xref: /freebsd/sys/dev/liquidio/base/lio_request_manager.c (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 /*
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2017 Cavium, Inc.. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Cavium, Inc. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER(S) OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 /*$FreeBSD$*/
34 
35 #include "lio_bsd.h"
36 #include "lio_common.h"
37 #include "lio_droq.h"
38 #include "lio_iq.h"
39 #include "lio_response_manager.h"
40 #include "lio_device.h"
41 #include "lio_main.h"
42 #include "lio_network.h"
43 #include "cn23xx_pf_device.h"
44 #include "lio_rxtx.h"
45 
46 struct lio_iq_post_status {
47 	int	status;
48 	int	index;
49 };
50 
51 static void	lio_check_db_timeout(void *arg, int pending);
52 static void	__lio_check_db_timeout(struct octeon_device *oct,
53 				       uint64_t iq_no);
54 
55 /* Return 0 on success, 1 on failure */
56 int
57 lio_init_instr_queue(struct octeon_device *oct, union octeon_txpciq txpciq,
58 		     uint32_t num_descs)
59 {
60 	struct lio_instr_queue	*iq;
61 	struct lio_iq_config	*conf = NULL;
62 	struct lio_tq		*db_tq;
63 	struct lio_request_list	*request_buf;
64 	bus_size_t		max_size;
65 	uint32_t		iq_no = (uint32_t)txpciq.s.q_no;
66 	uint32_t		q_size;
67 	int			error, i;
68 
69 	if (LIO_CN23XX_PF(oct))
70 		conf = &(LIO_GET_IQ_CFG(LIO_CHIP_CONF(oct, cn23xx_pf)));
71 	if (conf == NULL) {
72 		lio_dev_err(oct, "Unsupported Chip %x\n", oct->chip_id);
73 		return (1);
74 	}
75 
76 	q_size = (uint32_t)conf->instr_type * num_descs;
77 	iq = oct->instr_queue[iq_no];
78 	iq->oct_dev = oct;
79 
80 	max_size = LIO_CN23XX_PKI_MAX_FRAME_SIZE * num_descs;
81 
82 	error = bus_dma_tag_create(bus_get_dma_tag(oct->device),	/* parent */
83 				   1, 0,				/* alignment, bounds */
84 				   BUS_SPACE_MAXADDR,			/* lowaddr */
85 				   BUS_SPACE_MAXADDR,			/* highaddr */
86 				   NULL, NULL,				/* filter, filterarg */
87 				   max_size,				/* maxsize */
88 				   LIO_MAX_SG,				/* nsegments */
89 				   PAGE_SIZE,				/* maxsegsize */
90 				   0,					/* flags */
91 				   NULL,				/* lockfunc */
92 				   NULL,				/* lockfuncarg */
93 				   &iq->txtag);
94 	if (error) {
95 		lio_dev_err(oct, "Cannot allocate memory for instr queue %d\n",
96 			    iq_no);
97 		return (1);
98 	}
99 
100 	iq->base_addr = lio_dma_alloc(q_size, (vm_paddr_t *)&iq->base_addr_dma);
101 	if (!iq->base_addr) {
102 		lio_dev_err(oct, "Cannot allocate memory for instr queue %d\n",
103 			    iq_no);
104 		return (1);
105 	}
106 
107 	iq->max_count = num_descs;
108 
109 	/*
110 	 * Initialize a list to holds requests that have been posted to
111 	 * Octeon but has yet to be fetched by octeon
112 	 */
113 	iq->request_list = malloc(sizeof(*iq->request_list) * num_descs,
114 				  M_DEVBUF, M_NOWAIT | M_ZERO);
115 	if (iq->request_list == NULL) {
116 		lio_dev_err(oct, "Alloc failed for IQ[%d] nr free list\n",
117 			    iq_no);
118 		return (1);
119 	}
120 
121 	lio_dev_dbg(oct, "IQ[%d]: base: %p basedma: %llx count: %d\n",
