Linux Audio

Check our new training course

Linux BSP development engineering services

Need help to port Linux and bootloaders to your hardware?
Loading...
v4.6
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 *
 22 */
 23
 24#include <linux/slab.h>
 25#include <linux/mutex.h>
 26#include "kfd_device_queue_manager.h"
 27#include "kfd_kernel_queue.h"
 28#include "kfd_priv.h"
 29#include "kfd_pm4_headers.h"
 30#include "kfd_pm4_headers_vi.h"
 31#include "kfd_pm4_opcodes.h"
 32
 33static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
 34				unsigned int buffer_size_bytes)
 35{
 36	unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
 37
 38	BUG_ON((temp * sizeof(uint32_t)) > buffer_size_bytes);
 
 39	*wptr = temp;
 40}
 41
 42static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size)
 43{
 44	union PM4_MES_TYPE_3_HEADER header;
 45
 46	header.u32all = 0;
 47	header.opcode = opcode;
 48	header.count = packet_size/sizeof(uint32_t) - 2;
 49	header.type = PM4_TYPE_3;
 50
 51	return header.u32all;
 52}
 53
 54static void pm_calc_rlib_size(struct packet_manager *pm,
 55				unsigned int *rlib_size,
 56				bool *over_subscription)
 57{
 58	unsigned int process_count, queue_count;
 59	unsigned int map_queue_size;
 60
 61	BUG_ON(!pm || !rlib_size || !over_subscription);
 62
 63	process_count = pm->dqm->processes_count;
 64	queue_count = pm->dqm->queue_count;
 
 65
 66	/* check if there is over subscription*/
 
 
 
 
 67	*over_subscription = false;
 68	if ((process_count > 1) ||
 69		queue_count > PIPE_PER_ME_CP_SCHEDULING * QUEUES_PER_PIPE) {
 
 
 
 
 70		*over_subscription = true;
 71		pr_debug("kfd: over subscribed runlist\n");
 72	}
 73
 74	map_queue_size =
 75		(pm->dqm->dev->device_info->asic_family == CHIP_CARRIZO) ?
 76		sizeof(struct pm4_mes_map_queues) :
 77		sizeof(struct pm4_map_queues);
 78	/* calculate run list ib allocation size */
 79	*rlib_size = process_count * sizeof(struct pm4_map_process) +
 80		     queue_count * map_queue_size;
 81
 82	/*
 83	 * Increase the allocation size in case we need a chained run list
 84	 * when over subscription
 85	 */
 86	if (*over_subscription)
 87		*rlib_size += sizeof(struct pm4_runlist);
 88
 89	pr_debug("kfd: runlist ib size %d\n", *rlib_size);
 90}
 91
 92static int pm_allocate_runlist_ib(struct packet_manager *pm,
 93				unsigned int **rl_buffer,
 94				uint64_t *rl_gpu_buffer,
 95				unsigned int *rl_buffer_size,
 96				bool *is_over_subscription)
 97{
 98	int retval;
 99
100	BUG_ON(!pm);
101	BUG_ON(pm->allocated == true);
102	BUG_ON(is_over_subscription == NULL);
103
104	pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
105
106	retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
107					&pm->ib_buffer_obj);
108
109	if (retval != 0) {
110		pr_err("kfd: failed to allocate runlist IB\n");
111		return retval;
112	}
113
114	*(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
115	*rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
116
117	memset(*rl_buffer, 0, *rl_buffer_size);
118	pm->allocated = true;
119	return retval;
120}
121
122static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer,
123			uint64_t ib, size_t ib_size_in_dwords, bool chain)
124{
125	struct pm4_runlist *packet;
126
127	BUG_ON(!pm || !buffer || !ib);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
128
129	packet = (struct pm4_runlist *)buffer;
130
131	memset(buffer, 0, sizeof(struct pm4_runlist));
132	packet->header.u32all = build_pm4_header(IT_RUN_LIST,
133						sizeof(struct pm4_runlist));
134
135	packet->bitfields4.ib_size = ib_size_in_dwords;
136	packet->bitfields4.chain = chain ? 1 : 0;
137	packet->bitfields4.offload_polling = 0;
138	packet->bitfields4.valid = 1;
 
139	packet->ordinal2 = lower_32_bits(ib);
140	packet->bitfields3.ib_base_hi = upper_32_bits(ib);
141
142	return 0;
143}
144
145static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer,
146				struct qcm_process_device *qpd)
147{
148	struct pm4_map_process *packet;
149	struct queue *cur;
150	uint32_t num_queues;
151
152	BUG_ON(!pm || !buffer || !qpd);
153
154	packet = (struct pm4_map_process *)buffer;
155
156	pr_debug("kfd: In func %s\n", __func__);
157
158	memset(buffer, 0, sizeof(struct pm4_map_process));
159
160	packet->header.u32all = build_pm4_header(IT_MAP_PROCESS,
161					sizeof(struct pm4_map_process));
162	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
163	packet->bitfields2.process_quantum = 1;
164	packet->bitfields2.pasid = qpd->pqm->process->pasid;
165	packet->bitfields3.page_table_base = qpd->page_table_base;
166	packet->bitfields10.gds_size = qpd->gds_size;
167	packet->bitfields10.num_gws = qpd->num_gws;
168	packet->bitfields10.num_oac = qpd->num_oac;
169	num_queues = 0;
170	list_for_each_entry(cur, &qpd->queues_list, list)
171		num_queues++;
172	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
173
174	packet->sh_mem_config = qpd->sh_mem_config;
175	packet->sh_mem_bases = qpd->sh_mem_bases;
176	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
177	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
178
 
