status & MLX5_RXP_RESP_STATUS_MAX_LATENCY)
473 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_SCAN_TIMEOUT_F;
474 if (status & MLX5_RXP_RESP_STATUS_MAX_MATCH)
475 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_MATCH_F;
476 if (status & MLX5_RXP_RESP_STATUS_MAX_PREFIX)
477 op->rsp_flags |= RTE_REGEX_OPS_RSP_MAX_PREFIX_F;
478 if (status & MLX5_RXP_RESP_STATUS_MAX_PRI_THREADS)
479 op->rsp_flags |= RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
480 if (status & MLX5_RXP_RESP_STATUS_MAX_SEC_THREADS)
481 op->rsp_flags |= RTE_REGEX_OPS_RSP_RESOURCE_LIMIT_REACHED_F;
482 }
483
484 static inline volatile struct mlx5_cqe *
poll_one(struct mlx5_regex_cq * cq)485 poll_one(struct mlx5_regex_cq *cq)
486 {
487 volatile struct mlx5_cqe *cqe;
488 size_t next_cqe_offset;
489
490 next_cqe_offset = (cq->ci & (cq_size_get(cq) - 1));
491 cqe = (volatile struct mlx5_cqe *)(cq->cq_obj.cqes + next_cqe_offset);
492 rte_io_wmb();
493
494 int ret = check_cqe(cqe, cq_size_get(cq), cq->ci);
495
496 if (unlikely(ret == MLX5_CQE_STATUS_ERR)) {
497 DRV_LOG(ERR, "Completion with error on qp 0x%x", 0);
498 return NULL;
499 }
500
501 if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN))
502 return NULL;
503
504 return cqe;
505 }
506
507
508 /**
509 * DPDK callback for dequeue.
510 *
511 * @param dev
512 * Pointer to the regex dev structure.
513 * @param qp_id
514 * The queue to enqueue the traffic to.
515 * @param ops
516 * List of regex ops to dequeue.
517 * @param nb_ops
518 * Number of ops in ops parameter.
519 *
520 * @return
521 * Number of packets successfully dequeued (<= pkts_n).
522 */
523 uint16_t
mlx5_regexdev_dequeue(struct rte_regexdev * dev,uint16_t qp_id,struct rte_regex_ops ** ops,uint16_t nb_ops)524 mlx5_regexdev_dequeue(struct rte_regexdev *dev, uint16_t qp_id,
525 struct rte_regex_ops **ops, uint16_t nb_ops)
526 {
527 struct mlx5_regex_priv *priv = dev->data->dev_private;
528 struct mlx5_regex_qp *queue = &priv->qps[qp_id];
529 struct mlx5_regex_cq *cq = &queue->cq;
530 volatile struct mlx5_cqe *cqe;
531 size_t i = 0;
532
533 while ((cqe = poll_one(cq))) {
534 uint16_t wq_counter
535 = (rte_be_to_cpu_16(cqe->wqe_counter) + 1) &
536 MLX5_REGEX_MAX_WQE_INDEX;
537 size_t hw_qpid = cqe->user_index_bytes[2];
538 struct mlx5_regex_hw_qp *qp_obj = &queue->qps[hw_qpid];
539
540 /* UMR mode WQE counter move as WQE set(4 WQEBBS).*/
541 if (priv->has_umr)
542 wq_counter >>= 2;
543 while (qp_obj->ci != wq_counter) {
544 if (unlikely(i == nb_ops)) {
545 /* Return without updating cq->ci */
546 goto out;
547 }
548 uint32_t job_id = job_id_get(hw_qpid,
549 qp_size_get(qp_obj), qp_obj->ci);
550 extract_result(ops[i], &queue->jobs[job_id]);
551 qp_obj->ci = (qp_obj->ci + 1) & (priv->has_umr ?