122 		    iq_no, iq->base_addr, LIO_CAST64(iq->base_addr_dma),
123 		    iq->max_count);
124 
125 	/* Create the descriptor buffer dma maps */
126 	request_buf = iq->request_list;
127 	for (i = 0; i < num_descs; i++, request_buf++) {
128 		error = bus_dmamap_create(iq->txtag, 0, &request_buf->map);
129 		if (error) {
130 			lio_dev_err(oct, "Unable to create TX DMA map\n");
131 			return (1);
132 		}
133 	}
134 
135 	iq->txpciq.txpciq64 = txpciq.txpciq64;
136 	iq->fill_cnt = 0;
137 	iq->host_write_index = 0;
138 	iq->octeon_read_index = 0;
139 	iq->flush_index = 0;
140 	iq->last_db_time = 0;
141 	iq->db_timeout = (uint32_t)conf->db_timeout;
142 	atomic_store_rel_int(&iq->instr_pending, 0);
143 
144 	/* Initialize the lock for this instruction queue */
145 	mtx_init(&iq->lock, "Tx_lock", NULL, MTX_DEF);
146 	mtx_init(&iq->post_lock, "iq_post_lock", NULL, MTX_DEF);
147 	mtx_init(&iq->enq_lock, "enq_lock", NULL, MTX_DEF);
148 
149 	mtx_init(&iq->iq_flush_running_lock, "iq_flush_running_lock", NULL,
150 		 MTX_DEF);
151 
152 	oct->io_qmask.iq |= BIT_ULL(iq_no);
153 
154 	/* Set the 32B/64B mode for each input queue */
155 	oct->io_qmask.iq64B |= ((conf->instr_type == 64) << iq_no);
156 	iq->iqcmd_64B = (conf->instr_type == 64);
157 
158 	oct->fn_list.setup_iq_regs(oct, iq_no);
159 
160 	db_tq = &oct->check_db_tq[iq_no];
161 	db_tq->tq = taskqueue_create("lio_check_db_timeout", M_WAITOK,
162 				     taskqueue_thread_enqueue, &db_tq->tq);
163 	if (db_tq->tq == NULL) {
164 		lio_dev_err(oct, "check db wq create failed for iq %d\n",
165 			    iq_no);
166 		return (1);
167 	}
168 
169 	TIMEOUT_TASK_INIT(db_tq->tq, &db_tq->work, 0, lio_check_db_timeout,
170 			  (void *)db_tq);
171 	db_tq->ctxul = iq_no;
172 	db_tq->ctxptr = oct;
173 
174 	taskqueue_start_threads(&db_tq->tq, 1, PI_NET,
175 				"lio%d_check_db_timeout:%d",
176 				oct->octeon_id, iq_no);
177 	taskqueue_enqueue_timeout(db_tq->tq, &db_tq->work, 1);
178 
179 	/* Allocate a buf ring */
180 	oct->instr_queue[iq_no]->br =
181 		buf_ring_alloc(LIO_BR_SIZE, M_DEVBUF, M_WAITOK,
182 			       &oct->instr_queue[iq_no]->enq_lock);
183 	if (oct->instr_queue[iq_no]->br == NULL) {
184 		lio_dev_err(oct, "Critical Failure setting up buf ring\n");
185 		return (1);
186 	}
187 
188 	return (0);
189 }
190 
191 int
192 lio_delete_instr_queue(struct octeon_device *oct, uint32_t iq_no)
193 {
194 	struct lio_instr_queue		*iq = oct->instr_queue[iq_no];
195 	struct lio_request_list		*request_buf;
196 	struct lio_mbuf_free_info	*finfo;
197 	uint64_t			desc_size = 0, q_size;
198 	int				i;
199 
200 	lio_dev_dbg(oct, "%s[%d]\n", __func__, iq_no);
201 
202 	if (oct->check_db_tq[iq_no].tq != NULL) {
203 		while (taskqueue_cancel_timeout(oct->check_db_tq[iq_no].tq,
204 						&oct->check_db_tq[iq_no].work,
205 						NULL))
206 			taskqueue_drain_timeout(oct->check_db_tq[iq_no].tq,
207 						&oct->check_db_tq[iq_no].work);
208 		taskqueue_free(oct->check_db_tq[iq_no].tq);
209 		oct->check_db_tq[iq_no].tq = NULL;
210 	}
211 
212 	if (LIO_CN23XX_PF(oct))
213 		desc_size =
214 		    LIO_GET_IQ_INSTR_TYPE_CFG(LIO_CHIP_CONF(oct, cn23xx_pf));
