crypto/scheduler: change enqueue and dequeue functions
[dpdk.git] / drivers / crypto / scheduler / scheduler_roundrobin.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2017 Intel Corporation. All rights reserved.
5  *
6  *   Redistribution and use in source and binary forms, with or without
7  *   modification, are permitted provided that the following conditions
8  *   are met:
9  *
10  *     * Redistributions of source code must retain the above copyright
11  *       notice, this list of conditions and the following disclaimer.
12  *     * Redistributions in binary form must reproduce the above copyright
13  *       notice, this list of conditions and the following disclaimer in
14  *       the documentation and/or other materials provided with the
15  *       distribution.
16  *     * Neither the name of Intel Corporation nor the names of its
17  *       contributors may be used to endorse or promote products derived
18  *       from this software without specific prior written permission.
19  *
20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31  */
32
33 #include <rte_cryptodev.h>
34 #include <rte_malloc.h>
35
36 #include "rte_cryptodev_scheduler_operations.h"
37 #include "scheduler_pmd_private.h"
38
39 struct rr_scheduler_qp_ctx {
40         struct scheduler_slave slaves[MAX_SLAVES_NUM];
41         uint32_t nb_slaves;
42
43         uint32_t last_enq_slave_idx;
44         uint32_t last_deq_slave_idx;
45 };
46
47 static uint16_t
48 schedule_enqueue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
49 {
50         struct rr_scheduler_qp_ctx *rr_qp_ctx =
51                         ((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
52         uint32_t slave_idx = rr_qp_ctx->last_enq_slave_idx;
53         struct scheduler_slave *slave = &rr_qp_ctx->slaves[slave_idx];
54         uint16_t i, processed_ops;
55         struct rte_cryptodev_sym_session *sessions[nb_ops];
56         struct scheduler_session *sess0, *sess1, *sess2, *sess3;
57
58         if (unlikely(nb_ops == 0))
59                 return 0;
60
61         for (i = 0; i < nb_ops && i < 4; i++)
62                 rte_prefetch0(ops[i]->sym->session);
63
64         for (i = 0; (i < (nb_ops - 8)) && (nb_ops > 8); i += 4) {
65                 sess0 = (struct scheduler_session *)
66                                 ops[i]->sym->session->_private;
67                 sess1 = (struct scheduler_session *)
68                                 ops[i+1]->sym->session->_private;
69                 sess2 = (struct scheduler_session *)
70                                 ops[i+2]->sym->session->_private;
71                 sess3 = (struct scheduler_session *)
72                                 ops[i+3]->sym->session->_private;
73
74                 sessions[i] = ops[i]->sym->session;
75                 sessions[i + 1] = ops[i + 1]->sym->session;
76                 sessions[i + 2] = ops[i + 2]->sym->session;
77                 sessions[i + 3] = ops[i + 3]->sym->session;
78
79                 ops[i]->sym->session = sess0->sessions[slave_idx];
80                 ops[i + 1]->sym->session = sess1->sessions[slave_idx];
81                 ops[i + 2]->sym->session = sess2->sessions[slave_idx];
82                 ops[i + 3]->sym->session = sess3->sessions[slave_idx];
83
84                 rte_prefetch0(ops[i + 4]->sym->session);
85                 rte_prefetch0(ops[i + 5]->sym->session);
86                 rte_prefetch0(ops[i + 6]->sym->session);
87                 rte_prefetch0(ops[i + 7]->sym->session);
88         }
89
90         for (; i < nb_ops; i++) {
91                 sess0 = (struct scheduler_session *)
92                                 ops[i]->sym->session->_private;
93                 sessions[i] = ops[i]->sym->session;
94                 ops[i]->sym->session = sess0->sessions[slave_idx];
95         }
96
97         processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
98                         slave->qp_id, ops, nb_ops);
99
100         slave->nb_inflight_cops += processed_ops;
101
102         rr_qp_ctx->last_enq_slave_idx += 1;
103         rr_qp_ctx->last_enq_slave_idx %= rr_qp_ctx->nb_slaves;
104
105         /* recover session if enqueue is failed */
106         if (unlikely(processed_ops < nb_ops)) {
107                 for (i = processed_ops; i < nb_ops; i++)
108                         ops[i]->sym->session = sessions[i];
109         }
110
111         return processed_ops;