552 (MLX5_REGEX_MAX_WQE_INDEX >> 2) :
553 MLX5_REGEX_MAX_WQE_INDEX);
554 i++;
555 }
556 cq->ci = (cq->ci + 1) & 0xffffff;
557 rte_wmb();
558 cq->cq_obj.db_rec[0] = rte_cpu_to_be_32(cq->ci);
559 queue->free_qps |= (1 << hw_qpid);
560 }
561
562 out:
563 queue->ci += i;
564 return i;
565 }
566
567 static void
setup_qps(struct mlx5_regex_priv * priv,struct mlx5_regex_qp * queue)568 setup_qps(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *queue)
569 {
570 size_t hw_qpid, entry;
571 uint32_t job_id;
572 for (hw_qpid = 0; hw_qpid < queue->nb_obj; hw_qpid++) {
573 struct mlx5_regex_hw_qp *qp_obj = &queue->qps[hw_qpid];
574 uint8_t *wqe = (uint8_t *)(uintptr_t)qp_obj->qp_obj.wqes;
575 for (entry = 0 ; entry < qp_size_get(qp_obj); entry++) {
576 job_id = hw_qpid * qp_size_get(qp_obj) + entry;
577 struct mlx5_regex_job *job = &queue->jobs[job_id];
578
579 /* Fill UMR WQE with NOP in advanced. */
580 if (priv->has_umr) {
581 set_wqe_ctrl_seg
582 ((struct mlx5_wqe_ctrl_seg *)wqe,
583 entry * 2, MLX5_OPCODE_NOP, 0,
584 qp_obj->qp_obj.qp->id, 0, 12, 0, 0);
585 wqe += MLX5_REGEX_UMR_WQE_SIZE;
586 }
587 set_metadata_seg((struct mlx5_wqe_metadata_seg *)
588 (wqe + MLX5_REGEX_WQE_METADATA_OFFSET),
589 0, queue->metadata->lkey,
590 (uintptr_t)job->metadata);
591 set_data_seg((struct mlx5_wqe_data_seg *)
592 (wqe + MLX5_REGEX_WQE_SCATTER_OFFSET),
593 MLX5_REGEX_MAX_OUTPUT,
594 queue->outputs->lkey,
595 (uintptr_t)job->output);
596 wqe += 64;
597 }
598 queue->free_qps |= 1 << hw_qpid;
599 }
600 }
601
602 static int
setup_buffers(struct mlx5_regex_priv * priv,struct mlx5_regex_qp * qp)603 setup_buffers(struct mlx5_regex_priv *priv, struct mlx5_regex_qp *qp)
604 {
605 struct ibv_pd *pd = priv->cdev->pd;
606 uint32_t i;
607 int err;
608
609 void *ptr = rte_calloc(__func__, qp->nb_desc,
610 MLX5_REGEX_METADATA_SIZE,
611 MLX5_REGEX_METADATA_SIZE);
612 if (!ptr)
613 return -ENOMEM;
614
615 qp->metadata = mlx5_glue->reg_mr(pd, ptr,
616 MLX5_REGEX_METADATA_SIZE * qp->nb_desc,
617 IBV_ACCESS_LOCAL_WRITE);
618 if (!qp->metadata) {
619 DRV_LOG(ERR, "Failed to register metadata");
620 rte_free(ptr);
621 return -EINVAL;
622 }
623
624 ptr = rte_calloc(__func__, qp->nb_desc,
625 MLX5_REGEX_MAX_OUTPUT,
626 MLX5_REGEX_MAX_OUTPUT);
627 if (!ptr) {
628 err = -ENOMEM;
629 goto err_output;
630 }
631 qp->outputs = mlx5_glue->reg_mr(pd, ptr,
632 MLX5_REGEX_MAX_OUTPUT * qp->nb_desc,
633 IBV_ACCESS_LOCAL_WRITE);
634 if (!qp->outputs) {
635 rte_free(ptr);
636 DRV_LOG(ERR, "Failed to register output");
637 err = -EINVAL;
638 goto err_output;
639 }
640
641 if (priv->has_umr) {
642 ptr = rte_calloc(__func__, qp->nb_desc, MLX5_REGEX_KLMS_SIZE,
643 MLX5_REGEX_KLMS_SIZE);
644 if (!ptr) {
645 err = -ENOMEM;
646 goto err_imkey;
647 }
648 qp->imkey_addr = mlx5_glue->reg_mr(pd, ptr,
649 MLX5_REGEX_KLMS_SIZE * qp->nb_desc,
650 IBV_ACCESS_LOCAL_WRITE);
651 if (!qp->imkey_addr) {
652 rte_free(ptr);
