xref: /dpdk/drivers/crypto/scheduler/scheduler_roundrobin.c (revision f106eb44c0e50e0113a7da3da02cb196a09f3f3c)
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_ctx, 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_ctx)->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 		ops[i]->sym->session = sess0->sessions[slave_idx];
94 	}
95 
96 	processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
97 			slave->qp_id, ops, nb_ops);
98 
99 	slave->nb_inflight_cops += processed_ops;
100 
101 	rr_qp_ctx->last_enq_slave_idx += 1;
102 	rr_qp_ctx->last_enq_slave_idx %= rr_qp_ctx->nb_slaves;
103 
104 	/* recover session if enqueue is failed */
105 	if (unlikely(processed_ops < nb_ops)) {
106 		for (i = processed_ops; i < nb_ops; i++)
107 			ops[i]->sym->session = sessions[i];
108 	}
109 
110 	return processed_ops;
111 }
112 
113 static uint16_t
114 schedule_enqueue_ordering(void *qp_ctx, struct rte_crypto_op **ops,
115 		uint16_t nb_ops)
116 {
117 	struct scheduler_qp_ctx *gen_qp_ctx = qp_ctx;
118 	struct rr_scheduler_qp_ctx *rr_qp_ctx =
119 			gen_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 = gen_qp_ctx->seqn++;
151 		ops[i + 1]->sym->session = sess1->sessions[slave_idx];
152 		ops[i + 1]->sym->m_src->seqn = gen_qp_ctx->seqn++;
153 		ops[i + 2]->sym->session = sess2->sessions[slave_idx];
154 		ops[i + 2]->sym->m_src->seqn = gen_qp_ctx->seqn++;
155 		ops[i + 3]->sym->session = sess3->sessions[slave_idx];
156 		ops[i + 3]->sym->m_src->seqn = gen_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 		ops[i]->sym->session = sess0->sessions[slave_idx];
172 		ops[i]->sym->m_src->seqn = gen_qp_ctx->seqn++;
173 	}
174 
175 	processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
176 			slave->qp_id, ops, nb_ops);
177 
178 	slave->nb_inflight_cops += processed_ops;
179 
180 	rr_qp_ctx->last_enq_slave_idx += 1;
181 	rr_qp_ctx->last_enq_slave_idx %= rr_qp_ctx->nb_slaves;
182 
183 	/* recover session if enqueue is failed */
184 	if (unlikely(processed_ops < nb_ops)) {
185 		for (i = processed_ops; i < nb_ops; i++)
186 			ops[i]->sym->session = sessions[i];
187 	}
188 
189 	return processed_ops;
190 }
191 
192 
193 static uint16_t
194 schedule_dequeue(void *qp_ctx, struct rte_crypto_op **ops, uint16_t nb_ops)
195 {
196 	struct rr_scheduler_qp_ctx *rr_qp_ctx =
197 			((struct scheduler_qp_ctx *)qp_ctx)->private_qp_ctx;
198 	struct scheduler_slave *slave;
199 	uint32_t last_slave_idx = rr_qp_ctx->last_deq_slave_idx;
200 	uint16_t nb_deq_ops;
201 
202 	if (unlikely(rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops == 0)) {
203 		do {
204 			last_slave_idx += 1;
205 
206 			if (unlikely(last_slave_idx >= rr_qp_ctx->nb_slaves))
207 				last_slave_idx = 0;
208 			/* looped back, means no inflight cops in the queue */
209 			if (last_slave_idx == rr_qp_ctx->last_deq_slave_idx)
210 				return 0;
211 		} while (rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops
212 				== 0);
213 	}
214 
215 	slave = &rr_qp_ctx->slaves[last_slave_idx];
216 
217 	nb_deq_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
218 			slave->qp_id, ops, nb_ops);
219 
220 	last_slave_idx += 1;
221 	last_slave_idx %= rr_qp_ctx->nb_slaves;
222 
223 	rr_qp_ctx->last_deq_slave_idx = last_slave_idx;
