• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/ratelimit.h>
25 #include <linux/printk.h>
26 #include <linux/slab.h>
27 #include <linux/list.h>
28 #include <linux/types.h>
29 #include <linux/bitops.h>
30 #include <linux/sched.h>
31 #include "kfd_priv.h"
32 #include "kfd_device_queue_manager.h"
33 #include "kfd_mqd_manager.h"
34 #include "cik_regs.h"
35 #include "kfd_kernel_queue.h"
36 #include "amdgpu_amdkfd.h"
37 
38 /* Size of the per-pipe EOP queue */
39 #define CIK_HPD_EOP_BYTES_LOG2 11
40 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
41 
42 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
43 				  u32 pasid, unsigned int vmid);
44 
45 static int execute_queues_cpsch(struct device_queue_manager *dqm,
46 				enum kfd_unmap_queues_filter filter,
47 				uint32_t filter_param);
48 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
49 				enum kfd_unmap_queues_filter filter,
50 				uint32_t filter_param);
51 
52 static int map_queues_cpsch(struct device_queue_manager *dqm);
53 
54 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
55 				struct queue *q);
56 
57 static inline void deallocate_hqd(struct device_queue_manager *dqm,
58 				struct queue *q);
59 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
60 static int allocate_sdma_queue(struct device_queue_manager *dqm,
61 				struct queue *q);
62 static void kfd_process_hw_exception(struct work_struct *work);
63 
64 static inline
get_mqd_type_from_queue_type(enum kfd_queue_type type)65 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
66 {
67 	if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI)
68 		return KFD_MQD_TYPE_SDMA;
69 	return KFD_MQD_TYPE_CP;
70 }
71 
is_pipe_enabled(struct device_queue_manager * dqm,int mec,int pipe)72 static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
73 {
74 	int i;
75 	int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec
76 		+ pipe * dqm->dev->shared_resources.num_queue_per_pipe;
77 
78 	/* queue is available for KFD usage if bit is 1 */
79 	for (i = 0; i <  dqm->dev->shared_resources.num_queue_per_pipe; ++i)
80 		if (test_bit(pipe_offset + i,
81 			      dqm->dev->shared_resources.cp_queue_bitmap))
82 			return true;
83 	return false;
84 }
85 
get_cp_queues_num(struct device_queue_manager * dqm)86 unsigned int get_cp_queues_num(struct device_queue_manager *dqm)
87 {
88 	return bitmap_weight(dqm->dev->shared_resources.cp_queue_bitmap,
89 				KGD_MAX_QUEUES);
90 }
91 
get_queues_per_pipe(struct device_queue_manager * dqm)92 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
93 {
94 	return dqm->dev->shared_resources.num_queue_per_pipe;
95 }
96 
get_pipes_per_mec(struct device_queue_manager * dqm)97 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
98 {
99 	return dqm->dev->shared_resources.num_pipe_per_mec;
100 }
101 
get_num_sdma_engines(struct device_queue_manager * dqm)102 static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm)
103 {
104 	return dqm->dev->device_info->num_sdma_engines;
105 }
106 
get_num_xgmi_sdma_engines(struct device_queue_manager * dqm)107 static unsigned int get_num_xgmi_sdma_engines(struct device_queue_manager *dqm)
108 {
109 	return dqm->dev->device_info->num_xgmi_sdma_engines;
110 }
111 
get_num_all_sdma_engines(struct device_queue_manager * dqm)112 static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm)
113 {
114 	return get_num_sdma_engines(dqm) + get_num_xgmi_sdma_engines(dqm);
115 }
116 
get_num_sdma_queues(struct device_queue_manager * dqm)117 unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
118 {
119 	return dqm->dev->device_info->num_sdma_engines
120 			* dqm->dev->device_info->num_sdma_queues_per_engine;
121 }
122 
get_num_xgmi_sdma_queues(struct device_queue_manager * dqm)123 unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
124 {
125 	return dqm->dev->device_info->num_xgmi_sdma_engines
126 			* dqm->dev->device_info->num_sdma_queues_per_engine;
127 }
128 
program_sh_mem_settings(struct device_queue_manager * dqm,struct qcm_process_device * qpd)129 void program_sh_mem_settings(struct device_queue_manager *dqm,
130 					struct qcm_process_device *qpd)
131 {
132 	return dqm->dev->kfd2kgd->program_sh_mem_settings(
133 						dqm->dev->kgd, qpd->vmid,
134 						qpd->sh_mem_config,
135 						qpd->sh_mem_ape1_base,
136 						qpd->sh_mem_ape1_limit,
137 						qpd->sh_mem_bases);
138 }
139 
increment_queue_count(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)140 static void increment_queue_count(struct device_queue_manager *dqm,
141 				  struct qcm_process_device *qpd,
142 				  struct queue *q)
143 {
144 	dqm->active_queue_count++;
145 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
146 	    q->properties.type == KFD_QUEUE_TYPE_DIQ)
147 		dqm->active_cp_queue_count++;
148 
149 	if (q->properties.is_gws) {
150 		dqm->gws_queue_count++;
151 		qpd->mapped_gws_queue = true;
152 	}
153 }
154 
decrement_queue_count(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)155 static void decrement_queue_count(struct device_queue_manager *dqm,
156 				  struct qcm_process_device *qpd,
157 				  struct queue *q)
158 {
159 	dqm->active_queue_count--;
160 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
161 	    q->properties.type == KFD_QUEUE_TYPE_DIQ)
162 		dqm->active_cp_queue_count--;
163 
164 	if (q->properties.is_gws) {
165 		dqm->gws_queue_count--;
166 		qpd->mapped_gws_queue = false;
167 	}
168 }
169 
allocate_doorbell(struct qcm_process_device * qpd,struct queue * q)170 static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q)
171 {
172 	struct kfd_dev *dev = qpd->dqm->dev;
173 
174 	if (!KFD_IS_SOC15(dev->device_info->asic_family)) {
175 		/* On pre-SOC15 chips we need to use the queue ID to
176 		 * preserve the user mode ABI.
177 		 */
178 		q->doorbell_id = q->properties.queue_id;
179 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
180 			q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
181 		/* For SDMA queues on SOC15 with 8-byte doorbell, use static
182 		 * doorbell assignments based on the engine and queue id.
183 		 * The doobell index distance between RLC (2*i) and (2*i+1)
184 		 * for a SDMA engine is 512.
185 		 */
186 		uint32_t *idx_offset =
187 				dev->shared_resources.sdma_doorbell_idx;
188 
189 		q->doorbell_id = idx_offset[q->properties.sdma_engine_id]
190 			+ (q->properties.sdma_queue_id & 1)
191 			* KFD_QUEUE_DOORBELL_MIRROR_OFFSET
192 			+ (q->properties.sdma_queue_id >> 1);
193 	} else {
194 		/* For CP queues on SOC15 reserve a free doorbell ID */
195 		unsigned int found;
196 
197 		found = find_first_zero_bit(qpd->doorbell_bitmap,
198 					    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
199 		if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
200 			pr_debug("No doorbells available");
201 			return -EBUSY;
202 		}
203 		set_bit(found, qpd->doorbell_bitmap);
204 		q->doorbell_id = found;
205 	}
206 
207 	q->properties.doorbell_off =
208 		kfd_get_doorbell_dw_offset_in_bar(dev, qpd_to_pdd(qpd),
209 					  q->doorbell_id);
210 	return 0;
211 }
212 
deallocate_doorbell(struct qcm_process_device * qpd,struct queue * q)213 static void deallocate_doorbell(struct qcm_process_device *qpd,
214 				struct queue *q)
215 {
216 	unsigned int old;
217 	struct kfd_dev *dev = qpd->dqm->dev;
218 
219 	if (!KFD_IS_SOC15(dev->device_info->asic_family) ||
220 	    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
221 	    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
222 		return;
223 
224 	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
225 	WARN_ON(!old);
226 }
227 
allocate_vmid(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)228 static int allocate_vmid(struct device_queue_manager *dqm,
229 			struct qcm_process_device *qpd,
230 			struct queue *q)
231 {
232 	int allocated_vmid = -1, i;
233 
234 	for (i = dqm->dev->vm_info.first_vmid_kfd;
235 			i <= dqm->dev->vm_info.last_vmid_kfd; i++) {
236 		if (!dqm->vmid_pasid[i]) {
237 			allocated_vmid = i;
238 			break;
239 		}
240 	}
241 
242 	if (allocated_vmid < 0) {
243 		pr_err("no more vmid to allocate\n");
244 		return -ENOSPC;
245 	}
246 
247 	pr_debug("vmid allocated: %d\n", allocated_vmid);
248 
249 	dqm->vmid_pasid[allocated_vmid] = q->process->pasid;
250 
251 	set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid);
252 
253 	qpd->vmid = allocated_vmid;
254 	q->properties.vmid = allocated_vmid;
255 
256 	program_sh_mem_settings(dqm, qpd);
257 
258 	/* qpd->page_table_base is set earlier when register_process()
259 	 * is called, i.e. when the first queue is created.