 
179	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
180	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
181
182	return 0;
183}
184
185static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer,
186		struct queue *q, bool is_static)
187{
188	struct pm4_mes_map_queues *packet;
189	bool use_static = is_static;
190
191	BUG_ON(!pm || !buffer || !q);
192
193	pr_debug("kfd: In func %s\n", __func__);
194
195	packet = (struct pm4_mes_map_queues *)buffer;
196	memset(buffer, 0, sizeof(struct pm4_map_queues));
197
198	packet->header.u32all = build_pm4_header(IT_MAP_QUEUES,
199						sizeof(struct pm4_map_queues));
200	packet->bitfields2.alloc_format =
201		alloc_format__mes_map_queues__one_per_pipe_vi;
202	packet->bitfields2.num_queues = 1;
203	packet->bitfields2.queue_sel =
204		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
205
206	packet->bitfields2.engine_sel =
207		engine_sel__mes_map_queues__compute_vi;
208	packet->bitfields2.queue_type =
209		queue_type__mes_map_queues__normal_compute_vi;
210
211	switch (q->properties.type) {
212	case KFD_QUEUE_TYPE_COMPUTE:
213		if (use_static)
214			packet->bitfields2.queue_type =
215		queue_type__mes_map_queues__normal_latency_static_queue_vi;
216		break;
217	case KFD_QUEUE_TYPE_DIQ:
218		packet->bitfields2.queue_type =
219			queue_type__mes_map_queues__debug_interface_queue_vi;
220		break;
221	case KFD_QUEUE_TYPE_SDMA:
222		packet->bitfields2.engine_sel =
223				engine_sel__mes_map_queues__sdma0_vi;
224		use_static = false; /* no static queues under SDMA */
225		break;
226	default:
227		pr_err("kfd: in %s queue type %d\n", __func__,
228				q->properties.type);
229		BUG();
230		break;
231	}
232	packet->bitfields3.doorbell_offset =
233			q->properties.doorbell_off;
234
235	packet->mqd_addr_lo =
236			lower_32_bits(q->gart_mqd_addr);
237
238	packet->mqd_addr_hi =
239			upper_32_bits(q->gart_mqd_addr);
240
241	packet->wptr_addr_lo =
242			lower_32_bits((uint64_t)q->properties.write_ptr);
243
244	packet->wptr_addr_hi =
245			upper_32_bits((uint64_t)q->properties.write_ptr);
246
247	return 0;
248}
249
250static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer,
251				struct queue *q, bool is_static)
252{
253	struct pm4_map_queues *packet;
254	bool use_static = is_static;
255
256	BUG_ON(!pm || !buffer || !q);
257
258	pr_debug("kfd: In func %s\n", __func__);
259
260	packet = (struct pm4_map_queues *)buffer;
261	memset(buffer, 0, sizeof(struct pm4_map_queues));
262
263	packet->header.u32all = build_pm4_header(IT_MAP_QUEUES,
264						sizeof(struct pm4_map_queues));
265	packet->bitfields2.alloc_format =
266				alloc_format__mes_map_queues__one_per_pipe;
267	packet->bitfields2.num_queues = 1;
268	packet->bitfields2.queue_sel =
269		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots;
270
271	packet->bitfields2.vidmem = (q->properties.is_interop) ?
272			vidmem__mes_map_queues__uses_video_memory :
273			vidmem__mes_map_queues__uses_no_video_memory;
274
275	switch (q->properties.type) {
276	case KFD_QUEUE_TYPE_COMPUTE:
277	case KFD_QUEUE_TYPE_DIQ:
278		packet->bitfields2.engine_sel =
279				engine_sel__mes_map_queues__compute;
280		break;
281	case KFD_QUEUE_TYPE_SDMA:
282		packet->bitfields2.engine_sel =
283				engine_sel__mes_map_queues__sdma0;
284		use_static = false; /* no static queues under SDMA */
285		break;
286	default:
287		BUG();
288		break;
289	}
290
291	packet->mes_map_queues_ordinals[0].bitfields3.doorbell_offset =
292			q->properties.doorbell_off;
293
294	packet->mes_map_queues_ordinals[0].bitfields3.is_static =
295			(use_static == true) ? 1 : 0;
296
297	packet->mes_map_queues_ordinals[0].mqd_addr_lo =
298			lower_32_bits(q->gart_mqd_addr);
299
300	packet->mes_map_queues_ordinals[0].mqd_addr_hi =
301			upper_32_bits(q->gart_mqd_addr);
302
303	packet->mes_map_queues_ordinals[0].wptr_addr_lo =
304			lower_32_bits((uint64_t)q->properties.write_ptr);
305
306	packet->mes_map_queues_ordinals[0].wptr_addr_hi =
307			upper_32_bits((uint64_t)q->properties.write_ptr);
308
309	return 0;
310}
311
312static int pm_create_runlist_ib(struct packet_manager *pm,
313				struct list_head *queues,
314				uint64_t *rl_gpu_addr,
315				size_t *rl_size_bytes)
316{
317	unsigned int alloc_size_bytes;
318	unsigned int *rl_buffer, rl_wptr, i;
319	int retval, proccesses_mapped;
320	struct device_process_node *cur;
321	struct qcm_process_device *qpd;
322	struct queue *q;
323	struct kernel_queue *kq;
324	bool is_over_subscription;
325
326	BUG_ON(!pm || !queues || !rl_size_bytes || !rl_gpu_addr);
327
328	rl_wptr = retval = proccesses_mapped = 0;
329
330	retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
331				&alloc_size_bytes, &is_over_subscription);
332	if (retval != 0)
333		return retval;
334
335	*rl_size_bytes = alloc_size_bytes;
 