215 
216 	request_buf = iq->request_list;
217 	for (i = 0; i < iq->max_count; i++, request_buf++) {
218 		if ((request_buf->reqtype == LIO_REQTYPE_NORESP_NET) ||
219 		    (request_buf->reqtype == LIO_REQTYPE_NORESP_NET_SG)) {
220 			if (request_buf->buf != NULL) {
221 				finfo = request_buf->buf;
222 				bus_dmamap_sync(iq->txtag, request_buf->map,
223 						BUS_DMASYNC_POSTWRITE);
224 				bus_dmamap_unload(iq->txtag,
225 						  request_buf->map);
226 				m_freem(finfo->mb);
227 				request_buf->buf = NULL;
228 				if (request_buf->map != NULL) {
229 					bus_dmamap_destroy(iq->txtag,
230 							   request_buf->map);
231 					request_buf->map = NULL;
232 				}
233 			} else if (request_buf->map != NULL) {
234 				bus_dmamap_unload(iq->txtag, request_buf->map);
235 				bus_dmamap_destroy(iq->txtag, request_buf->map);
236 				request_buf->map = NULL;
237 			}
238 		}
239 	}
240 
241 	if (iq->br != NULL) {
242 		buf_ring_free(iq->br, M_DEVBUF);
243 		iq->br = NULL;
244 	}
245 
246 	if (iq->request_list != NULL) {
247 		free(iq->request_list, M_DEVBUF);
248 		iq->request_list = NULL;
249 	}
250 
251 	if (iq->txtag != NULL) {
252 		bus_dma_tag_destroy(iq->txtag);
253 		iq->txtag = NULL;
254 	}
255 
256 	if (iq->base_addr) {
257 		q_size = iq->max_count * desc_size;
258 		lio_dma_free((uint32_t)q_size, iq->base_addr);
259 
260 		oct->io_qmask.iq &= ~(1ULL << iq_no);
261 		bzero(oct->instr_queue[iq_no], sizeof(struct lio_instr_queue));
262 		oct->num_iqs--;
263 
264 		return (0);
265 	}
266 
267 	return (1);
268 }
269 
270 /* Return 0 on success, 1 on failure */
271 int
272 lio_setup_iq(struct octeon_device *oct, int ifidx, int q_index,
273 	     union octeon_txpciq txpciq, uint32_t num_descs)
274 {
275 	uint32_t	iq_no = (uint32_t)txpciq.s.q_no;
276 
277 	if (oct->instr_queue[iq_no]->oct_dev != NULL) {
278 		lio_dev_dbg(oct, "IQ is in use. Cannot create the IQ: %d again\n",
279 			    iq_no);
280 		oct->instr_queue[iq_no]->txpciq.txpciq64 = txpciq.txpciq64;
281 		return (0);
282 	}
283 
284 	oct->instr_queue[iq_no]->q_index = q_index;
285 	oct->instr_queue[iq_no]->ifidx = ifidx;
286 
287 	if (lio_init_instr_queue(oct, txpciq, num_descs)) {
288 		lio_delete_instr_queue(oct, iq_no);
289 		return (1);
290 	}
291 
292 	oct->num_iqs++;
293 	if (oct->fn_list.enable_io_queues(oct))
294 		return (1);
295 
296 	return (0);
297 }
298 
299 int
300 lio_wait_for_instr_fetch(struct octeon_device *oct)
301 {
302 	int	i, retry = 1000, pending, instr_cnt = 0;
303 
304 	do {
305 		instr_cnt = 0;
306 
307 		for (i = 0; i < LIO_MAX_INSTR_QUEUES(oct); i++) {
308 			if (!(oct->io_qmask.iq & BIT_ULL(i)))
309 				continue;
310 			pending = atomic_load_acq_int(
311 					&oct->instr_queue[i]->instr_pending);
312 			if (pending)
313 				__lio_check_db_timeout(oct, i);
314 			instr_cnt += pending;
315 		}
316 
317 		if (instr_cnt == 0)
318 			break;
319 
320 		lio_sleep_timeout(1);
321 
322 	} while (retry-- && instr_cnt);
323 
324 	return (instr_cnt);
325 }
326 
327 static inline void
328 lio_ring_doorbell(struct octeon_device *oct, struct lio_instr_queue *iq)