260 	 */
261 	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd,
262 			qpd->vmid,
263 			qpd->page_table_base);
264 	/* invalidate the VM context after pasid and vmid mapping is set up */
265 	kfd_flush_tlb(qpd_to_pdd(qpd));
266 
267 	if (dqm->dev->kfd2kgd->set_scratch_backing_va)
268 		dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->kgd,
269 				qpd->sh_hidden_private_base, qpd->vmid);
270 
271 	return 0;
272 }
273 
flush_texture_cache_nocpsch(struct kfd_dev * kdev,struct qcm_process_device * qpd)274 static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
275 				struct qcm_process_device *qpd)
276 {
277 	const struct packet_manager_funcs *pmf = qpd->dqm->packets.pmf;
278 	int ret;
279 
280 	if (!qpd->ib_kaddr)
281 		return -ENOMEM;
282 
283 	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
284 	if (ret)
285 		return ret;
286 
287 	return amdgpu_amdkfd_submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid,
288 				qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
289 				pmf->release_mem_size / sizeof(uint32_t));
290 }
291 
deallocate_vmid(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)292 static void deallocate_vmid(struct device_queue_manager *dqm,
293 				struct qcm_process_device *qpd,
294 				struct queue *q)
295 {
296 	/* On GFX v7, CP doesn't flush TC at dequeue */
297 	if (q->device->device_info->asic_family == CHIP_HAWAII)
298 		if (flush_texture_cache_nocpsch(q->device, qpd))
299 			pr_err("Failed to flush TC\n");
300 
301 	kfd_flush_tlb(qpd_to_pdd(qpd));
302 
303 	/* Release the vmid mapping */
304 	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
305 	dqm->vmid_pasid[qpd->vmid] = 0;
306 
307 	qpd->vmid = 0;
308 	q->properties.vmid = 0;
309 }
310 
create_queue_nocpsch(struct device_queue_manager * dqm,struct queue * q,struct qcm_process_device * qpd)311 static int create_queue_nocpsch(struct device_queue_manager *dqm,
312 				struct queue *q,
313 				struct qcm_process_device *qpd)
314 {
315 	struct mqd_manager *mqd_mgr;
316 	int retval;
317 
318 	dqm_lock(dqm);
319 
320 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
321 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
322 				dqm->total_queue_count);
323 		retval = -EPERM;
324 		goto out_unlock;
325 	}
326 
327 	if (list_empty(&qpd->queues_list)) {
328 		retval = allocate_vmid(dqm, qpd, q);
329 		if (retval)
330 			goto out_unlock;
331 	}
332 	q->properties.vmid = qpd->vmid;
333 	/*
334 	 * Eviction state logic: mark all queues as evicted, even ones
335 	 * not currently active. Restoring inactive queues later only
336 	 * updates the is_evicted flag but is a no-op otherwise.
337 	 */
338 	q->properties.is_evicted = !!qpd->evicted;
339 
340 	q->properties.tba_addr = qpd->tba_addr;
341 	q->properties.tma_addr = qpd->tma_addr;
342 
343 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
344 			q->properties.type)];
345 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
346 		retval = allocate_hqd(dqm, q);
347 		if (retval)
348 			goto deallocate_vmid;
349 		pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
350 			q->pipe, q->queue);
351 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
352 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
353 		retval = allocate_sdma_queue(dqm, q);
354 		if (retval)
355 			goto deallocate_vmid;
356 		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
357 	}
358 
359 	retval = allocate_doorbell(qpd, q);
360 	if (retval)
361 		goto out_deallocate_hqd;
362 
363 	/* Temporarily release dqm lock to avoid a circular lock dependency */
364 	dqm_unlock(dqm);
365 	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
366 	dqm_lock(dqm);
367 
368 	if (!q->mqd_mem_obj) {
369 		retval = -ENOMEM;
370 		goto out_deallocate_doorbell;
371 	}
372 	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
373 				&q->gart_mqd_addr, &q->properties);
374 	if (q->properties.is_active) {
375 		if (!dqm->sched_running) {
376 			WARN_ONCE(1, "Load non-HWS mqd while stopped\n");
377 			goto add_queue_to_list;
378 		}
379 
380 		if (WARN(q->process->mm != current->mm,
381 					"should only run in user thread"))
382 			retval = -EFAULT;
383 		else
384 			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
385 					q->queue, &q->properties, current->mm);
386 		if (retval)
387 			goto out_free_mqd;
388 	}
389 
390 add_queue_to_list:
391 	list_add(&q->list, &qpd->queues_list);
392 	qpd->queue_count++;
393 	if (q->properties.is_active)
394 		increment_queue_count(dqm, qpd, q);
395 
396 	/*
397 	 * Unconditionally increment this counter, regardless of the queue's
398 	 * type or whether the queue is active.
399 	 */
400 	dqm->total_queue_count++;
401 	pr_debug("Total of %d queues are accountable so far\n",
402 			dqm->total_queue_count);
403 	goto out_unlock;
404 
405 out_free_mqd:
406 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
407 out_deallocate_doorbell:
408 	deallocate_doorbell(qpd, q);
409 out_deallocate_hqd:
410 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
411 		deallocate_hqd(dqm, q);
412 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
413 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
414 		deallocate_sdma_queue(dqm, q);
415 deallocate_vmid:
416 	if (list_empty(&qpd->queues_list))
417 		deallocate_vmid(dqm, qpd, q);
418 out_unlock:
419 	dqm_unlock(dqm);
420 	return retval;
421 }
422 
allocate_hqd(struct device_queue_manager * dqm,struct queue * q)423 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
424 {
425 	bool set;
426 	int pipe, bit, i;
427 
428 	set = false;
429 
430 	for (pipe = dqm->next_pipe_to_allocate, i = 0;
431 			i < get_pipes_per_mec(dqm);
432 			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
433 
434 		if (!is_pipe_enabled(dqm, 0, pipe))
435 			continue;
436 
437 		if (dqm->allocated_queues[pipe] != 0) {
438 			bit = ffs(dqm->allocated_queues[pipe]) - 1;
439 			dqm->allocated_queues[pipe] &= ~(1 << bit);
440 			q->pipe = pipe;
441 			q->queue = bit;
442 			set = true;
443 			break;
444 		}
445 	}
446 
447 	if (!set)
448 		return -EBUSY;
449 
450 	pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
451 	/* horizontal hqd allocation */
452 	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
453 
454 	return 0;
455 }
456 
deallocate_hqd(struct device_queue_manager * dqm,struct queue * q)457 static inline void deallocate_hqd(struct device_queue_manager *dqm,
458 				struct queue *q)
459 {
460 	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
461 }
462 
463 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
464  * to avoid asynchronized access
465  */
destroy_queue_nocpsch_locked(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)466 static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
467 				struct qcm_process_device *qpd,
468 				struct queue *q)
469 {
470 	int retval;
471 	struct mqd_manager *mqd_mgr;
472 
473 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
474 			q->properties.type)];
475 
476 	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
477 		deallocate_hqd(dqm, q);
478 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
479 		deallocate_sdma_queue(dqm, q);
480 	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
481 		deallocate_sdma_queue(dqm, q);
482 	else {
483 		pr_debug("q->properties.type %d is invalid\n",
484 				q->properties.type);
485 		return -EINVAL;
486 	}
487 	dqm->total_queue_count--;
488 
489 	deallocate_doorbell(qpd, q);
490 
491 	if (!dqm->sched_running) {
492 		WARN_ONCE(1, "Destroy non-HWS queue while stopped\n");
493 		return 0;
494 	}
495 
496 	retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
497 				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
498 				KFD_UNMAP_LATENCY_MS,
499 				q->pipe, q->queue);
500 	if (retval == -ETIME)
501 		qpd->reset_wavefronts = true;
502 
503 	list_del(&q->list);
504 	if (list_empty(&qpd->queues_list)) {
505 		if (qpd->reset_wavefronts) {
506 			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
507 					dqm->dev);
508 			/* dbgdev_wave_reset_wavefronts has to be called before
509 			 * deallocate_vmid(), i.e. when vmid is still in use.