336
337	pr_debug("kfd: In func %s\n", __func__);
338	pr_debug("kfd: building runlist ib process count: %d queues count %d\n",
339		pm->dqm->processes_count, pm->dqm->queue_count);
340
341	/* build the run list ib packet */
342	list_for_each_entry(cur, queues, list) {
343		qpd = cur->qpd;
344		/* build map process packet */
345		if (proccesses_mapped >= pm->dqm->processes_count) {
346			pr_debug("kfd: not enough space left in runlist IB\n");
347			pm_release_ib(pm);
348			return -ENOMEM;
349		}
350
351		retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd);
352		if (retval != 0)
353			return retval;
354
355		proccesses_mapped++;
356		inc_wptr(&rl_wptr, sizeof(struct pm4_map_process),
357				alloc_size_bytes);
358
359		list_for_each_entry(kq, &qpd->priv_queue_list, list) {
360			if (kq->queue->properties.is_active != true)
361				continue;
362
363			pr_debug("kfd: static_queue, mapping kernel q %d, is debug status %d\n",
364				kq->queue->queue, qpd->is_debug);
365
366			if (pm->dqm->dev->device_info->asic_family ==
367					CHIP_CARRIZO)
368				retval = pm_create_map_queue_vi(pm,
369						&rl_buffer[rl_wptr],
370						kq->queue,
371						qpd->is_debug);
372			else
373				retval = pm_create_map_queue(pm,
374						&rl_buffer[rl_wptr],
375						kq->queue,
376						qpd->is_debug);
377			if (retval != 0)
378				return retval;
379
380			inc_wptr(&rl_wptr,
381				sizeof(struct pm4_map_queues),
382				alloc_size_bytes);
383		}
384
385		list_for_each_entry(q, &qpd->queues_list, list) {
386			if (q->properties.is_active != true)
387				continue;
388
389			pr_debug("kfd: static_queue, mapping user queue %d, is debug status %d\n",
390				q->queue, qpd->is_debug);
391
392			if (pm->dqm->dev->device_info->asic_family ==
393					CHIP_CARRIZO)
394				retval = pm_create_map_queue_vi(pm,
395						&rl_buffer[rl_wptr],
396						q,
397						qpd->is_debug);
398			else
399				retval = pm_create_map_queue(pm,
400						&rl_buffer[rl_wptr],
401						q,
402						qpd->is_debug);
403
404			if (retval != 0)
405				return retval;
406
407			inc_wptr(&rl_wptr,
408				sizeof(struct pm4_map_queues),
409				alloc_size_bytes);
410		}
411	}
412
413	pr_debug("kfd: finished map process and queues to runlist\n");
414
415	if (is_over_subscription)
416		pm_create_runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr,
417				alloc_size_bytes / sizeof(uint32_t), true);
 
 
418
419	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
420		pr_debug("0x%2X ", rl_buffer[i]);
421	pr_debug("\n");
422
423	return 0;
424}
425
426int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
 
 
 
 
 
 
427{
428	BUG_ON(!dqm);
 
 
 
 
 
429
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
430	pm->dqm = dqm;
431	mutex_init(&pm->lock);
432	pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
433	if (pm->priv_queue == NULL) {
434		mutex_destroy(&pm->lock);
435		return -ENOMEM;
436	}
437	pm->allocated = false;
438
439	return 0;
440}
441
442void pm_uninit(struct packet_manager *pm)
443{
444	BUG_ON(!pm);
445
446	mutex_destroy(&pm->lock);
447	kernel_queue_uninit(pm->priv_queue);
448}
449
450int pm_send_set_resources(struct packet_manager *pm,
451				struct scheduling_resources *res)
452{
453	struct pm4_set_resources *packet;
454
455	BUG_ON(!pm || !res);
456
457	pr_debug("kfd: In func %s\n", __func__);
458
459	mutex_lock(&pm->lock);
460	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
461					sizeof(*packet) / sizeof(uint32_t),
462			(unsigned int **)&packet);
463	if (packet == NULL) {
464		mutex_unlock(&pm->lock);
465		pr_err("kfd: failed to allocate buffer on kernel queue\n");
466		return -ENOMEM;
467	}
468
469	memset(packet, 0, sizeof(struct pm4_set_resources));
470	packet->header.u32all = build_pm4_header(IT_SET_RESOURCES,
471					sizeof(struct pm4_set_resources));
472
473	packet->bitfields2.queue_type =
474			queue_type__mes_set_resources__hsa_interface_queue_hiq;
475	packet->bitfields2.vmid_mask = res->vmid_mask;
476	packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY;
477	packet->bitfields7.oac_mask = res->oac_mask;
478	packet->bitfields8.gds_heap_base = res->gds_heap_base;
479	packet->bitfields8.gds_heap_size = res->gds_heap_size;
480
481	packet->gws_mask_lo = lower_32_bits(res->gws_mask);
482	packet->gws_mask_hi = upper_32_bits(res->gws_mask);
483
484	packet->queue_mask_lo = lower_32_bits(res->queue_mask);
485	packet->queue_mask_hi = upper_32_bits(res->queue_mask);
486
487	pm->priv_queue->ops.submit_packet(pm->priv_queue);
488
 