329 {
330 
331 	if (atomic_load_acq_int(&oct->status) == LIO_DEV_RUNNING) {
332 		lio_write_csr32(oct, iq->doorbell_reg, iq->fill_cnt);
333 		/* make sure doorbell write goes through */
334 		__compiler_membar();
335 		iq->fill_cnt = 0;
336 		iq->last_db_time = ticks;
337 		return;
338 	}
339 }
340 
341 static inline void
342 __lio_copy_cmd_into_iq(struct lio_instr_queue *iq, uint8_t *cmd)
343 {
344 	uint8_t	*iqptr, cmdsize;
345 
346 	cmdsize = ((iq->iqcmd_64B) ? 64 : 32);
347 	iqptr = iq->base_addr + (cmdsize * iq->host_write_index);
348 
349 	memcpy(iqptr, cmd, cmdsize);
350 }
351 
352 static inline struct lio_iq_post_status
353 __lio_post_command2(struct lio_instr_queue *iq, uint8_t *cmd)
354 {
355 	struct lio_iq_post_status	st;
356 
357 	st.status = LIO_IQ_SEND_OK;
358 
359 	/*
360 	 * This ensures that the read index does not wrap around to the same
361 	 * position if queue gets full before Octeon could fetch any instr.
362 	 */
363 	if (atomic_load_acq_int(&iq->instr_pending) >=
364 	    (int32_t)(iq->max_count - 1)) {
365 		st.status = LIO_IQ_SEND_FAILED;
366 		st.index = -1;
367 		return (st);
368 	}
369 
370 	if (atomic_load_acq_int(&iq->instr_pending) >=
371 	    (int32_t)(iq->max_count - 2))
372 		st.status = LIO_IQ_SEND_STOP;
373 
374 	__lio_copy_cmd_into_iq(iq, cmd);
375 
376 	/* "index" is returned, host_write_index is modified. */
377 	st.index = iq->host_write_index;
378 	iq->host_write_index = lio_incr_index(iq->host_write_index, 1,
379 					      iq->max_count);
380 	iq->fill_cnt++;
381 
382 	/*
383 	 * Flush the command into memory. We need to be sure the data is in
384 	 * memory before indicating that the instruction is pending.
385 	 */
386 	wmb();
387 
388 	atomic_add_int(&iq->instr_pending, 1);
389 
390 	return (st);
391 }
392 
393 static inline void
394 __lio_add_to_request_list(struct lio_instr_queue *iq, int idx, void *buf,
395 			  int reqtype)
396 {
397 
398 	iq->request_list[idx].buf = buf;
399 	iq->request_list[idx].reqtype = reqtype;
400 }
401 
402 /* Can only run in process context */
403 int
404 lio_process_iq_request_list(struct octeon_device *oct,
405 			    struct lio_instr_queue *iq, uint32_t budget)
406 {
407 	struct lio_soft_command		*sc;
408 	struct octeon_instr_irh		*irh = NULL;
409 	void				*buf;
410 	uint32_t			inst_count = 0;
411 	uint32_t			old = iq->flush_index;
412 	int				reqtype;
413 
414 	while (old != iq->octeon_read_index) {
415 		reqtype = iq->request_list[old].reqtype;
416 		buf = iq->request_list[old].buf;
417 
418 		if (reqtype == LIO_REQTYPE_NONE)
419 			goto skip_this;
420 
421 		switch (reqtype) {
422 		case LIO_REQTYPE_NORESP_NET:
423 			lio_free_mbuf(iq, buf);
424 			break;
425 		case LIO_REQTYPE_NORESP_NET_SG:
426 			lio_free_sgmbuf(iq, buf);
427 			break;
428 		case LIO_REQTYPE_RESP_NET:
429 		case LIO_REQTYPE_SOFT_COMMAND:
430 			sc = buf;
431 			if (LIO_CN23XX_PF(oct))
432 				irh = (struct octeon_instr_irh *)
433 					&sc->cmd.cmd3.irh;
434 			if (irh->rflag) {
435 				/*
436 				 * We're expecting a response from Octeon.