510 			 */
511 			dbgdev_wave_reset_wavefronts(dqm->dev,
512 					qpd->pqm->process);
513 			qpd->reset_wavefronts = false;
514 		}
515 
516 		deallocate_vmid(dqm, qpd, q);
517 	}
518 	qpd->queue_count--;
519 	if (q->properties.is_active)
520 		decrement_queue_count(dqm, qpd, q);
521 
522 	return retval;
523 }
524 
destroy_queue_nocpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)525 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
526 				struct qcm_process_device *qpd,
527 				struct queue *q)
528 {
529 	int retval;
530 	uint64_t sdma_val = 0;
531 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
532 	struct mqd_manager *mqd_mgr =
533 		dqm->mqd_mgrs[get_mqd_type_from_queue_type(q->properties.type)];
534 
535 	/* Get the SDMA queue stats */
536 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
537 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
538 		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
539 							&sdma_val);
540 		if (retval)
541 			pr_err("Failed to read SDMA queue counter for queue: %d\n",
542 				q->properties.queue_id);
543 	}
544 
545 	dqm_lock(dqm);
546 	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
547 	if (!retval)
548 		pdd->sdma_past_activity_counter += sdma_val;
549 	dqm_unlock(dqm);
550 
551 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
552 
553 	return retval;
554 }
555 
update_queue(struct device_queue_manager * dqm,struct queue * q)556 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
557 {
558 	int retval = 0;
559 	struct mqd_manager *mqd_mgr;
560 	struct kfd_process_device *pdd;
561 	bool prev_active = false;
562 
563 	dqm_lock(dqm);
564 	pdd = kfd_get_process_device_data(q->device, q->process);
565 	if (!pdd) {
566 		retval = -ENODEV;
567 		goto out_unlock;
568 	}
569 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
570 			q->properties.type)];
571 
572 	/* Save previous activity state for counters */
573 	prev_active = q->properties.is_active;
574 
575 	/* Make sure the queue is unmapped before updating the MQD */
576 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
577 		retval = unmap_queues_cpsch(dqm,
578 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
579 		if (retval) {
580 			pr_err("unmap queue failed\n");
581 			goto out_unlock;
582 		}
583 	} else if (prev_active &&
584 		   (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
585 		    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
586 		    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
587 
588 		if (!dqm->sched_running) {
589 			WARN_ONCE(1, "Update non-HWS queue while stopped\n");
590 			goto out_unlock;
591 		}
592 
593 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
594 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
595 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
596 		if (retval) {
597 			pr_err("destroy mqd failed\n");
598 			goto out_unlock;
599 		}
600 	}
601 
602 	mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties);
603 
604 	/*
605 	 * check active state vs. the previous state and modify
606 	 * counter accordingly. map_queues_cpsch uses the
607 	 * dqm->active_queue_count to determine whether a new runlist must be
608 	 * uploaded.
609 	 */
610 	if (q->properties.is_active && !prev_active) {
611 		increment_queue_count(dqm, &pdd->qpd, q);
612 	} else if (!q->properties.is_active && prev_active) {
613 		decrement_queue_count(dqm, &pdd->qpd, q);
614 	} else if (q->gws && !q->properties.is_gws) {
615 		if (q->properties.is_active) {
616 			dqm->gws_queue_count++;
617 			pdd->qpd.mapped_gws_queue = true;
618 		}
619 		q->properties.is_gws = true;
620 	} else if (!q->gws && q->properties.is_gws) {
621 		if (q->properties.is_active) {
622 			dqm->gws_queue_count--;
623 			pdd->qpd.mapped_gws_queue = false;
624 		}
625 		q->properties.is_gws = false;
626 	}
627 
628 	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
629 		retval = map_queues_cpsch(dqm);
630 	else if (q->properties.is_active &&
631 		 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
632 		  q->properties.type == KFD_QUEUE_TYPE_SDMA ||
633 		  q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
634 		if (WARN(q->process->mm != current->mm,
635 			 "should only run in user thread"))
636 			retval = -EFAULT;
637 		else
638 			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
639 						   q->pipe, q->queue,
640 						   &q->properties, current->mm);
641 	}
642 
643 out_unlock:
644 	dqm_unlock(dqm);
645 	return retval;
646 }
647 
evict_process_queues_nocpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)648 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
649 					struct qcm_process_device *qpd)
650 {
651 	struct queue *q;
652 	struct mqd_manager *mqd_mgr;
653 	struct kfd_process_device *pdd;
654 	int retval, ret = 0;
655 
656 	dqm_lock(dqm);
657 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
658 		goto out;
659 
660 	pdd = qpd_to_pdd(qpd);
661 	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
662 			    pdd->process->pasid);
663 
664 	pdd->last_evict_timestamp = get_jiffies_64();
665 	/* Mark all queues as evicted. Deactivate all active queues on
666 	 * the qpd.
667 	 */
668 	list_for_each_entry(q, &qpd->queues_list, list) {
669 		q->properties.is_evicted = true;
670 		if (!q->properties.is_active)
671 			continue;
672 
673 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
674 				q->properties.type)];
675 		q->properties.is_active = false;
676 		decrement_queue_count(dqm, qpd, q);
677 
678 		if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n"))
679 			continue;
680 
681 		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
682 				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
683 				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
684 		if (retval && !ret)
685 			/* Return the first error, but keep going to
686 			 * maintain a consistent eviction state
687 			 */
688 			ret = retval;
689 	}
690 
691 out:
692 	dqm_unlock(dqm);
693 	return ret;
694 }
695 
evict_process_queues_cpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)696 static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
697 				      struct qcm_process_device *qpd)
698 {
699 	struct queue *q;
700 	struct kfd_process_device *pdd;
701 	int retval = 0;
702 
703 	dqm_lock(dqm);
704 	if (qpd->evicted++ > 0) /* already evicted, do nothing */
705 		goto out;
706 
707 	pdd = qpd_to_pdd(qpd);
708 	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
709 			    pdd->process->pasid);
710 
711 	/* Mark all queues as evicted. Deactivate all active queues on
712 	 * the qpd.
713 	 */
714 	list_for_each_entry(q, &qpd->queues_list, list) {
715 		q->properties.is_evicted = true;
716 		if (!q->properties.is_active)
717 			continue;
718 
719 		q->properties.is_active = false;
720 		decrement_queue_count(dqm, qpd, q);
721 	}
722 	pdd->last_evict_timestamp = get_jiffies_64();
723 	retval = execute_queues_cpsch(dqm,
724 				qpd->is_debug ?
725 				KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
726 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
727 
728 out:
729 	dqm_unlock(dqm);
730 	return retval;
731 }
732 
restore_process_queues_nocpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)733 static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
734 					  struct qcm_process_device *qpd)
735 {
736 	struct mm_struct *mm = NULL;
737 	struct queue *q;
738 	struct mqd_manager *mqd_mgr;
739 	struct kfd_process_device *pdd;
740 	uint64_t pd_base;
741 	uint64_t eviction_duration;
742 	int retval, ret = 0;
743 
744 	pdd = qpd_to_pdd(qpd);
745 	/* Retrieve PD base */
746 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
747 
748 	dqm_lock(dqm);
749 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
750 		goto out;
751 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
752 		qpd->evicted--;
753 		goto out;
754 	}
755 
756 	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
757 			    pdd->process->pasid);
758 
759 	/* Update PD Base in QPD */
760 	qpd->page_table_base = pd_base;
761 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
762 
763 	if (!list_empty(&qpd->queues_list)) {
764 		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
765 				dqm->dev->kgd,
766 				qpd->vmid,
767 				qpd->page_table_base);
768 		kfd_flush_tlb(pdd);
769 	}
770 
771 	/* Take a safe reference to the mm_struct, which may otherwise
772 	 * disappear even while the kfd_process is still referenced.