112 }
113
114 static uint16_t
115 schedule_enqueue_ordering(void *qp, struct rte_crypto_op **ops,
116                 uint16_t nb_ops)
117 {
118         struct scheduler_qp_ctx *qp_ctx = qp;
119         struct rr_scheduler_qp_ctx *rr_qp_ctx = qp_ctx->private_qp_ctx;
120         uint32_t slave_idx = rr_qp_ctx->last_enq_slave_idx;
121         struct scheduler_slave *slave = &rr_qp_ctx->slaves[slave_idx];
122         uint16_t i, processed_ops;
123         struct rte_cryptodev_sym_session *sessions[nb_ops];
124         struct scheduler_session *sess0, *sess1, *sess2, *sess3;
125
126         if (unlikely(nb_ops == 0))
127                 return 0;
128
129         for (i = 0; i < nb_ops && i < 4; i++) {
130                 rte_prefetch0(ops[i]->sym->session);
131                 rte_prefetch0(ops[i]->sym->m_src);
132         }
133
134         for (i = 0; (i < (nb_ops - 8)) && (nb_ops > 8); i += 4) {
135                 sess0 = (struct scheduler_session *)
136                                 ops[i]->sym->session->_private;
137                 sess1 = (struct scheduler_session *)
138                                 ops[i+1]->sym->session->_private;
139                 sess2 = (struct scheduler_session *)
140                                 ops[i+2]->sym->session->_private;
141                 sess3 = (struct scheduler_session *)
142                                 ops[i+3]->sym->session->_private;
143
144                 sessions[i] = ops[i]->sym->session;
145                 sessions[i + 1] = ops[i + 1]->sym->session;
146                 sessions[i + 2] = ops[i + 2]->sym->session;
147                 sessions[i + 3] = ops[i + 3]->sym->session;
148
149                 ops[i]->sym->session = sess0->sessions[slave_idx];
150                 ops[i]->sym->m_src->seqn = qp_ctx->seqn++;
151                 ops[i + 1]->sym->session = sess1->sessions[slave_idx];
152                 ops[i + 1]->sym->m_src->seqn = qp_ctx->seqn++;
153                 ops[i + 2]->sym->session = sess2->sessions[slave_idx];
154                 ops[i + 2]->sym->m_src->seqn = qp_ctx->seqn++;
155                 ops[i + 3]->sym->session = sess3->sessions[slave_idx];
156                 ops[i + 3]->sym->m_src->seqn = qp_ctx->seqn++;
157
158                 rte_prefetch0(ops[i + 4]->sym->session);
159                 rte_prefetch0(ops[i + 4]->sym->m_src);
160                 rte_prefetch0(ops[i + 5]->sym->session);
161                 rte_prefetch0(ops[i + 5]->sym->m_src);
162                 rte_prefetch0(ops[i + 6]->sym->session);
163                 rte_prefetch0(ops[i + 6]->sym->m_src);
164                 rte_prefetch0(ops[i + 7]->sym->session);
165                 rte_prefetch0(ops[i + 7]->sym->m_src);
166         }
167
168         for (; i < nb_ops; i++) {
169                 sess0 = (struct scheduler_session *)
170                                 ops[i]->sym->session->_private;
171                 sessions[i] = ops[i]->sym->session;
172                 ops[i]->sym->session = sess0->sessions[slave_idx];
173                 ops[i]->sym->m_src->seqn = qp_ctx->seqn++;
174         }
175
176         processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
177                         slave->qp_id, ops, nb_ops);
178
179         slave->nb_inflight_cops += processed_ops;
180
181         rr_qp_ctx->last_enq_slave_idx += 1;
182         rr_qp_ctx->last_enq_slave_idx %= rr_qp_ctx->nb_slaves;
183
184         /* recover session if enqueue is failed */
185         if (unlikely(processed_ops < nb_ops)) {
186                 for (i = processed_ops; i < nb_ops; i++)
187                         ops[i]->sym->session = sessions[i];
188         }
189
190         return processed_ops;
191 }
192
193
194 static uint16_t
195 schedule_dequeue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
196 {
197         struct rr_scheduler_qp_ctx *rr_qp_ctx =
198                         ((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
199         struct scheduler_slave *slave;
200         uint32_t last_slave_idx = rr_qp_ctx->last_deq_slave_idx;
201         uint16_t nb_deq_ops;
202
203         if (unlikely(rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops == 0)) {
204                 do {
205                         last_slave_idx += 1;
206
207                         if (unlikely(last_slave_idx >= rr_qp_ctx->nb_slaves))
208                                 last_slave_idx = 0;