224 
225 	slave->nb_inflight_cops -= nb_deq_ops;
226 
227 	return nb_deq_ops;
228 }
229 
230 static uint16_t
231 schedule_dequeue_ordering(void *qp_ctx, struct rte_crypto_op **ops,
232 		uint16_t nb_ops)
233 {
234 	struct scheduler_qp_ctx *gen_qp_ctx = (struct scheduler_qp_ctx *)qp_ctx;
235 	struct rr_scheduler_qp_ctx *rr_qp_ctx = (gen_qp_ctx->private_qp_ctx);
236 	struct scheduler_slave *slave;
237 	struct rte_reorder_buffer *reorder_buff = gen_qp_ctx->reorder_buf;
238 	struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3;
239 	uint16_t nb_deq_ops, nb_drained_mbufs;
240 	const uint16_t nb_op_ops = nb_ops;
241 	struct rte_crypto_op *op_ops[nb_op_ops];
242 	struct rte_mbuf *reorder_mbufs[nb_op_ops];
243 	uint32_t last_slave_idx = rr_qp_ctx->last_deq_slave_idx;
244 	uint16_t i;
245 
246 	if (unlikely(rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops == 0)) {
247 		do {
248 			last_slave_idx += 1;
249 
250 			if (unlikely(last_slave_idx >= rr_qp_ctx->nb_slaves))
251 				last_slave_idx = 0;
252 			/* looped back, means no inflight cops in the queue */
253 			if (last_slave_idx == rr_qp_ctx->last_deq_slave_idx)
254 				return 0;
255 		} while (rr_qp_ctx->slaves[last_slave_idx].nb_inflight_cops
256 				== 0);
257 	}
258 
259 	slave = &rr_qp_ctx->slaves[last_slave_idx];
260 
261 	nb_deq_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
262 			slave->qp_id, op_ops, nb_ops);
263 
264 	rr_qp_ctx->last_deq_slave_idx += 1;
265 	rr_qp_ctx->last_deq_slave_idx %= rr_qp_ctx->nb_slaves;
266 
267 	slave->nb_inflight_cops -= nb_deq_ops;
268 
269 	for (i = 0; i < nb_deq_ops && i < 4; i++)
270 		rte_prefetch0(op_ops[i]->sym->m_src);
271 
272 	for (i = 0; (i < (nb_deq_ops - 8)) && (nb_deq_ops > 8); i += 4) {
273 		mbuf0 = op_ops[i]->sym->m_src;
274 		mbuf1 = op_ops[i + 1]->sym->m_src;
275 		mbuf2 = op_ops[i + 2]->sym->m_src;
276 		mbuf3 = op_ops[i + 3]->sym->m_src;
277 
278 		mbuf0->userdata = op_ops[i];
279 		mbuf1->userdata = op_ops[i + 1];
280 		mbuf2->userdata = op_ops[i + 2];
281 		mbuf3->userdata = op_ops[i + 3];
282 
283 		rte_reorder_insert(reorder_buff, mbuf0);
284 		rte_reorder_insert(reorder_buff, mbuf1);
285 		rte_reorder_insert(reorder_buff, mbuf2);
286 		rte_reorder_insert(reorder_buff, mbuf3);
287 
288 		rte_prefetch0(op_ops[i + 4]->sym->m_src);
289 		rte_prefetch0(op_ops[i + 5]->sym->m_src);
290 		rte_prefetch0(op_ops[i + 6]->sym->m_src);
291 		rte_prefetch0(op_ops[i + 7]->sym->m_src);
292 	}
293 
294 	for (; i < nb_deq_ops; i++) {
295 		mbuf0 = op_ops[i]->sym->m_src;
296 		mbuf0->userdata = op_ops[i];
297 		rte_reorder_insert(reorder_buff, mbuf0);
298 	}
299 
300 	nb_drained_mbufs = rte_reorder_drain(reorder_buff, reorder_mbufs,
301 			nb_ops);
302 	for (i = 0; i < nb_drained_mbufs && i < 4; i++)
303 		rte_prefetch0(reorder_mbufs[i]);
304 
305 	for (i = 0; (i < (nb_drained_mbufs - 8)) && (nb_drained_mbufs > 8);
306 			i += 4) {
307 		ops[i] = *(struct rte_crypto_op **)reorder_mbufs[i]->userdata;
308 		ops[i + 1] = *(struct rte_crypto_op **)
309 			reorder_mbufs[i + 1]->userdata;
310 		ops[i + 2] = *(struct rte_crypto_op **)
311 			reorder_mbufs[i + 2]->userdata;
312 		ops[i + 3] = *(struct rte_crypto_op **)
313 			reorder_mbufs[i + 3]->userdata;
314 
315 		reorder_mbufs[i]->userdata = NULL;