773 	 */
774 	mm = get_task_mm(pdd->process->lead_thread);
775 	if (!mm) {
776 		ret = -EFAULT;
777 		goto out;
778 	}
779 
780 	/* Remove the eviction flags. Activate queues that are not
781 	 * inactive for other reasons.
782 	 */
783 	list_for_each_entry(q, &qpd->queues_list, list) {
784 		q->properties.is_evicted = false;
785 		if (!QUEUE_IS_ACTIVE(q->properties))
786 			continue;
787 
788 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
789 				q->properties.type)];
790 		q->properties.is_active = true;
791 		increment_queue_count(dqm, qpd, q);
792 
793 		if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n"))
794 			continue;
795 
796 		retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
797 				       q->queue, &q->properties, mm);
798 		if (retval && !ret)
799 			/* Return the first error, but keep going to
800 			 * maintain a consistent eviction state
801 			 */
802 			ret = retval;
803 	}
804 	qpd->evicted = 0;
805 	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
806 	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
807 out:
808 	if (mm)
809 		mmput(mm);
810 	dqm_unlock(dqm);
811 	return ret;
812 }
813 
restore_process_queues_cpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)814 static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
815 					struct qcm_process_device *qpd)
816 {
817 	struct queue *q;
818 	struct kfd_process_device *pdd;
819 	uint64_t pd_base;
820 	uint64_t eviction_duration;
821 	int retval = 0;
822 
823 	pdd = qpd_to_pdd(qpd);
824 	/* Retrieve PD base */
825 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
826 
827 	dqm_lock(dqm);
828 	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
829 		goto out;
830 	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
831 		qpd->evicted--;
832 		goto out;
833 	}
834 
835 	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
836 			    pdd->process->pasid);
837 
838 	/* Update PD Base in QPD */
839 	qpd->page_table_base = pd_base;
840 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
841 
842 	/* activate all active queues on the qpd */
843 	list_for_each_entry(q, &qpd->queues_list, list) {
844 		q->properties.is_evicted = false;
845 		if (!QUEUE_IS_ACTIVE(q->properties))
846 			continue;
847 
848 		q->properties.is_active = true;
849 		increment_queue_count(dqm, &pdd->qpd, q);
850 	}
851 	retval = execute_queues_cpsch(dqm,
852 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
853 	qpd->evicted = 0;
854 	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
855 	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
856 out:
857 	dqm_unlock(dqm);
858 	return retval;
859 }
860 
register_process(struct device_queue_manager * dqm,struct qcm_process_device * qpd)861 static int register_process(struct device_queue_manager *dqm,
862 					struct qcm_process_device *qpd)
863 {
864 	struct device_process_node *n;
865 	struct kfd_process_device *pdd;
866 	uint64_t pd_base;
867 	int retval;
868 
869 	n = kzalloc(sizeof(*n), GFP_KERNEL);
870 	if (!n)
871 		return -ENOMEM;
872 
873 	n->qpd = qpd;
874 
875 	pdd = qpd_to_pdd(qpd);
876 	/* Retrieve PD base */
877 	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
878 
879 	dqm_lock(dqm);
880 	list_add(&n->list, &dqm->queues);
881 
882 	/* Update PD Base in QPD */
883 	qpd->page_table_base = pd_base;
884 	pr_debug("Updated PD address to 0x%llx\n", pd_base);
885 
886 	retval = dqm->asic_ops.update_qpd(dqm, qpd);
887 
888 	dqm->processes_count++;
889 
890 	dqm_unlock(dqm);
891 
892 	/* Outside the DQM lock because under the DQM lock we can't do
893 	 * reclaim or take other locks that others hold while reclaiming.
894 	 */
895 	kfd_inc_compute_active(dqm->dev);
896 
897 	return retval;
898 }
899 
unregister_process(struct device_queue_manager * dqm,struct qcm_process_device * qpd)900 static int unregister_process(struct device_queue_manager *dqm,
901 					struct qcm_process_device *qpd)
902 {
903 	int retval;
904 	struct device_process_node *cur, *next;
905 
906 	pr_debug("qpd->queues_list is %s\n",
907 			list_empty(&qpd->queues_list) ? "empty" : "not empty");
908 
909 	retval = 0;
910 	dqm_lock(dqm);
911 
912 	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
913 		if (qpd == cur->qpd) {
914 			list_del(&cur->list);
915 			kfree(cur);
916 			dqm->processes_count--;
917 			goto out;
918 		}
919 	}
920 	/* qpd not found in dqm list */
921 	retval = 1;
922 out:
923 	dqm_unlock(dqm);
924 
925 	/* Outside the DQM lock because under the DQM lock we can't do
926 	 * reclaim or take other locks that others hold while reclaiming.
927 	 */
928 	if (!retval)
929 		kfd_dec_compute_active(dqm->dev);
930 
931 	return retval;
932 }
933 
934 static int
set_pasid_vmid_mapping(struct device_queue_manager * dqm,u32 pasid,unsigned int vmid)935 set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid,
936 			unsigned int vmid)
937 {
938 	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
939 						dqm->dev->kgd, pasid, vmid);
940 }
941 
init_interrupts(struct device_queue_manager * dqm)942 static void init_interrupts(struct device_queue_manager *dqm)
943 {
944 	unsigned int i;
945 
946 	for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
947 		if (is_pipe_enabled(dqm, 0, i))
948 			dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
949 }
950 
initialize_nocpsch(struct device_queue_manager * dqm)951 static int initialize_nocpsch(struct device_queue_manager *dqm)
952 {
953 	int pipe, queue;
954 
955 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
956 
957 	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
958 					sizeof(unsigned int), GFP_KERNEL);
959 	if (!dqm->allocated_queues)
960 		return -ENOMEM;
961 
962 	mutex_init(&dqm->lock_hidden);
963 	INIT_LIST_HEAD(&dqm->queues);
964 	dqm->active_queue_count = dqm->next_pipe_to_allocate = 0;
965 	dqm->active_cp_queue_count = 0;
966 	dqm->gws_queue_count = 0;
967 
968 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
969 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
970 
971 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
972 			if (test_bit(pipe_offset + queue,
973 				     dqm->dev->shared_resources.cp_queue_bitmap))
974 				dqm->allocated_queues[pipe] |= 1 << queue;
975 	}
976 
977 	memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid));
978 
979 	dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm));
980 	dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm));
981 
982 	return 0;
983 }
984 
uninitialize(struct device_queue_manager * dqm)985 static void uninitialize(struct device_queue_manager *dqm)
986 {
987 	int i;
988 
989 	WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0);
990 
991 	kfree(dqm->allocated_queues);
992 	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
993 		kfree(dqm->mqd_mgrs[i]);
994 	mutex_destroy(&dqm->lock_hidden);
995 }
996 
start_nocpsch(struct device_queue_manager * dqm)997 static int start_nocpsch(struct device_queue_manager *dqm)
998 {
999 	pr_info("SW scheduler is used");
1000 	init_interrupts(dqm);
1001 
1002 	if (dqm->dev->device_info->asic_family == CHIP_HAWAII)
1003 		return pm_init(&dqm->packets, dqm);
1004 	dqm->sched_running = true;
1005 
1006 	return 0;
1007 }
1008 
stop_nocpsch(struct device_queue_manager * dqm)1009 static int stop_nocpsch(struct device_queue_manager *dqm)
1010 {
1011 	if (dqm->dev->device_info->asic_family == CHIP_HAWAII)
1012 		pm_uninit(&dqm->packets, false);
1013 	dqm->sched_running = false;
1014 
1015 	return 0;
1016 }
1017 