653 DRV_LOG(ERR, "Failed to register output");
654 err = -EINVAL;
655 goto err_imkey;
656 }
657 }
658
659 /* distribute buffers to jobs */
660 for (i = 0; i < qp->nb_desc; i++) {
661 qp->jobs[i].output =
662 (uint8_t *)qp->outputs->addr +
663 (i % qp->nb_desc) * MLX5_REGEX_MAX_OUTPUT;
664 qp->jobs[i].metadata =
665 (uint8_t *)qp->metadata->addr +
666 (i % qp->nb_desc) * MLX5_REGEX_METADATA_SIZE;
667 if (qp->imkey_addr)
668 qp->jobs[i].imkey_array = (struct mlx5_klm *)
669 qp->imkey_addr->addr +
670 (i % qp->nb_desc) * MLX5_REGEX_MAX_KLM_NUM;
671 }
672
673 return 0;
674
675 err_imkey:
676 ptr = qp->outputs->addr;
677 rte_free(ptr);
678 mlx5_glue->dereg_mr(qp->outputs);
679 err_output:
680 ptr = qp->metadata->addr;
681 rte_free(ptr);
682 mlx5_glue->dereg_mr(qp->metadata);
683 return err;
684 }
685
686 int
mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv * priv,uint32_t qp_id)687 mlx5_regexdev_setup_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
688 {
689 struct mlx5_regex_qp *qp = &priv->qps[qp_id];
690 struct mlx5_klm klm = { 0 };
691 struct mlx5_devx_mkey_attr attr = {
692 .klm_array = &klm,
693 .klm_num = 1,
694 .umr_en = 1,
695 };
696 uint32_t i;
697 int err = 0;
698
699 qp->jobs = rte_calloc(__func__, qp->nb_desc, sizeof(*qp->jobs), 64);
700 if (!qp->jobs)
701 return -ENOMEM;
702 err = setup_buffers(priv, qp);
703 if (err) {
704 rte_free(qp->jobs);
705 qp->jobs = NULL;
706 return err;
707 }
708
709 setup_qps(priv, qp);
710
711 if (priv->has_umr) {
712 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
713 attr.pd = priv->cdev->pdn;
714 #endif
715 for (i = 0; i < qp->nb_desc; i++) {
716 attr.klm_num = MLX5_REGEX_MAX_KLM_NUM;
717 attr.klm_array = qp->jobs[i].imkey_array;
718 qp->jobs[i].imkey = mlx5_devx_cmd_mkey_create
719 (priv->cdev->ctx, &attr);
720 if (!qp->jobs[i].imkey) {
721 err = -rte_errno;
722 DRV_LOG(ERR, "Failed to allocate imkey.");
723 mlx5_regexdev_teardown_fastpath(priv, qp_id);
724 }
725 }
726 }
727 return err;
728 }
729
730 static void
free_buffers(struct mlx5_regex_qp * qp)731 free_buffers(struct mlx5_regex_qp *qp)
732 {
733 if (qp->imkey_addr) {
734 mlx5_glue->dereg_mr(qp->imkey_addr);
735 rte_free(qp->imkey_addr->addr);
736 }
737 if (qp->metadata) {
738 mlx5_glue->dereg_mr(qp->metadata);
739 rte_free(qp->metadata->addr);
740 }
741 if (qp->outputs) {
742 mlx5_glue->dereg_mr(qp->outputs);
743 rte_free(qp->outputs->addr);
744 }
745 }
746
747 void
mlx5_regexdev_teardown_fastpath(struct mlx5_regex_priv * priv,uint32_t qp_id)748 mlx5_regexdev_teardown_fastpath(struct mlx5_regex_priv *priv, uint32_t qp_id)
749 {
750 struct mlx5_regex_qp *qp = &priv->qps[qp_id];
751 uint32_t i;
752
753 if (qp->jobs) {
754 for (i = 0; i < qp->nb_desc; i++) {
755 if (qp->jobs[i].imkey)
756 claim_zero(mlx5_devx_cmd_destroy
757 (qp->jobs[i].imkey));
758 }
759 free_buffers(qp);
760 rte_free(qp->jobs);
761 qp->jobs = NULL;
762 }
763 }
764