489	mutex_unlock(&pm->lock);
490
491	return 0;
492}
493
494int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
495{
496	uint64_t rl_gpu_ib_addr;
497	uint32_t *rl_buffer;
498	size_t rl_ib_size, packet_size_dwords;
499	int retval;
500
501	BUG_ON(!pm || !dqm_queues);
502
503	retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
504					&rl_ib_size);
505	if (retval != 0)
506		goto fail_create_runlist_ib;
507
508	pr_debug("kfd: runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
509
510	packet_size_dwords = sizeof(struct pm4_runlist) / sizeof(uint32_t);
511	mutex_lock(&pm->lock);
512
513	retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
514					packet_size_dwords, &rl_buffer);
515	if (retval != 0)
516		goto fail_acquire_packet_buffer;
517
518	retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr,
519					rl_ib_size / sizeof(uint32_t), false);
520	if (retval != 0)
521		goto fail_create_runlist;
522
523	pm->priv_queue->ops.submit_packet(pm->priv_queue);
524
525	mutex_unlock(&pm->lock);
526
527	return retval;
528
529fail_create_runlist:
530	pm->priv_queue->ops.rollback_packet(pm->priv_queue);
531fail_acquire_packet_buffer:
532	mutex_unlock(&pm->lock);
533fail_create_runlist_ib:
534	if (pm->allocated == true)
535		pm_release_ib(pm);
536	return retval;
537}
538
539int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
540			uint32_t fence_value)
541{
542	int retval;
543	struct pm4_query_status *packet;
544
545	BUG_ON(!pm || !fence_address);
 
546
547	mutex_lock(&pm->lock);
548	retval = pm->priv_queue->ops.acquire_packet_buffer(
549			pm->priv_queue,
550			sizeof(struct pm4_query_status) / sizeof(uint32_t),
551			(unsigned int **)&packet);
552	if (retval != 0)
553		goto fail_acquire_packet_buffer;
554
555	packet->header.u32all = build_pm4_header(IT_QUERY_STATUS,
556					sizeof(struct pm4_query_status));
557
558	packet->bitfields2.context_id = 0;
559	packet->bitfields2.interrupt_sel =
560			interrupt_sel__mes_query_status__completion_status;
561	packet->bitfields2.command =
562			command__mes_query_status__fence_only_after_write_ack;
563
564	packet->addr_hi = upper_32_bits((uint64_t)fence_address);
565	packet->addr_lo = lower_32_bits((uint64_t)fence_address);
566	packet->data_hi = upper_32_bits((uint64_t)fence_value);
567	packet->data_lo = lower_32_bits((uint64_t)fence_value);
568
569	pm->priv_queue->ops.submit_packet(pm->priv_queue);
570	mutex_unlock(&pm->lock);
571
572	return 0;
573
574fail_acquire_packet_buffer:
575	mutex_unlock(&pm->lock);
576	return retval;
577}
578
579int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
580			enum kfd_preempt_type_filter mode,
581			uint32_t filter_param, bool reset,
582			unsigned int sdma_engine)
583{
584	int retval;
585	uint32_t *buffer;
586	struct pm4_unmap_queues *packet;
587
588	BUG_ON(!pm);
589
590	mutex_lock(&pm->lock);
591	retval = pm->priv_queue->ops.acquire_packet_buffer(
592			pm->priv_queue,
593			sizeof(struct pm4_unmap_queues) / sizeof(uint32_t),
594			&buffer);
595	if (retval != 0)
596		goto err_acquire_packet_buffer;
597
598	packet = (struct pm4_unmap_queues *)buffer;
599	memset(buffer, 0, sizeof(struct pm4_unmap_queues));
600	pr_debug("kfd: static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n",
601		mode, reset, type);
602	packet->header.u32all = build_pm4_header(IT_UNMAP_QUEUES,
603					sizeof(struct pm4_unmap_queues));
604	switch (type) {
605	case KFD_QUEUE_TYPE_COMPUTE:
606	case KFD_QUEUE_TYPE_DIQ:
607		packet->bitfields2.engine_sel =
608			engine_sel__mes_unmap_queues__compute;
609		break;
610	case KFD_QUEUE_TYPE_SDMA:
611		packet->bitfields2.engine_sel =
612			engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
613		break;
614	default:
615		BUG();
616		break;
 
617	}
618
619	if (reset)
620		packet->bitfields2.action =
621				action__mes_unmap_queues__reset_queues;
622	else
623		packet->bitfields2.action =
624				action__mes_unmap_queues__preempt_queues;
625
626	switch (mode) {
627	case KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE:
628		packet->bitfields2.queue_sel =
629				queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
630		packet->bitfields2.num_queues = 1;
631		packet->bitfields3b.doorbell_offset0 = filter_param;
632		break;
633	case KFD_PREEMPT_TYPE_FILTER_BY_PASID:
634		packet->bitfields2.queue_sel =
635				queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
636		packet->bitfields3a.pasid = filter_param;
637		break;
638	case KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES:
639		packet->bitfields2.queue_sel =
640				queue_sel__mes_unmap_queues__perform_request_on_all_active_queues;
641		break;
642	case KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES:
643		/* in this case, we do not preempt static queues */
644		packet->bitfields2.queue_sel =
645				queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only;
646		break;
647	default:
648		BUG();
649		break;
650	};
 
651
652	pm->priv_queue->ops.submit_packet(pm->priv_queue);
653
654	mutex_unlock(&pm->lock);
655	return 0;
656
 