pre_reset(struct device_queue_manager * dqm)1018 static void pre_reset(struct device_queue_manager *dqm)
1019 {
1020 	dqm_lock(dqm);
1021 	dqm->is_resetting = true;
1022 	dqm_unlock(dqm);
1023 }
1024 
allocate_sdma_queue(struct device_queue_manager * dqm,struct queue * q)1025 static int allocate_sdma_queue(struct device_queue_manager *dqm,
1026 				struct queue *q)
1027 {
1028 	int bit;
1029 
1030 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1031 		if (dqm->sdma_bitmap == 0) {
1032 			pr_err("No more SDMA queue to allocate\n");
1033 			return -ENOMEM;
1034 		}
1035 
1036 		bit = __ffs64(dqm->sdma_bitmap);
1037 		dqm->sdma_bitmap &= ~(1ULL << bit);
1038 		q->sdma_id = bit;
1039 		q->properties.sdma_engine_id = q->sdma_id %
1040 				get_num_sdma_engines(dqm);
1041 		q->properties.sdma_queue_id = q->sdma_id /
1042 				get_num_sdma_engines(dqm);
1043 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1044 		if (dqm->xgmi_sdma_bitmap == 0) {
1045 			pr_err("No more XGMI SDMA queue to allocate\n");
1046 			return -ENOMEM;
1047 		}
1048 		bit = __ffs64(dqm->xgmi_sdma_bitmap);
1049 		dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
1050 		q->sdma_id = bit;
1051 		/* sdma_engine_id is sdma id including
1052 		 * both PCIe-optimized SDMAs and XGMI-
1053 		 * optimized SDMAs. The calculation below
1054 		 * assumes the first N engines are always
1055 		 * PCIe-optimized ones
1056 		 */
1057 		q->properties.sdma_engine_id = get_num_sdma_engines(dqm) +
1058 				q->sdma_id % get_num_xgmi_sdma_engines(dqm);
1059 		q->properties.sdma_queue_id = q->sdma_id /
1060 				get_num_xgmi_sdma_engines(dqm);
1061 	}
1062 
1063 	pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
1064 	pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
1065 
1066 	return 0;
1067 }
1068 
deallocate_sdma_queue(struct device_queue_manager * dqm,struct queue * q)1069 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
1070 				struct queue *q)
1071 {
1072 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1073 		if (q->sdma_id >= get_num_sdma_queues(dqm))
1074 			return;
1075 		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
1076 	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1077 		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
1078 			return;
1079 		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
1080 	}
1081 }
1082 
1083 /*
1084  * Device Queue Manager implementation for cp scheduler
1085  */
1086 
set_sched_resources(struct device_queue_manager * dqm)1087 static int set_sched_resources(struct device_queue_manager *dqm)
1088 {
1089 	int i, mec;
1090 	struct scheduling_resources res;
1091 
1092 	res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
1093 
1094 	res.queue_mask = 0;
1095 	for (i = 0; i < KGD_MAX_QUEUES; ++i) {
1096 		mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
1097 			/ dqm->dev->shared_resources.num_pipe_per_mec;
1098 
1099 		if (!test_bit(i, dqm->dev->shared_resources.cp_queue_bitmap))
1100 			continue;
1101 
1102 		/* only acquire queues from the first MEC */
1103 		if (mec > 0)
1104 			continue;
1105 
1106 		/* This situation may be hit in the future if a new HW
1107 		 * generation exposes more than 64 queues. If so, the
1108 		 * definition of res.queue_mask needs updating
1109 		 */
1110 		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
1111 			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
1112 			break;
1113 		}
1114 
1115 		res.queue_mask |= 1ull
1116 			<< amdgpu_queue_mask_bit_to_set_resource_bit(
1117 				(struct amdgpu_device *)dqm->dev->kgd, i);
1118 	}
1119 	res.gws_mask = ~0ull;
1120 	res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1121 
1122 	pr_debug("Scheduling resources:\n"
1123 			"vmid mask: 0x%8X\n"
1124 			"queue mask: 0x%8llX\n",
1125 			res.vmid_mask, res.queue_mask);
1126 
1127 	return pm_send_set_resources(&dqm->packets, &res);
1128 }
1129 
initialize_cpsch(struct device_queue_manager * dqm)1130 static int initialize_cpsch(struct device_queue_manager *dqm)
1131 {
1132 	uint64_t num_sdma_queues;
1133 	uint64_t num_xgmi_sdma_queues;
1134 
1135 	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1136 
1137 	mutex_init(&dqm->lock_hidden);
1138 	INIT_LIST_HEAD(&dqm->queues);
1139 	dqm->active_queue_count = dqm->processes_count = 0;
1140 	dqm->active_cp_queue_count = 0;
1141 	dqm->gws_queue_count = 0;
1142 	dqm->active_runlist = false;
1143 
1144 	num_sdma_queues = get_num_sdma_queues(dqm);
1145 	if (num_sdma_queues >= BITS_PER_TYPE(dqm->sdma_bitmap))
1146 		dqm->sdma_bitmap = ULLONG_MAX;
1147 	else
1148 		dqm->sdma_bitmap = (BIT_ULL(num_sdma_queues) - 1);
1149 
1150 	num_xgmi_sdma_queues = get_num_xgmi_sdma_queues(dqm);
1151 	if (num_xgmi_sdma_queues >= BITS_PER_TYPE(dqm->xgmi_sdma_bitmap))
1152 		dqm->xgmi_sdma_bitmap = ULLONG_MAX;
1153 	else
1154 		dqm->xgmi_sdma_bitmap = (BIT_ULL(num_xgmi_sdma_queues) - 1);
1155 
1156 	INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
1157 
1158 	return 0;
1159 }
1160 
start_cpsch(struct device_queue_manager * dqm)1161 static int start_cpsch(struct device_queue_manager *dqm)
1162 {
1163 	int retval;
1164 
1165 	retval = 0;
1166 
1167 	retval = pm_init(&dqm->packets, dqm);
1168 	if (retval)
1169 		goto fail_packet_manager_init;
1170 
1171 	retval = set_sched_resources(dqm);
1172 	if (retval)
1173 		goto fail_set_sched_resources;
1174 
1175 	pr_debug("Allocating fence memory\n");
1176 
1177 	/* allocate fence memory on the gart */
1178 	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1179 					&dqm->fence_mem);
1180 
1181 	if (retval)
1182 		goto fail_allocate_vidmem;
1183 
1184 	dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr;
1185 	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1186 
1187 	init_interrupts(dqm);
1188 
1189 	dqm_lock(dqm);
1190 	/* clear hang status when driver try to start the hw scheduler */
1191 	dqm->is_hws_hang = false;
1192 	dqm->is_resetting = false;
1193 	dqm->sched_running = true;
1194 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1195 	dqm_unlock(dqm);
1196 
1197 	return 0;
1198 fail_allocate_vidmem:
1199 fail_set_sched_resources:
1200 	pm_uninit(&dqm->packets, false);
1201 fail_packet_manager_init:
1202 	return retval;
1203 }
1204 
stop_cpsch(struct device_queue_manager * dqm)1205 static int stop_cpsch(struct device_queue_manager *dqm)
1206 {
1207 	bool hanging;
1208 
1209 	dqm_lock(dqm);
1210 	if (!dqm->sched_running) {
1211 		dqm_unlock(dqm);
1212 		return 0;
1213 	}
1214 
1215 	if (!dqm->is_hws_hang)
1216 		unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1217 	hanging = dqm->is_hws_hang || dqm->is_resetting;
1218 	dqm->sched_running = false;
1219 	dqm_unlock(dqm);
1220 
1221 	pm_release_ib(&dqm->packets);
1222 
1223 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1224 	pm_uninit(&dqm->packets, hanging);
1225 
1226 	return 0;
1227 }
1228 
create_kernel_queue_cpsch(struct device_queue_manager * dqm,struct kernel_queue * kq,struct qcm_process_device * qpd)1229 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1230 					struct kernel_queue *kq,
1231 					struct qcm_process_device *qpd)
1232 {
1233 	dqm_lock(dqm);
1234 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1235 		pr_warn("Can't create new kernel queue because %d queues were already created\n",
1236 				dqm->total_queue_count);
1237 		dqm_unlock(dqm);
1238 		return -EPERM;
1239 	}
1240 
1241 	/*
1242 	 * Unconditionally increment this counter, regardless of the queue's
1243 	 * type or whether the queue is active.