437 				 * It's up to lio_process_ordered_list() to
438 				 * process  sc. Add sc to the ordered soft
439 				 * command response list because we expect
440 				 * a response from Octeon.
441 				 */
442 				mtx_lock(&oct->response_list
443 					 [LIO_ORDERED_SC_LIST].lock);
444 				atomic_add_int(&oct->response_list
445 					       [LIO_ORDERED_SC_LIST].
446 					       pending_req_count, 1);
447 				STAILQ_INSERT_TAIL(&oct->response_list
448 						   [LIO_ORDERED_SC_LIST].
449 						   head, &sc->node, entries);
450 				mtx_unlock(&oct->response_list
451 					   [LIO_ORDERED_SC_LIST].lock);
452 			} else {
453 				if (sc->callback != NULL) {
454 					/* This callback must not sleep */
455 					sc->callback(oct, LIO_REQUEST_DONE,
456 						     sc->callback_arg);
457 				}
458 			}
459 
460 			break;
461 		default:
462 			lio_dev_err(oct, "%s Unknown reqtype: %d buf: %p at idx %d\n",
463 				    __func__, reqtype, buf, old);
464 		}
465 
466 		iq->request_list[old].buf = NULL;
467 		iq->request_list[old].reqtype = 0;
468 
469 skip_this:
470 		inst_count++;
471 		old = lio_incr_index(old, 1, iq->max_count);
472 
473 		if ((budget) && (inst_count >= budget))
474 			break;
475 	}
476 
477 	iq->flush_index = old;
478 
479 	return (inst_count);
480 }
481 
482 /* Can only be called from process context */
483 int
484 lio_flush_iq(struct octeon_device *oct, struct lio_instr_queue *iq,
485 	     uint32_t budget)
486 {
487 	uint32_t	inst_processed = 0;
488 	uint32_t	tot_inst_processed = 0;
489 	int		tx_done = 1;
490 
491 	if (!mtx_trylock(&iq->iq_flush_running_lock))
492 		return (tx_done);
493 
494 	mtx_lock(&iq->lock);
495 
496 	iq->octeon_read_index = oct->fn_list.update_iq_read_idx(iq);
497 
498 	do {
499 		/* Process any outstanding IQ packets. */
500 		if (iq->flush_index == iq->octeon_read_index)
501 			break;
502 
503 		if (budget)
504 			inst_processed =
505 				lio_process_iq_request_list(oct, iq,
506 							    budget -
507 							    tot_inst_processed);
508 		else
509 			inst_processed =
510 				lio_process_iq_request_list(oct, iq, 0);
511 
512 		if (inst_processed) {
513 			atomic_subtract_int(&iq->instr_pending, inst_processed);
514 			iq->stats.instr_processed += inst_processed;
515 		}
516 		tot_inst_processed += inst_processed;
517 		inst_processed = 0;
518 
519 	} while (tot_inst_processed < budget);
520 
521 	if (budget && (tot_inst_processed >= budget))
522 		tx_done = 0;
523 
524 	iq->last_db_time = ticks;
525 
526 	mtx_unlock(&iq->lock);
527 
528 	mtx_unlock(&iq->iq_flush_running_lock);
529 
530 	return (tx_done);
531 }
532 
533 /*
534  * Process instruction queue after timeout.
535  * This routine gets called from a taskqueue or when removing the module.
536  */
537 static void
538 __lio_check_db_timeout(struct octeon_device *oct, uint64_t iq_no)
539 {
540 	struct lio_instr_queue	*iq;
541 	uint64_t		next_time;
542 
543 	if (oct == NULL)
544 		return;
545 
546 	iq = oct->instr_queue[iq_no];
547 	if (iq == NULL)
548 		return;
549 
550 	if (atomic_load_acq_int(&iq->instr_pending)) {
551 		/* If ticks - last_db_time < db_timeout do nothing  */
552 		next_time = iq->last_db_time + lio_ms_to_ticks(iq->db_timeout);
553 		if (!lio_check_timeout(ticks, next_time))
554 			return;
555 
556 		iq->last_db_time = ticks;
557 
558 		/* Flush the instruction queue */
559 		lio_flush_iq(oct, iq, 0);
560 
561 		lio_enable_irq(NULL, iq);
562 	}
563 
564 	if (oct->props.ifp != NULL && iq->br != NULL) {
565 		if (mtx_trylock(&iq->enq_lock)) {
566 			if (!drbr_empty(oct->props.ifp, iq->br))
567 				lio_mq_start_locked(oct->props.ifp, iq);
568 
569 			mtx_unlock(&iq->enq_lock);
570 		}
571 	}
572 }
573 
574 /*
575  * Called by the Poll thread at regular intervals to check the instruction
576  * queue for commands to be posted and for commands that were fetched by Octeon.