 
657err_acquire_packet_buffer:
658	mutex_unlock(&pm->lock);
659	return retval;
660}
661
662void pm_release_ib(struct packet_manager *pm)
663{
664	BUG_ON(!pm);
665
666	mutex_lock(&pm->lock);
667	if (pm->allocated) {
668		kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
669		pm->allocated = false;
670	}
671	mutex_unlock(&pm->lock);
672}
v4.17
  1/*
  2 * Copyright 2014 Advanced Micro Devices, Inc.
  3 *
  4 * Permission is hereby granted, free of charge, to any person obtaining a
  5 * copy of this software and associated documentation files (the "Software"),
  6 * to deal in the Software without restriction, including without limitation
  7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  8 * and/or sell copies of the Software, and to permit persons to whom the
  9 * Software is furnished to do so, subject to the following conditions:
 10 *
 11 * The above copyright notice and this permission notice shall be included in
 12 * all copies or substantial portions of the Software.
 13 *
 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 20 * OTHER DEALINGS IN THE SOFTWARE.
 21 *
 22 */
 23
 24#include <linux/slab.h>
 25#include <linux/mutex.h>
 26#include "kfd_device_queue_manager.h"
 27#include "kfd_kernel_queue.h"
 28#include "kfd_priv.h"
 
 29#include "kfd_pm4_headers_vi.h"
 30#include "kfd_pm4_opcodes.h"
 31
 32static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
 33				unsigned int buffer_size_bytes)
 34{
 35	unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
 36
 37	WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
 38	     "Runlist IB overflow");
 39	*wptr = temp;
 40}
 41
 42static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size)
 43{
 44	union PM4_MES_TYPE_3_HEADER header;
 45
 46	header.u32All = 0;
 47	header.opcode = opcode;
 48	header.count = packet_size / 4 - 2;
 49	header.type = PM4_TYPE_3;
 50
 51	return header.u32All;
 52}
 53
 54static void pm_calc_rlib_size(struct packet_manager *pm,
 55				unsigned int *rlib_size,
 56				bool *over_subscription)
 57{
 58	unsigned int process_count, queue_count, compute_queue_count;
 59	unsigned int map_queue_size;
 60	unsigned int max_proc_per_quantum = 1;
 61	struct kfd_dev *dev = pm->dqm->dev;
 62
 63	process_count = pm->dqm->processes_count;
 64	queue_count = pm->dqm->queue_count;
 65	compute_queue_count = queue_count - pm->dqm->sdma_queue_count;
 66
 67	/* check if there is over subscription
 68	 * Note: the arbitration between the number of VMIDs and
 69	 * hws_max_conc_proc has been done in
 70	 * kgd2kfd_device_init().
 71	 */
 72	*over_subscription = false;
 73
 74	if (dev->max_proc_per_quantum > 1)
 75		max_proc_per_quantum = dev->max_proc_per_quantum;
 76
 77	if ((process_count > max_proc_per_quantum) ||
 78	    compute_queue_count > get_queues_num(pm->dqm)) {
 79		*over_subscription = true;
 80		pr_debug("Over subscribed runlist\n");
 81	}
 82
 83	map_queue_size = sizeof(struct pm4_mes_map_queues);
 
 
 
 84	/* calculate run list ib allocation size */
 85	*rlib_size = process_count * sizeof(struct pm4_mes_map_process) +
 86		     queue_count * map_queue_size;
 87
 88	/*
 89	 * Increase the allocation size in case we need a chained run list
 90	 * when over subscription
 91	 */
 92	if (*over_subscription)
 93		*rlib_size += sizeof(struct pm4_mes_runlist);
 94
 95	pr_debug("runlist ib size %d\n", *rlib_size);
 96}
 97
 98static int pm_allocate_runlist_ib(struct packet_manager *pm,
 99				unsigned int **rl_buffer,
100				uint64_t *rl_gpu_buffer,
101				unsigned int *rl_buffer_size,
102				bool *is_over_subscription)
103{
104	int retval;
105
106	if (WARN_ON(pm->allocated))
107		return -EINVAL;
 
108
109	pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
110
111	retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
112					&pm->ib_buffer_obj);
113
114	if (retval) {
115		pr_err("Failed to allocate runlist IB\n");
116		return retval;
117	}
118
119	*(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
120	*rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
121
122	memset(*rl_buffer, 0, *rl_buffer_size);
123	pm->allocated = true;
124	return retval;
125}
126
127static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer,
128			uint64_t ib, size_t ib_size_in_dwords, bool chain)
129{
130	struct pm4_mes_runlist *packet;
131	int concurrent_proc_cnt = 0;
132	struct kfd_dev *kfd = pm->dqm->dev;
133
134	if (WARN_ON(!ib))
135		return -EFAULT;
136
137	/* Determine the number of processes to map together to HW:
138	 * it can not exceed the number of VMIDs available to the
139	 * scheduler, and it is determined by the smaller of the number
140	 * of processes in the runlist and kfd module parameter
141	 * hws_max_conc_proc.
142	 * Note: the arbitration between the number of VMIDs and
143	 * hws_max_conc_proc has been done in
144	 * kgd2kfd_device_init().
145	 */
146	concurrent_proc_cnt = min(pm->dqm->processes_count,
147			kfd->max_proc_per_quantum);
148
149	packet = (struct pm4_mes_runlist *)buffer;
150
151	memset(buffer, 0, sizeof(struct pm4_mes_runlist));
152	packet->header.u32All = build_pm4_header(IT_RUN_LIST,
153						sizeof(struct pm4_mes_runlist));
154
155	packet->bitfields4.ib_size = ib_size_in_dwords;
156	packet->bitfields4.chain = chain ? 1 : 0;
157	packet->bitfields4.offload_polling = 0;
158	packet->bitfields4.valid = 1;
159	packet->bitfields4.process_cnt = concurrent_proc_cnt;
160	packet->ordinal2 = lower_32_bits(ib);
161	packet->bitfields3.ib_base_hi = upper_32_bits(ib);
162
163	return 0;
164}
165
166static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer,
167				struct qcm_process_device *qpd)
168{
169	struct pm4_mes_map_process *packet;
 