1244 	 */
1245 	dqm->total_queue_count++;
1246 	pr_debug("Total of %d queues are accountable so far\n",
1247 			dqm->total_queue_count);
1248 
1249 	list_add(&kq->list, &qpd->priv_queue_list);
1250 	increment_queue_count(dqm, qpd, kq->queue);
1251 	qpd->is_debug = true;
1252 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1253 	dqm_unlock(dqm);
1254 
1255 	return 0;
1256 }
1257 
destroy_kernel_queue_cpsch(struct device_queue_manager * dqm,struct kernel_queue * kq,struct qcm_process_device * qpd)1258 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1259 					struct kernel_queue *kq,
1260 					struct qcm_process_device *qpd)
1261 {
1262 	dqm_lock(dqm);
1263 	list_del(&kq->list);
1264 	decrement_queue_count(dqm, qpd, kq->queue);
1265 	qpd->is_debug = false;
1266 	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1267 	/*
1268 	 * Unconditionally decrement this counter, regardless of the queue's
1269 	 * type.
1270 	 */
1271 	dqm->total_queue_count--;
1272 	pr_debug("Total of %d queues are accountable so far\n",
1273 			dqm->total_queue_count);
1274 	dqm_unlock(dqm);
1275 }
1276 
create_queue_cpsch(struct device_queue_manager * dqm,struct queue * q,struct qcm_process_device * qpd)1277 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1278 			struct qcm_process_device *qpd)
1279 {
1280 	int retval;
1281 	struct mqd_manager *mqd_mgr;
1282 
1283 	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1284 		pr_warn("Can't create new usermode queue because %d queues were already created\n",
1285 				dqm->total_queue_count);
1286 		retval = -EPERM;
1287 		goto out;
1288 	}
1289 
1290 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1291 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1292 		dqm_lock(dqm);
1293 		retval = allocate_sdma_queue(dqm, q);
1294 		dqm_unlock(dqm);
1295 		if (retval)
1296 			goto out;
1297 	}
1298 
1299 	retval = allocate_doorbell(qpd, q);
1300 	if (retval)
1301 		goto out_deallocate_sdma_queue;
1302 
1303 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1304 			q->properties.type)];
1305 
1306 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1307 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1308 		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1309 	q->properties.tba_addr = qpd->tba_addr;
1310 	q->properties.tma_addr = qpd->tma_addr;
1311 	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
1312 	if (!q->mqd_mem_obj) {
1313 		retval = -ENOMEM;
1314 		goto out_deallocate_doorbell;
1315 	}
1316 
1317 	dqm_lock(dqm);
1318 	/*
1319 	 * Eviction state logic: mark all queues as evicted, even ones
1320 	 * not currently active. Restoring inactive queues later only
1321 	 * updates the is_evicted flag but is a no-op otherwise.
1322 	 */
1323 	q->properties.is_evicted = !!qpd->evicted;
1324 	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1325 				&q->gart_mqd_addr, &q->properties);
1326 
1327 	list_add(&q->list, &qpd->queues_list);
1328 	qpd->queue_count++;
1329 
1330 	if (q->properties.is_active) {
1331 		increment_queue_count(dqm, qpd, q);
1332 
1333 		execute_queues_cpsch(dqm,
1334 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1335 	}
1336 
1337 	/*
1338 	 * Unconditionally increment this counter, regardless of the queue's
1339 	 * type or whether the queue is active.
1340 	 */
1341 	dqm->total_queue_count++;
1342 
1343 	pr_debug("Total of %d queues are accountable so far\n",
1344 			dqm->total_queue_count);
1345 
1346 	dqm_unlock(dqm);
1347 	return retval;
1348 
1349 out_deallocate_doorbell:
1350 	deallocate_doorbell(qpd, q);
1351 out_deallocate_sdma_queue:
1352 	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1353 		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1354 		dqm_lock(dqm);
1355 		deallocate_sdma_queue(dqm, q);
1356 		dqm_unlock(dqm);
1357 	}
1358 out:
1359 	return retval;
1360 }
1361 
amdkfd_fence_wait_timeout(uint64_t * fence_addr,uint64_t fence_value,unsigned int timeout_ms)1362 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1363 				uint64_t fence_value,
1364 				unsigned int timeout_ms)
1365 {
1366 	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
1367 
1368 	while (*fence_addr != fence_value) {
1369 		if (time_after(jiffies, end_jiffies)) {
1370 			pr_err("qcm fence wait loop timeout expired\n");
1371 			/* In HWS case, this is used to halt the driver thread
1372 			 * in order not to mess up CP states before doing
1373 			 * scandumps for FW debugging.
1374 			 */
1375 			while (halt_if_hws_hang)
1376 				schedule();
1377 
1378 			return -ETIME;
1379 		}
1380 		schedule();
1381 	}
1382 
1383 	return 0;
1384 }
1385 
1386 /* dqm->lock mutex has to be locked before calling this function */
map_queues_cpsch(struct device_queue_manager * dqm)1387 static int map_queues_cpsch(struct device_queue_manager *dqm)
1388 {
1389 	int retval;
1390 
1391 	if (!dqm->sched_running)
1392 		return 0;
1393 	if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0)
1394 		return 0;
1395 	if (dqm->active_runlist)
1396 		return 0;
1397 
1398 	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1399 	pr_debug("%s sent runlist\n", __func__);
1400 	if (retval) {
1401 		pr_err("failed to execute runlist\n");
1402 		return retval;
1403 	}
1404 	dqm->active_runlist = true;
1405 
1406 	return retval;
1407 }
1408 
1409 /* dqm->lock mutex has to be locked before calling this function */
unmap_queues_cpsch(struct device_queue_manager * dqm,enum kfd_unmap_queues_filter filter,uint32_t filter_param)1410 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1411 				enum kfd_unmap_queues_filter filter,
1412 				uint32_t filter_param)
1413 {
1414 	int retval = 0;
1415 
1416 	if (!dqm->sched_running)
1417 		return 0;
1418 	if (dqm->is_hws_hang)
1419 		return -EIO;
1420 	if (!dqm->active_runlist)
1421 		return retval;
1422 
1423 	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
1424 			filter, filter_param, false, 0);
1425 	if (retval)
1426 		return retval;
1427 
1428 	*dqm->fence_addr = KFD_FENCE_INIT;
1429 	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
1430 				KFD_FENCE_COMPLETED);
1431 	/* should be timed out */
1432 	retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
1433 				queue_preemption_timeout_ms);
1434 	if (retval) {
1435 		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1436 		dqm->is_hws_hang = true;
1437 		/* It's possible we're detecting a HWS hang in the
1438 		 * middle of a GPU reset. No need to schedule another
1439 		 * reset in this case.