209                         /* looped back, means no inflight cops in the queue */
210                         if (last_slave_idx == rr_qp_ctx->last_deq_slave_idx)
211                                 return 0;
212                 } while (rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops
213                                 == 0);
214         }
215
216         slave = &rr_qp_ctx->slaves[last_slave_idx];
217
218         nb_deq_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
219                         slave->qp_id, ops, nb_ops);
220
221         last_slave_idx += 1;
222         last_slave_idx %= rr_qp_ctx->nb_slaves;
223
224         rr_qp_ctx->last_deq_slave_idx = last_slave_idx;
225
226         slave->nb_inflight_cops -= nb_deq_ops;
227
228         return nb_deq_ops;
229 }
230
231 static uint16_t
232 schedule_dequeue_ordering(void *qp, struct rte_crypto_op **ops,
233                 uint16_t nb_ops)
234 {
235         struct scheduler_qp_ctx *qp_ctx = (struct scheduler_qp_ctx *)qp;
236         struct rr_scheduler_qp_ctx *rr_qp_ctx = (qp_ctx->private_qp_ctx);
237         struct scheduler_slave *slave;
238         struct rte_reorder_buffer *reorder_buff = qp_ctx->reorder_buf;
239         struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3;
240         uint16_t nb_deq_ops, nb_drained_mbufs;
241         const uint16_t nb_op_ops = nb_ops;
242         struct rte_crypto_op *op_ops[nb_op_ops];
243         struct rte_mbuf *reorder_mbufs[nb_op_ops];
244         uint32_t last_slave_idx = rr_qp_ctx->last_deq_slave_idx;
245         uint16_t i;
246
247         if (unlikely(rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops == 0)) {
248                 do {
249                         last_slave_idx += 1;
250
251                         if (unlikely(last_slave_idx >= rr_qp_ctx->nb_slaves))
252                                 last_slave_idx = 0;
253                         /* looped back, means no inflight cops in the queue */
254                         if (last_slave_idx == rr_qp_ctx->last_deq_slave_idx)
255                                 return 0;
256                 } while (rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops
257                                 == 0);
258         }
259
260         slave = &rr_qp_ctx->slaves[last_slave_idx];
261
262         nb_deq_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
263                         slave->qp_id, op_ops, nb_ops);
264
265         rr_qp_ctx->last_deq_slave_idx += 1;
266         rr_qp_ctx->last_deq_slave_idx %= rr_qp_ctx->nb_slaves;
267
268         slave->nb_inflight_cops -= nb_deq_ops;
269
270         for (i = 0; i < nb_deq_ops && i < 4; i++)
271                 rte_prefetch0(op_ops[i]->sym->m_src);
272
273         for (i = 0; (i < (nb_deq_ops - 8)) && (nb_deq_ops > 8); i += 4) {
274                 mbuf0 = op_ops[i]->sym->m_src;
275                 mbuf1 = op_ops[i + 1]->sym->m_src;
276                 mbuf2 = op_ops[i + 2]->sym->m_src;
277                 mbuf3 = op_ops[i + 3]->sym->m_src;
278
279                 mbuf0->userdata = op_ops[i];
280                 mbuf1->userdata = op_ops[i + 1];
281                 mbuf2->userdata = op_ops[i + 2];
282                 mbuf3->userdata = op_ops[i + 3];
283
284                 rte_reorder_insert(reorder_buff, mbuf0);
285                 rte_reorder_insert(reorder_buff, mbuf1);
286                 rte_reorder_insert(reorder_buff, mbuf2);
287                 rte_reorder_insert(reorder_buff, mbuf3);
288
289                 rte_prefetch0(op_ops[i + 4]->sym->m_src);
290                 rte_prefetch0(op_ops[i + 5]->sym->m_src);
291                 rte_prefetch0(op_ops[i + 6]->sym->m_src);
292                 rte_prefetch0(op_ops[i + 7]->sym->m_src);
293         }
294
295         for (; i < nb_deq_ops; i++) {
296                 mbuf0 = op_ops[i]->sym->m_src;
297                 mbuf0->userdata = op_ops[i];
298                 rte_reorder_insert(reorder_buff, mbuf0);
299         }
300
301         nb_drained_mbufs = rte_reorder_drain(reorder_buff, reorder_mbufs,
302                         nb_ops);
303         for (i = 0; i < nb_drained_mbufs && i < 4; i++)
304                 rte_prefetch0(reorder_mbufs[i]);
305
306         for (i = 0; (i < (nb_drained_mbufs - 8)) && (nb_drained_mbufs > 8);
307                         i += 4) {
308                 ops[i] = *(struct rte_crypto_op **)reorder_mbufs[i]->userdata;