 
 
 
170
171	packet = (struct pm4_mes_map_process *)buffer;
172
173	memset(buffer, 0, sizeof(struct pm4_mes_map_process));
174
175	packet->header.u32All = build_pm4_header(IT_MAP_PROCESS,
176					sizeof(struct pm4_mes_map_process));
 
 
177	packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
178	packet->bitfields2.process_quantum = 1;
179	packet->bitfields2.pasid = qpd->pqm->process->pasid;
180	packet->bitfields3.page_table_base = qpd->page_table_base;
181	packet->bitfields10.gds_size = qpd->gds_size;
182	packet->bitfields10.num_gws = qpd->num_gws;
183	packet->bitfields10.num_oac = qpd->num_oac;
184	packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count;
 
 
 
185
186	packet->sh_mem_config = qpd->sh_mem_config;
187	packet->sh_mem_bases = qpd->sh_mem_bases;
188	packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
189	packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
190
191	packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
192
193	packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
194	packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
195
196	return 0;
197}
198
199static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer,
200		struct queue *q, bool is_static)
201{
202	struct pm4_mes_map_queues *packet;
203	bool use_static = is_static;
204
 
 
 
 
205	packet = (struct pm4_mes_map_queues *)buffer;
206	memset(buffer, 0, sizeof(struct pm4_mes_map_queues));
207
208	packet->header.u32All = build_pm4_header(IT_MAP_QUEUES,
209						sizeof(struct pm4_mes_map_queues));
210	packet->bitfields2.alloc_format =
211		alloc_format__mes_map_queues__one_per_pipe_vi;
212	packet->bitfields2.num_queues = 1;
213	packet->bitfields2.queue_sel =
214		queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
215
216	packet->bitfields2.engine_sel =
217		engine_sel__mes_map_queues__compute_vi;
218	packet->bitfields2.queue_type =
219		queue_type__mes_map_queues__normal_compute_vi;
220
221	switch (q->properties.type) {
222	case KFD_QUEUE_TYPE_COMPUTE:
223		if (use_static)
224			packet->bitfields2.queue_type =
225		queue_type__mes_map_queues__normal_latency_static_queue_vi;
226		break;
227	case KFD_QUEUE_TYPE_DIQ:
228		packet->bitfields2.queue_type =
229			queue_type__mes_map_queues__debug_interface_queue_vi;
230		break;
231	case KFD_QUEUE_TYPE_SDMA:
232		packet->bitfields2.engine_sel = q->properties.sdma_engine_id +
233				engine_sel__mes_map_queues__sdma0_vi;
234		use_static = false; /* no static queues under SDMA */
235		break;
236	default:
237		WARN(1, "queue type %d", q->properties.type);
238		return -EINVAL;
 
 
239	}
240	packet->bitfields3.doorbell_offset =
241			q->properties.doorbell_off;
242
243	packet->mqd_addr_lo =
244			lower_32_bits(q->gart_mqd_addr);
245
246	packet->mqd_addr_hi =
247			upper_32_bits(q->gart_mqd_addr);
248
249	packet->wptr_addr_lo =
250			lower_32_bits((uint64_t)q->properties.write_ptr);
251
252	packet->wptr_addr_hi =
253			upper_32_bits((uint64_t)q->properties.write_ptr);
254
255	return 0;
256}
257
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
258static int pm_create_runlist_ib(struct packet_manager *pm,
259				struct list_head *queues,
260				uint64_t *rl_gpu_addr,
261				size_t *rl_size_bytes)
262{
263	unsigned int alloc_size_bytes;
264	unsigned int *rl_buffer, rl_wptr, i;
265	int retval, proccesses_mapped;
266	struct device_process_node *cur;
267	struct qcm_process_device *qpd;
268	struct queue *q;
269	struct kernel_queue *kq;
270	bool is_over_subscription;
271
 
 
272	rl_wptr = retval = proccesses_mapped = 0;
273
274	retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
275				&alloc_size_bytes, &is_over_subscription);
276	if (retval)
277		return retval;
278
279	*rl_size_bytes = alloc_size_bytes;
280	pm->ib_size_bytes = alloc_size_bytes;
281
282	pr_debug("Building runlist ib process count: %d queues count %d\n",
 
283		pm->dqm->processes_count, pm->dqm->queue_count);
284
285	/* build the run list ib packet */
286	list_for_each_entry(cur, queues, list) {
287		qpd = cur->qpd;
288		/* build map process packet */
289		if (proccesses_mapped >= pm->dqm->processes_count) {
290			pr_debug("Not enough space left in runlist IB\n");
291			pm_release_ib(pm);
292			return -ENOMEM;
293		}
294
295		retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd);
296		if (retval)
297			return retval;
298
299		proccesses_mapped++;
300		inc_wptr(&rl_wptr, sizeof(struct pm4_mes_map_process),
301				alloc_size_bytes);
302
303		list_for_each_entry(kq, &qpd->priv_queue_list, list) {
304			if (!kq->queue->properties.is_active)
305				continue;
306
307			pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
308				kq->queue->queue, qpd->is_debug);
309
310			retval = pm_create_map_queue(pm,
 
 
311						&rl_buffer[rl_wptr],
312						kq->queue,
313						qpd->is_debug);
314			if (retval)
 
 
 
 
 