577  */
578 static void
579 lio_check_db_timeout(void *arg, int pending)
580 {
581 	struct lio_tq		*db_tq = (struct lio_tq *)arg;
582 	struct octeon_device	*oct = db_tq->ctxptr;
583 	uint64_t		iq_no = db_tq->ctxul;
584 	uint32_t		delay = 10;
585 
586 	__lio_check_db_timeout(oct, iq_no);
587 	taskqueue_enqueue_timeout(db_tq->tq, &db_tq->work,
588 				  lio_ms_to_ticks(delay));
589 }
590 
591 int
592 lio_send_command(struct octeon_device *oct, uint32_t iq_no,
593 		 uint32_t force_db, void *cmd, void *buf,
594 		 uint32_t datasize, uint32_t reqtype)
595 {
596 	struct lio_iq_post_status	st;
597 	struct lio_instr_queue		*iq = oct->instr_queue[iq_no];
598 
599 	/*
600 	 * Get the lock and prevent other tasks and tx interrupt handler
601 	 * from running.
602 	 */
603 	mtx_lock(&iq->post_lock);
604 
605 	st = __lio_post_command2(iq, cmd);
606 
607 	if (st.status != LIO_IQ_SEND_FAILED) {
608 		__lio_add_to_request_list(iq, st.index, buf, reqtype);
609 		LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, bytes_sent, datasize);
610 		LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_posted, 1);
611 
612 		if (force_db || (st.status == LIO_IQ_SEND_STOP))
613 			lio_ring_doorbell(oct, iq);
614 	} else {
615 		LIO_INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_dropped, 1);
616 	}
617 
618 	mtx_unlock(&iq->post_lock);
619 
620 	/*
621 	 * This is only done here to expedite packets being flushed for
622 	 * cases where there are no IQ completion interrupts.
623 	 */
624 
625 	return (st.status);
626 }
627 
628 void
629 lio_prepare_soft_command(struct octeon_device *oct, struct lio_soft_command *sc,
630 			 uint8_t opcode, uint8_t subcode, uint32_t irh_ossp,
631 			 uint64_t ossp0, uint64_t ossp1)
632 {
633 	struct octeon_instr_ih3		*ih3;
634 	struct octeon_instr_pki_ih3	*pki_ih3;
635 	struct octeon_instr_irh		*irh;
636 	struct octeon_instr_rdp		*rdp;
637 
638 	KASSERT(opcode <= 15, ("%s, %d, opcode > 15", __func__, __LINE__));
639 	KASSERT(subcode <= 127, ("%s, %d, opcode > 127", __func__, __LINE__));
640 
641 	if (LIO_CN23XX_PF(oct)) {
642 		ih3 = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
643 
644 		ih3->pkind = oct->instr_queue[sc->iq_no]->txpciq.s.pkind;
645 
646 		pki_ih3 = (struct octeon_instr_pki_ih3 *)&sc->cmd.cmd3.pki_ih3;
647 
648 		pki_ih3->w = 1;
649 		pki_ih3->raw = 1;
650 		pki_ih3->utag = 1;
651 		pki_ih3->uqpg = oct->instr_queue[sc->iq_no]->txpciq.s.use_qpg;
652 		pki_ih3->utt = 1;
653 		pki_ih3->tag = LIO_CONTROL;
654 		pki_ih3->tagtype = LIO_ATOMIC_TAG;
655 		pki_ih3->qpg = oct->instr_queue[sc->iq_no]->txpciq.s.qpg;
656 		pki_ih3->pm = 0x7;
657 		pki_ih3->sl = 8;
658 
659 		if (sc->datasize)
660 			ih3->dlengsz = sc->datasize;
661 
662 		irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