1440 		 */
1441 		if (!dqm->is_resetting)
1442 			schedule_work(&dqm->hw_exception_work);
1443 		return retval;
1444 	}
1445 
1446 	pm_release_ib(&dqm->packets);
1447 	dqm->active_runlist = false;
1448 
1449 	return retval;
1450 }
1451 
1452 /* dqm->lock mutex has to be locked before calling this function */
execute_queues_cpsch(struct device_queue_manager * dqm,enum kfd_unmap_queues_filter filter,uint32_t filter_param)1453 static int execute_queues_cpsch(struct device_queue_manager *dqm,
1454 				enum kfd_unmap_queues_filter filter,
1455 				uint32_t filter_param)
1456 {
1457 	int retval;
1458 
1459 	if (dqm->is_hws_hang)
1460 		return -EIO;
1461 	retval = unmap_queues_cpsch(dqm, filter, filter_param);
1462 	if (retval)
1463 		return retval;
1464 
1465 	return map_queues_cpsch(dqm);
1466 }
1467 
destroy_queue_cpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd,struct queue * q)1468 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1469 				struct qcm_process_device *qpd,
1470 				struct queue *q)
1471 {
1472 	int retval;
1473 	struct mqd_manager *mqd_mgr;
1474 	uint64_t sdma_val = 0;
1475 	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
1476 
1477 	/* Get the SDMA queue stats */
1478 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1479 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1480 		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
1481 							&sdma_val);
1482 		if (retval)
1483 			pr_err("Failed to read SDMA queue counter for queue: %d\n",
1484 				q->properties.queue_id);
1485 	}
1486 
1487 	retval = 0;
1488 
1489 	/* remove queue from list to prevent rescheduling after preemption */
1490 	dqm_lock(dqm);
1491 
1492 	if (qpd->is_debug) {
1493 		/*
1494 		 * error, currently we do not allow to destroy a queue
1495 		 * of a currently debugged process
1496 		 */
1497 		retval = -EBUSY;
1498 		goto failed_try_destroy_debugged_queue;
1499 
1500 	}
1501 
1502 	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1503 			q->properties.type)];
1504 
1505 	deallocate_doorbell(qpd, q);
1506 
1507 	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1508 	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1509 		deallocate_sdma_queue(dqm, q);
1510 		pdd->sdma_past_activity_counter += sdma_val;
1511 	}
1512 
1513 	list_del(&q->list);
1514 	qpd->queue_count--;
1515 	if (q->properties.is_active) {
1516 		decrement_queue_count(dqm, qpd, q);
1517 		retval = execute_queues_cpsch(dqm,
1518 				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1519 		if (retval == -ETIME)
1520 			qpd->reset_wavefronts = true;
1521 	}
1522 
1523 	/*
1524 	 * Unconditionally decrement this counter, regardless of the queue's
1525 	 * type
1526 	 */
1527 	dqm->total_queue_count--;
1528 	pr_debug("Total of %d queues are accountable so far\n",
1529 			dqm->total_queue_count);
1530 
1531 	dqm_unlock(dqm);
1532 
1533 	/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
1534 	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1535 
1536 	return retval;
1537 
1538 failed_try_destroy_debugged_queue:
1539 
1540 	dqm_unlock(dqm);
1541 	return retval;
1542 }
1543 
1544 /*
1545  * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1546  * stay in user mode.
1547  */
1548 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1549 /* APE1 limit is inclusive and 64K aligned. */
1550 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1551 
set_cache_memory_policy(struct device_queue_manager * dqm,struct qcm_process_device * qpd,enum cache_policy default_policy,enum cache_policy alternate_policy,void __user * alternate_aperture_base,uint64_t alternate_aperture_size)1552 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1553 				   struct qcm_process_device *qpd,
1554 				   enum cache_policy default_policy,
1555 				   enum cache_policy alternate_policy,
1556 				   void __user *alternate_aperture_base,
1557 				   uint64_t alternate_aperture_size)
1558 {
1559 	bool retval = true;
1560 
1561 	if (!dqm->asic_ops.set_cache_memory_policy)
1562 		return retval;
1563 
1564 	dqm_lock(dqm);
1565 
1566 	if (alternate_aperture_size == 0) {
1567 		/* base > limit disables APE1 */
1568 		qpd->sh_mem_ape1_base = 1;
1569 		qpd->sh_mem_ape1_limit = 0;
1570 	} else {
1571 		/*
1572 		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1573 		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1574 		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1575 		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1576 		 * Verify that the base and size parameters can be
1577 		 * represented in this format and convert them.
1578 		 * Additionally restrict APE1 to user-mode addresses.
1579 		 */
1580 
1581 		uint64_t base = (uintptr_t)alternate_aperture_base;
1582 		uint64_t limit = base + alternate_aperture_size - 1;
1583 
1584 		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1585 		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1586 			retval = false;
1587 			goto out;
1588 		}
1589 
1590 		qpd->sh_mem_ape1_base = base >> 16;
1591 		qpd->sh_mem_ape1_limit = limit >> 16;
1592 	}
1593 
1594 	retval = dqm->asic_ops.set_cache_memory_policy(
1595 			dqm,
1596 			qpd,
1597 			default_policy,
1598 			alternate_policy,
1599 			alternate_aperture_base,
1600 			alternate_aperture_size);
1601 
1602 	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1603 		program_sh_mem_settings(dqm, qpd);
1604 
1605 	pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1606 		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1607 		qpd->sh_mem_ape1_limit);
1608 
1609 out:
1610 	dqm_unlock(dqm);
1611 	return retval;
1612 }
1613 
set_trap_handler(struct device_queue_manager * dqm,struct qcm_process_device * qpd,uint64_t tba_addr,uint64_t tma_addr)1614 static int set_trap_handler(struct device_queue_manager *dqm,
1615 				struct qcm_process_device *qpd,
1616 				uint64_t tba_addr,
1617 				uint64_t tma_addr)
1618 {
1619 	uint64_t *tma;
1620 
1621 	if (dqm->dev->cwsr_enabled) {
1622 		/* Jump from CWSR trap handler to user trap */
1623 		tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1624 		tma[0] = tba_addr;
1625 		tma[1] = tma_addr;
1626 	} else {
1627 		qpd->tba_addr = tba_addr;
1628 		qpd->tma_addr = tma_addr;
1629 	}
1630 
1631 	return 0;
1632 }
1633 
process_termination_nocpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)1634 static int process_termination_nocpsch(struct device_queue_manager *dqm,
1635 		struct qcm_process_device *qpd)
1636 {
1637 	struct queue *q;
1638 	struct device_process_node *cur, *next_dpn;
1639 	int retval = 0;
1640 	bool found = false;
1641 
1642 	dqm_lock(dqm);
1643 
1644 	/* Clear all user mode queues */
1645 	while (!list_empty(&qpd->queues_list)) {
1646 		struct mqd_manager *mqd_mgr;
1647 		int ret;
1648 
1649 		q = list_first_entry(&qpd->queues_list, struct queue, list);
1650 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1651 				q->properties.type)];
1652 		ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
1653 		if (ret)
1654 			retval = ret;
1655 		dqm_unlock(dqm);
1656 		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1657 		dqm_lock(dqm);
1658 	}
1659 
1660 	/* Unregister process */
1661 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1662 		if (qpd == cur->qpd) {
1663 			list_del(&cur->list);
1664 			kfree(cur);
1665 			dqm->processes_count--;
1666 			found = true;
1667 			break;
1668 		}
1669 	}
1670 
1671 	dqm_unlock(dqm);
1672 
1673 	/* Outside the DQM lock because under the DQM lock we can't do
1674 	 * reclaim or take other locks that others hold while reclaiming.
1675 	 */
1676 	if (found)
1677 		kfd_dec_compute_active(dqm->dev);
1678 
1679 	return retval;
1680 }
1681 
get_wave_state(struct device_queue_manager * dqm,struct queue * q,void __user * ctl_stack,u32 * ctl_stack_used_size,u32 * save_area_used_size)1682 static int get_wave_state(struct device_queue_manager *dqm,
1683 			  struct queue *q,
1684 			  void __user *ctl_stack,
1685 			  u32 *ctl_stack_used_size,
1686 			  u32 *save_area_used_size)
1687 {
1688 	struct mqd_manager *mqd_mgr;
1689 
1690 	dqm_lock(dqm);
1691 
1692 	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
1693 
1694 	if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
1695 	    q->properties.is_active || !q->device->cwsr_enabled ||
1696 	    !mqd_mgr->get_wave_state) {
1697 		dqm_unlock(dqm);
1698 		return -EINVAL;
1699 	}
1700 
1701 	dqm_unlock(dqm);
1702 
1703 	/*
1704 	 * get_wave_state is outside the dqm lock to prevent circular locking
1705 	 * and the queue should be protected against destruction by the process
1706 	 * lock.