316 		reorder_mbufs[i + 1]->userdata = NULL;
317 		reorder_mbufs[i + 2]->userdata = NULL;
318 		reorder_mbufs[i + 3]->userdata = NULL;
319 
320 		rte_prefetch0(reorder_mbufs[i + 4]);
321 		rte_prefetch0(reorder_mbufs[i + 5]);
322 		rte_prefetch0(reorder_mbufs[i + 6]);
323 		rte_prefetch0(reorder_mbufs[i + 7]);
324 	}
325 
326 	for (; i < nb_drained_mbufs; i++) {
327 		ops[i] = *(struct rte_crypto_op **)
328 			reorder_mbufs[i]->userdata;
329 		reorder_mbufs[i]->userdata = NULL;
330 	}
331 
332 	return nb_drained_mbufs;
333 }
334 
335 static int
336 slave_attach(__rte_unused struct rte_cryptodev *dev,
337 		__rte_unused uint8_t slave_id)
338 {
339 	return 0;
340 }
341 
342 static int
343 slave_detach(__rte_unused struct rte_cryptodev *dev,
344 		__rte_unused uint8_t slave_id)
345 {
346 	return 0;
347 }
348 
349 static int
350 scheduler_start(struct rte_cryptodev *dev)
351 {
352 	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
353 	uint16_t i;
354 
355 	for (i = 0; i < dev->data->nb_queue_pairs; i++) {
356 		struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
357 		struct rr_scheduler_qp_ctx *rr_qp_ctx =
358 				qp_ctx->private_qp_ctx;
359 		uint32_t j;
360 
361 		memset(rr_qp_ctx->slaves, 0, MAX_SLAVES_NUM *
362 				sizeof(struct scheduler_slave));
363 		for (j = 0; j < sched_ctx->nb_slaves; j++) {
364 			rr_qp_ctx->slaves[j].dev_id =
365 					sched_ctx->slaves[j].dev_id;
366 			rr_qp_ctx->slaves[j].qp_id = i;
367 		}
368 
369 		rr_qp_ctx->nb_slaves = sched_ctx->nb_slaves;
370 
371 		rr_qp_ctx->last_enq_slave_idx = 0;
372 		rr_qp_ctx->last_deq_slave_idx = 0;
373 
374 		if (sched_ctx->reordering_enabled) {
375 			qp_ctx->schedule_enqueue = &schedule_enqueue_ordering;
376 			qp_ctx->schedule_dequeue = &schedule_dequeue_ordering;
377 		} else {
378 			qp_ctx->schedule_enqueue = &schedule_enqueue;
379 			qp_ctx->schedule_dequeue = &schedule_dequeue;
380 		}
381 	}
382 
383 	return 0;
384 }
385 
386 static int
387 scheduler_stop(__rte_unused struct rte_cryptodev *dev)
388 {
389 	return 0;
390 }
391 
392 static int
393 scheduler_config_qp(struct rte_cryptodev *dev, uint16_t qp_id)
394 {
395 	struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[qp_id];
396 	struct rr_scheduler_qp_ctx *rr_qp_ctx;
397 
398 	rr_qp_ctx = rte_zmalloc_socket(NULL, sizeof(*rr_qp_ctx), 0,
399 			rte_socket_id());
400 	if (!rr_qp_ctx) {
401 		CS_LOG_ERR("failed allocate memory for private queue pair");
402 		return -ENOMEM;
403 	}
404 
405 	qp_ctx->private_qp_ctx = (void *)rr_qp_ctx;
406 
407 	return 0;
408 }
409 
410 static int
411 scheduler_create_private_ctx(__rte_unused struct rte_cryptodev *dev)
412 {
413 	return 0;
414 }
415 
416 struct rte_cryptodev_scheduler_ops scheduler_rr_ops = {
417 	slave_attach,
418 	slave_detach,
419 	scheduler_start,
420 	scheduler_stop,
421 	scheduler_config_qp,
422 	scheduler_create_private_ctx
423 };
424 
425 struct rte_cryptodev_scheduler scheduler = {
426 		.name = "roundrobin-scheduler",
427 		.description = "scheduler which will round robin burst across "
428 				"slave crypto devices",
429 		.mode = CDEV_SCHED_MODE_ROUNDROBIN,
430 		.ops = &scheduler_rr_ops
431 };
432 
433 struct rte_cryptodev_scheduler *roundrobin_scheduler = &scheduler;
434