663 		irh->opcode = opcode;
664 		irh->subcode = subcode;
665 
666 		/* opcode/subcode specific parameters (ossp) */
667 		irh->ossp = irh_ossp;
668 		sc->cmd.cmd3.ossp[0] = ossp0;
669 		sc->cmd.cmd3.ossp[1] = ossp1;
670 
671 		if (sc->rdatasize) {
672 			rdp = (struct octeon_instr_rdp *)&sc->cmd.cmd3.rdp;
673 			rdp->pcie_port = oct->pcie_port;
674 			rdp->rlen = sc->rdatasize;
675 
676 			irh->rflag = 1;
677 			/* PKI IH3 */
678 			/* pki_ih3 irh+ossp[0]+ossp[1]+rdp+rptr = 48 bytes */
679 			ih3->fsz = LIO_SOFTCMDRESP_IH3;
680 		} else {
681 			irh->rflag = 0;
682 			/* PKI IH3 */
683 			/* pki_h3 + irh + ossp[0] + ossp[1] = 32 bytes */
684 			ih3->fsz = LIO_PCICMD_O3;
685 		}
686 	}
687 }
688 
689 int
690 lio_send_soft_command(struct octeon_device *oct, struct lio_soft_command *sc)
691 {
692 	struct octeon_instr_ih3	*ih3;
693 	struct octeon_instr_irh	*irh;
694 	uint32_t		len = 0;
695 
696 	if (LIO_CN23XX_PF(oct)) {
697 		ih3 = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
698 		if (ih3->dlengsz) {
699 			KASSERT(sc->dmadptr, ("%s, %d, sc->dmadptr is NULL",
700 					      __func__, __LINE__));
701 			sc->cmd.cmd3.dptr = sc->dmadptr;
702 		}
703 
704 		irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
705 		if (irh->rflag) {
706 			KASSERT(sc->dmarptr, ("%s, %d, sc->dmarptr is NULL",
707 					      __func__, __LINE__));
708 			KASSERT(sc->status_word, ("%s, %d, sc->status_word is NULL",
709 						  __func__, __LINE__));
710 			*sc->status_word = COMPLETION_WORD_INIT;
711 			sc->cmd.cmd3.rptr = sc->dmarptr;
712 		}
713 		len = (uint32_t)ih3->dlengsz;
714 	}
715 	if (sc->wait_time)
716 		sc->timeout = ticks + lio_ms_to_ticks(sc->wait_time);
717 
718 	return (lio_send_command(oct, sc->iq_no, 1, &sc->cmd, sc,
719 				 len, LIO_REQTYPE_SOFT_COMMAND));
720 }
721 
722 int
723 lio_setup_sc_buffer_pool(struct octeon_device *oct)
724 {
725 	struct lio_soft_command	*sc;
726 	uint64_t		dma_addr;
727 	int			i;
728 
729 	STAILQ_INIT(&oct->sc_buf_pool.head);
730 	mtx_init(&oct->sc_buf_pool.lock, "sc_pool_lock", NULL, MTX_DEF);
731 	atomic_store_rel_int(&oct->sc_buf_pool.alloc_buf_count, 0);
732 
733 	for (i = 0; i < LIO_MAX_SOFT_COMMAND_BUFFERS; i++) {
734 		sc = (struct lio_soft_command *)
735 			lio_dma_alloc(LIO_SOFT_COMMAND_BUFFER_SIZE, (vm_paddr_t *)&dma_addr);
736 		if (sc == NULL) {
737 			lio_free_sc_buffer_pool(oct);
738 			return (1);
739 		}
740 
741 		sc->dma_addr = dma_addr;
742 		sc->size = LIO_SOFT_COMMAND_BUFFER_SIZE;
743 
744 		STAILQ_INSERT_TAIL(&oct->sc_buf_pool.head, &sc->node, entries);
745 	}
746 
747 	return (0);
748 }
749 
750 int
751 lio_free_sc_buffer_pool(struct octeon_device *oct)
752 {
753 	struct lio_stailq_node	*tmp, *tmp2;
754 	struct lio_soft_command	*sc;
755 
756 	mtx_lock(&oct->sc_buf_pool.lock);