315				return retval;
316
317			inc_wptr(&rl_wptr,
318				sizeof(struct pm4_mes_map_queues),
319				alloc_size_bytes);
320		}
321
322		list_for_each_entry(q, &qpd->queues_list, list) {
323			if (!q->properties.is_active)
324				continue;
325
326			pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
327				q->queue, qpd->is_debug);
328
329			retval = pm_create_map_queue(pm,
 
 
 
 
 
 
 
330						&rl_buffer[rl_wptr],
331						q,
332						qpd->is_debug);
333
334			if (retval)
335				return retval;
336
337			inc_wptr(&rl_wptr,
338				sizeof(struct pm4_mes_map_queues),
339				alloc_size_bytes);
340		}
341	}
342
343	pr_debug("Finished map process and queues to runlist\n");
344
345	if (is_over_subscription)
346		retval = pm_create_runlist(pm, &rl_buffer[rl_wptr],
347					*rl_gpu_addr,
348					alloc_size_bytes / sizeof(uint32_t),
349					true);
350
351	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
352		pr_debug("0x%2X ", rl_buffer[i]);
353	pr_debug("\n");
354
355	return retval;
356}
357
358/* pm_create_release_mem - Create a RELEASE_MEM packet and return the size
359 *     of this packet
360 *     @gpu_addr - GPU address of the packet. It's a virtual address.
361 *     @buffer - buffer to fill up with the packet. It's a CPU kernel pointer
362 *     Return - length of the packet
363 */
364uint32_t pm_create_release_mem(uint64_t gpu_addr, uint32_t *buffer)
365{
366	struct pm4_mec_release_mem *packet;
367
368	WARN_ON(!buffer);
369
370	packet = (struct pm4_mec_release_mem *)buffer;
371	memset(buffer, 0, sizeof(*packet));
372
373	packet->header.u32All = build_pm4_header(IT_RELEASE_MEM,
374						 sizeof(*packet));
375
376	packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
377	packet->bitfields2.event_index = event_index___release_mem__end_of_pipe;
378	packet->bitfields2.tcl1_action_ena = 1;
379	packet->bitfields2.tc_action_ena = 1;
380	packet->bitfields2.cache_policy = cache_policy___release_mem__lru;
381	packet->bitfields2.atc = 0;
382
383	packet->bitfields3.data_sel = data_sel___release_mem__send_32_bit_low;
384	packet->bitfields3.int_sel =
385		int_sel___release_mem__send_interrupt_after_write_confirm;
386
387	packet->bitfields4.address_lo_32b = (gpu_addr & 0xffffffff) >> 2;
388	packet->address_hi = upper_32_bits(gpu_addr);
389
390	packet->data_lo = 0;
391
392	return sizeof(*packet) / sizeof(unsigned int);
393}
394
395int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
396{
397	pm->dqm = dqm;
398	mutex_init(&pm->lock);
399	pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
400	if (!pm->priv_queue) {
401		mutex_destroy(&pm->lock);
402		return -ENOMEM;
403	}
404	pm->allocated = false;
405
406	return 0;
407}
408
409void pm_uninit(struct packet_manager *pm)
410{
 
 
411	mutex_destroy(&pm->lock);
412	kernel_queue_uninit(pm->priv_queue);
413}
414
415int pm_send_set_resources(struct packet_manager *pm,
416				struct scheduling_resources *res)
417{
418	struct pm4_mes_set_resources *packet;
419	int retval = 0;
 
 
 
420
421	mutex_lock(&pm->lock);
422	pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
423					sizeof(*packet) / sizeof(uint32_t),
424					(unsigned int **)&packet);
425	if (!packet) {
426		pr_err("Failed to allocate buffer on kernel queue\n");
427		retval = -ENOMEM;
428		goto out;
429	}
430
431	memset(packet, 0, sizeof(struct pm4_mes_set_resources));
432	packet->header.u32All = build_pm4_header(IT_SET_RESOURCES,
433					sizeof(struct pm4_mes_set_resources));
434
435	packet->bitfields2.queue_type =
436			queue_type__mes_set_resources__hsa_interface_queue_hiq;
437	packet->bitfields2.vmid_mask = res->vmid_mask;
438	packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY_MS / 100;
439	packet->bitfields7.oac_mask = res->oac_mask;
440	packet->bitfields8.gds_heap_base = res->gds_heap_base;
441	packet->bitfields8.gds_heap_size = res->gds_heap_size;
442
443	packet->gws_mask_lo = lower_32_bits(res->gws_mask);
444	packet->gws_mask_hi = upper_32_bits(res->gws_mask);
445
446	packet->queue_mask_lo = lower_32_bits(res->queue_mask);
447	packet->queue_mask_hi = upper_32_bits(res->queue_mask);
448
449	pm->priv_queue->ops.submit_packet(pm->priv_queue);
450
451out:
452	mutex_unlock(&pm->lock);
453
454	return retval;
455}
456
457int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
458{
459	uint64_t rl_gpu_ib_addr;
460	uint32_t *rl_buffer;
461	size_t rl_ib_size, packet_size_dwords;
462	int retval;
463
 
 
464	retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
465					&rl_ib_size);
466	if (retval)
467		goto fail_create_runlist_ib;
468
469	pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
470
471	packet_size_dwords = sizeof(struct pm4_mes_runlist) / sizeof(uint32_t);
472	mutex_lock(&pm->lock);
473
474	retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
475					packet_size_dwords, &rl_buffer);
476	if (retval)
477		goto fail_acquire_packet_buffer;
478
479	retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr,
480					rl_ib_size / sizeof(uint32_t), false);
481	if (retval)
482		goto fail_create_runlist;
483
484	pm->priv_queue->ops.submit_packet(pm->priv_queue);
485
486	mutex_unlock(&pm->lock);
487
488	return retval;
489
490fail_create_runlist:
491	pm->priv_queue->ops.rollback_packet(pm->priv_queue);
492fail_acquire_packet_buffer:
493	mutex_unlock(&pm->lock);
494fail_create_runlist_ib:
495	pm_release_ib(pm);
 