309                 ops[i + 1] = *(struct rte_crypto_op **)
310                         reorder_mbufs[i + 1]->userdata;
311                 ops[i + 2] = *(struct rte_crypto_op **)
312                         reorder_mbufs[i + 2]->userdata;
313                 ops[i + 3] = *(struct rte_crypto_op **)
314                         reorder_mbufs[i + 3]->userdata;
315
316                 reorder_mbufs[i]->userdata = NULL;
317                 reorder_mbufs[i + 1]->userdata = NULL;
318                 reorder_mbufs[i + 2]->userdata = NULL;
319                 reorder_mbufs[i + 3]->userdata = NULL;
320
321                 rte_prefetch0(reorder_mbufs[i + 4]);
322                 rte_prefetch0(reorder_mbufs[i + 5]);
323                 rte_prefetch0(reorder_mbufs[i + 6]);
324                 rte_prefetch0(reorder_mbufs[i + 7]);
325         }
326
327         for (; i < nb_drained_mbufs; i++) {
328                 ops[i] = *(struct rte_crypto_op **)
329                         reorder_mbufs[i]->userdata;
330                 reorder_mbufs[i]->userdata = NULL;
331         }
332
333         return nb_drained_mbufs;
334 }
335
336 static int
337 slave_attach(__rte_unused struct rte_cryptodev *dev,
338                 __rte_unused uint8_t slave_id)
339 {
340         return 0;
341 }
342
343 static int
344 slave_detach(__rte_unused struct rte_cryptodev *dev,
345                 __rte_unused uint8_t slave_id)
346 {
347         return 0;
348 }
349
350 static int
351 scheduler_start(struct rte_cryptodev *dev)
352 {
353         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
354         uint16_t i;
355
356         if (sched_ctx->reordering_enabled) {
357                 dev->enqueue_burst = &schedule_enqueue_ordering;
358                 dev->dequeue_burst = &schedule_dequeue_ordering;
359         } else {
360                 dev->enqueue_burst = &schedule_enqueue;
361                 dev->dequeue_burst = &schedule_dequeue;
362         }
363
364         for (i = 0; i < dev->data->nb_queue_pairs; i++) {
365                 struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
366                 struct rr_scheduler_qp_ctx *rr_qp_ctx =
367                                 qp_ctx->private_qp_ctx;
368                 uint32_t j;
369
370                 memset(rr_qp_ctx->slaves, 0, MAX_SLAVES_NUM *
371                                 sizeof(struct scheduler_slave));
372                 for (j = 0; j < sched_ctx->nb_slaves; j++) {
373                         rr_qp_ctx->slaves[j].dev_id =
374                                         sched_ctx->slaves[j].dev_id;
375                         rr_qp_ctx->slaves[j].qp_id = i;
376                 }
377
378                 rr_qp_ctx->nb_slaves = sched_ctx->nb_slaves;
379
380                 rr_qp_ctx->last_enq_slave_idx = 0;
381                 rr_qp_ctx->last_deq_slave_idx = 0;
382         }
383
384         return 0;
385 }
386
387 static int
388 scheduler_stop(__rte_unused struct rte_cryptodev *dev)
389 {
390         return 0;
391 }
392
393 static int
394 scheduler_config_qp(struct rte_cryptodev *dev, uint16_t qp_id)
395 {
396         struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[qp_id];
397         struct rr_scheduler_qp_ctx *rr_qp_ctx;
398
399         rr_qp_ctx = rte_zmalloc_socket(NULL, sizeof(*rr_qp_ctx), 0,
400                         rte_socket_id());
401         if (!rr_qp_ctx) {
402                 CS_LOG_ERR("failed allocate memory for private queue pair");
403                 return -ENOMEM;
404         }
405
406         qp_ctx->private_qp_ctx = (void *)rr_qp_ctx;
407
408         return 0;
409 }
410
411 static int
412 scheduler_create_private_ctx(__rte_unused struct rte_cryptodev *dev)
413 {
414         return 0;
415 }
416
417 struct rte_cryptodev_scheduler_ops scheduler_rr_ops = {
418         slave_attach,
419         slave_detach,
420         scheduler_start,
421         scheduler_stop,
422         scheduler_config_qp,
423         scheduler_create_private_ctx
424 };
425
426 struct rte_cryptodev_scheduler scheduler = {
427                 .name = "roundrobin-scheduler",
428                 .description = "scheduler which will round robin burst across "
429                                 "slave crypto devices",
430                 .mode = CDEV_SCHED_MODE_ROUNDROBIN,
431                 .ops = &scheduler_rr_ops
432 };
433
434 struct rte_cryptodev_scheduler *roundrobin_scheduler = &scheduler;