757 
758 	STAILQ_FOREACH_SAFE(tmp, &oct->sc_buf_pool.head, entries, tmp2) {
759 		sc = LIO_STAILQ_FIRST_ENTRY(&oct->sc_buf_pool.head,
760 					    struct lio_soft_command, node);
761 
762 		STAILQ_REMOVE_HEAD(&oct->sc_buf_pool.head, entries);
763 
764 		lio_dma_free(sc->size, sc);
765 	}
766 
767 	STAILQ_INIT(&oct->sc_buf_pool.head);
768 
769 	mtx_unlock(&oct->sc_buf_pool.lock);
770 
771 	return (0);
772 }
773 
774 struct lio_soft_command *
775 lio_alloc_soft_command(struct octeon_device *oct, uint32_t datasize,
776 		       uint32_t rdatasize, uint32_t ctxsize)
777 {
778 	struct lio_soft_command	*sc = NULL;
779 	struct lio_stailq_node	*tmp;
780 	uint64_t		dma_addr;
781 	uint32_t		size;
782 	uint32_t		offset = sizeof(struct lio_soft_command);
783 
784 	KASSERT((offset + datasize + rdatasize + ctxsize) <=
785 		LIO_SOFT_COMMAND_BUFFER_SIZE,
786 		("%s, %d, offset + datasize + rdatasize + ctxsize > LIO_SOFT_COMMAND_BUFFER_SIZE",
787 		 __func__, __LINE__));
788 
789 	mtx_lock(&oct->sc_buf_pool.lock);
790 
791 	if (STAILQ_EMPTY(&oct->sc_buf_pool.head)) {
792 		mtx_unlock(&oct->sc_buf_pool.lock);
793 		return (NULL);
794 	}
795 	tmp = STAILQ_LAST(&oct->sc_buf_pool.head, lio_stailq_node, entries);
796 
797 	STAILQ_REMOVE(&oct->sc_buf_pool.head, tmp, lio_stailq_node, entries);
798 
799 	atomic_add_int(&oct->sc_buf_pool.alloc_buf_count, 1);
800 
801 	mtx_unlock(&oct->sc_buf_pool.lock);
802 
803 	sc = (struct lio_soft_command *)tmp;
804 
805 	dma_addr = sc->dma_addr;
806 	size = sc->size;
807 
808 	bzero(sc, sc->size);
809 
810 	sc->dma_addr = dma_addr;
811 	sc->size = size;
812 
813 	if (ctxsize) {
814 		sc->ctxptr = (uint8_t *)sc + offset;
815 		sc->ctxsize = ctxsize;
816 	}
817 
818 	/* Start data at 128 byte boundary */
819 	offset = (offset + ctxsize + 127) & 0xffffff80;
820 
821 	if (datasize) {
822 		sc->virtdptr = (uint8_t *)sc + offset;
823 		sc->dmadptr = dma_addr + offset;
824 		sc->datasize = datasize;
825 	}
826 	/* Start rdata at 128 byte boundary */
827 	offset = (offset + datasize + 127) & 0xffffff80;
828 
829 	if (rdatasize) {
830 		KASSERT(rdatasize >= 16, ("%s, %d, rdatasize < 16", __func__,
831 					  __LINE__));
832 		sc->virtrptr = (uint8_t *)sc + offset;
833 		sc->dmarptr = dma_addr + offset;
834 		sc->rdatasize = rdatasize;
835 		sc->status_word = (uint64_t *)((uint8_t *)(sc->virtrptr) +
836 					       rdatasize - 8);
837 	}
838 	return (sc);
839 }
840 
841 void
842 lio_free_soft_command(struct octeon_device *oct,
843 		      struct lio_soft_command *sc)
844 {
845 
846 	mtx_lock(&oct->sc_buf_pool.lock);
847 
848 	STAILQ_INSERT_TAIL(&oct->sc_buf_pool.head, &sc->node, entries);
849 
850 	atomic_subtract_int(&oct->sc_buf_pool.alloc_buf_count, 1);
851 
852 	mtx_unlock(&oct->sc_buf_pool.lock);
853 }
854