496	return retval;
497}
498
499int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
500			uint32_t fence_value)
501{
502	int retval;
503	struct pm4_mes_query_status *packet;
504
505	if (WARN_ON(!fence_address))
506		return -EFAULT;
507
508	mutex_lock(&pm->lock);
509	retval = pm->priv_queue->ops.acquire_packet_buffer(
510			pm->priv_queue,
511			sizeof(struct pm4_mes_query_status) / sizeof(uint32_t),
512			(unsigned int **)&packet);
513	if (retval)
514		goto fail_acquire_packet_buffer;
515
516	packet->header.u32All = build_pm4_header(IT_QUERY_STATUS,
517					sizeof(struct pm4_mes_query_status));
518
519	packet->bitfields2.context_id = 0;
520	packet->bitfields2.interrupt_sel =
521			interrupt_sel__mes_query_status__completion_status;
522	packet->bitfields2.command =
523			command__mes_query_status__fence_only_after_write_ack;
524
525	packet->addr_hi = upper_32_bits((uint64_t)fence_address);
526	packet->addr_lo = lower_32_bits((uint64_t)fence_address);
527	packet->data_hi = upper_32_bits((uint64_t)fence_value);
528	packet->data_lo = lower_32_bits((uint64_t)fence_value);
529
530	pm->priv_queue->ops.submit_packet(pm->priv_queue);
 
 
 
531
532fail_acquire_packet_buffer:
533	mutex_unlock(&pm->lock);
534	return retval;
535}
536
537int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
538			enum kfd_unmap_queues_filter filter,
539			uint32_t filter_param, bool reset,
540			unsigned int sdma_engine)
541{
542	int retval;
543	uint32_t *buffer;
544	struct pm4_mes_unmap_queues *packet;
 
 
545
546	mutex_lock(&pm->lock);
547	retval = pm->priv_queue->ops.acquire_packet_buffer(
548			pm->priv_queue,
549			sizeof(struct pm4_mes_unmap_queues) / sizeof(uint32_t),
550			&buffer);
551	if (retval)
552		goto err_acquire_packet_buffer;
553
554	packet = (struct pm4_mes_unmap_queues *)buffer;
555	memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues));
556	pr_debug("static_queue: unmapping queues: filter is %d , reset is %d , type is %d\n",
557		filter, reset, type);
558	packet->header.u32All = build_pm4_header(IT_UNMAP_QUEUES,
559					sizeof(struct pm4_mes_unmap_queues));
560	switch (type) {
561	case KFD_QUEUE_TYPE_COMPUTE:
562	case KFD_QUEUE_TYPE_DIQ:
563		packet->bitfields2.engine_sel =
564			engine_sel__mes_unmap_queues__compute;
565		break;
566	case KFD_QUEUE_TYPE_SDMA:
567		packet->bitfields2.engine_sel =
568			engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
569		break;
570	default:
571		WARN(1, "queue type %d", type);
572		retval = -EINVAL;
573		goto err_invalid;
574	}
575
576	if (reset)
577		packet->bitfields2.action =
578				action__mes_unmap_queues__reset_queues;
579	else
580		packet->bitfields2.action =
581				action__mes_unmap_queues__preempt_queues;
582
583	switch (filter) {
584	case KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE:
585		packet->bitfields2.queue_sel =
586				queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
587		packet->bitfields2.num_queues = 1;
588		packet->bitfields3b.doorbell_offset0 = filter_param;
589		break;
590	case KFD_UNMAP_QUEUES_FILTER_BY_PASID:
591		packet->bitfields2.queue_sel =
592				queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
593		packet->bitfields3a.pasid = filter_param;
594		break;
595	case KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES:
596		packet->bitfields2.queue_sel =
597				queue_sel__mes_unmap_queues__unmap_all_queues;
598		break;
599	case KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES:
600		/* in this case, we do not preempt static queues */
601		packet->bitfields2.queue_sel =
602				queue_sel__mes_unmap_queues__unmap_all_non_static_queues;
603		break;
604	default:
605		WARN(1, "filter %d", filter);
606		retval = -EINVAL;
607		goto err_invalid;
608	}
609
610	pm->priv_queue->ops.submit_packet(pm->priv_queue);
611
612	mutex_unlock(&pm->lock);
613	return 0;
614
615err_invalid:
616	pm->priv_queue->ops.rollback_packet(pm->priv_queue);
617err_acquire_packet_buffer:
618	mutex_unlock(&pm->lock);
619	return retval;
620}
621
622void pm_release_ib(struct packet_manager *pm)
623{
 
 
624	mutex_lock(&pm->lock);
625	if (pm->allocated) {
626		kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
627		pm->allocated = false;
628	}
629	mutex_unlock(&pm->lock);
630}
631
632#if defined(CONFIG_DEBUG_FS)
633
634int pm_debugfs_runlist(struct seq_file *m, void *data)
635{
636	struct packet_manager *pm = data;
637
638	mutex_lock(&pm->lock);
639
640	if (!pm->allocated) {
641		seq_puts(m, "  No active runlist\n");
642		goto out;
643	}
644
645	seq_hex_dump(m, "  ", DUMP_PREFIX_OFFSET, 32, 4,
646		     pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
647
648out:
649	mutex_unlock(&pm->lock);
650	return 0;
651}
652
653#endif