1707 	 */
1708 	return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack,
1709 			ctl_stack_used_size, save_area_used_size);
1710 }
1711 
process_termination_cpsch(struct device_queue_manager * dqm,struct qcm_process_device * qpd)1712 static int process_termination_cpsch(struct device_queue_manager *dqm,
1713 		struct qcm_process_device *qpd)
1714 {
1715 	int retval;
1716 	struct queue *q;
1717 	struct kernel_queue *kq, *kq_next;
1718 	struct mqd_manager *mqd_mgr;
1719 	struct device_process_node *cur, *next_dpn;
1720 	enum kfd_unmap_queues_filter filter =
1721 		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
1722 	bool found = false;
1723 
1724 	retval = 0;
1725 
1726 	dqm_lock(dqm);
1727 
1728 	/* Clean all kernel queues */
1729 	list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
1730 		list_del(&kq->list);
1731 		decrement_queue_count(dqm, qpd, kq->queue);
1732 		qpd->is_debug = false;
1733 		dqm->total_queue_count--;
1734 		filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
1735 	}
1736 
1737 	/* Clear all user mode queues */
1738 	list_for_each_entry(q, &qpd->queues_list, list) {
1739 		if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1740 			deallocate_sdma_queue(dqm, q);
1741 		else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1742 			deallocate_sdma_queue(dqm, q);
1743 
1744 		if (q->properties.is_active)
1745 			decrement_queue_count(dqm, qpd, q);
1746 
1747 		dqm->total_queue_count--;
1748 	}
1749 
1750 	/* Unregister process */
1751 	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1752 		if (qpd == cur->qpd) {
1753 			list_del(&cur->list);
1754 			kfree(cur);
1755 			dqm->processes_count--;
1756 			found = true;
1757 			break;
1758 		}
1759 	}
1760 
1761 	retval = execute_queues_cpsch(dqm, filter, 0);
1762 	if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) {
1763 		pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
1764 		dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
1765 		qpd->reset_wavefronts = false;
1766 	}
1767 
1768 	/* Lastly, free mqd resources.
1769 	 * Do free_mqd() after dqm_unlock to avoid circular locking.
1770 	 */
1771 	while (!list_empty(&qpd->queues_list)) {
1772 		q = list_first_entry(&qpd->queues_list, struct queue, list);
1773 		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1774 				q->properties.type)];
1775 		list_del(&q->list);
1776 		qpd->queue_count--;
1777 		dqm_unlock(dqm);
1778 		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1779 		dqm_lock(dqm);
1780 	}
1781 	dqm_unlock(dqm);
1782 
1783 	/* Outside the DQM lock because under the DQM lock we can't do
1784 	 * reclaim or take other locks that others hold while reclaiming.
1785 	 */
1786 	if (found)
1787 		kfd_dec_compute_active(dqm->dev);
1788 
1789 	return retval;
1790 }
1791 
init_mqd_managers(struct device_queue_manager * dqm)1792 static int init_mqd_managers(struct device_queue_manager *dqm)
1793 {
1794 	int i, j;
1795 	struct mqd_manager *mqd_mgr;
1796 
1797 	for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
1798 		mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
1799 		if (!mqd_mgr) {
1800 			pr_err("mqd manager [%d] initialization failed\n", i);
1801 			goto out_free;
1802 		}
1803 		dqm->mqd_mgrs[i] = mqd_mgr;
1804 	}
1805 
1806 	return 0;
1807 
1808 out_free:
1809 	for (j = 0; j < i; j++) {
1810 		kfree(dqm->mqd_mgrs[j]);
1811 		dqm->mqd_mgrs[j] = NULL;
1812 	}
1813 
1814 	return -ENOMEM;
1815 }
1816 
1817 /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
allocate_hiq_sdma_mqd(struct device_queue_manager * dqm)1818 static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
1819 {
1820 	int retval;
1821 	struct kfd_dev *dev = dqm->dev;
1822 	struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd;
1823 	uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size *
1824 		get_num_all_sdma_engines(dqm) *
1825 		dev->device_info->num_sdma_queues_per_engine +
1826 		dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
1827 
1828 	retval = amdgpu_amdkfd_alloc_gtt_mem(dev->kgd, size,
1829 		&(mem_obj->gtt_mem), &(mem_obj->gpu_addr),
1830 		(void *)&(mem_obj->cpu_ptr), false);
1831 
1832 	return retval;
1833 }
1834 
device_queue_manager_init(struct kfd_dev * dev)1835 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1836 {
1837 	struct device_queue_manager *dqm;
1838 
1839 	pr_debug("Loading device queue manager\n");
1840 
1841 	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
1842 	if (!dqm)
1843 		return NULL;
1844 
1845 	switch (dev->device_info->asic_family) {
1846 	/* HWS is not available on Hawaii. */
1847 	case CHIP_HAWAII:
1848 	/* HWS depends on CWSR for timely dequeue. CWSR is not
1849 	 * available on Tonga.
1850 	 *
1851 	 * FIXME: This argument also applies to Kaveri.
1852 	 */
1853 	case CHIP_TONGA:
1854 		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
1855 		break;
1856 	default:
1857 		dqm->sched_policy = sched_policy;
1858 		break;
1859 	}
1860 
1861 	dqm->dev = dev;
1862 	switch (dqm->sched_policy) {
1863 	case KFD_SCHED_POLICY_HWS:
1864 	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1865 		/* initialize dqm for cp scheduling */
1866 		dqm->ops.create_queue = create_queue_cpsch;
1867 		dqm->ops.initialize = initialize_cpsch;
1868 		dqm->ops.start = start_cpsch;
1869 		dqm->ops.stop = stop_cpsch;
1870 		dqm->ops.pre_reset = pre_reset;
1871 		dqm->ops.destroy_queue = destroy_queue_cpsch;
1872 		dqm->ops.update_queue = update_queue;
1873 		dqm->ops.register_process = register_process;
1874 		dqm->ops.unregister_process = unregister_process;
1875 		dqm->ops.uninitialize = uninitialize;
1876 		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1877 		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1878 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1879 		dqm->ops.set_trap_handler = set_trap_handler;
1880 		dqm->ops.process_termination = process_termination_cpsch;
1881 		dqm->ops.evict_process_queues = evict_process_queues_cpsch;
1882 		dqm->ops.restore_process_queues = restore_process_queues_cpsch;
1883 		dqm->ops.get_wave_state = get_wave_state;
1884 		break;
1885 	case KFD_SCHED_POLICY_NO_HWS:
1886 		/* initialize dqm for no cp scheduling */
1887 		dqm->ops.start = start_nocpsch;
1888 		dqm->ops.stop = stop_nocpsch;
1889 		dqm->ops.pre_reset = pre_reset;
1890 		dqm->ops.create_queue = create_queue_nocpsch;
1891 		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1892 		dqm->ops.update_queue = update_queue;
1893 		dqm->ops.register_process = register_process;
1894 		dqm->ops.unregister_process = unregister_process;
1895 		dqm->ops.initialize = initialize_nocpsch;
1896 		dqm->ops.uninitialize = uninitialize;
1897 		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1898 		dqm->ops.set_trap_handler = set_trap_handler;
1899 		dqm->ops.process_termination = process_termination_nocpsch;
1900 		dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
1901 		dqm->ops.restore_process_queues =
1902 			restore_process_queues_nocpsch;
1903 		dqm->ops.get_wave_state = get_wave_state;
1904 		break;
1905 	default:
1906 		pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
1907 		goto out_free;
1908 	}
1909 
1910 	switch (dev->device_info->asic_family) {
1911 	case CHIP_CARRIZO:
1912 		device_queue_manager_init_vi(&dqm->asic_ops);
1913 		break;
1914 
1915 	case CHIP_KAVERI:
1916 		device_queue_manager_init_cik(&dqm->asic_ops);
1917 		break;
1918 
1919 	case CHIP_HAWAII:
1920 		device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
1921 		break;
1922 
1923 	case CHIP_TONGA:
1924 	case CHIP_FIJI:
1925 	case CHIP_POLARIS10:
1926 	case CHIP_POLARIS11:
1927 	case CHIP_POLARIS12:
1928 	case CHIP_VEGAM:
1929 		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
1930 		break;
1931 
1932 	case CHIP_VEGA10:
1933 	case CHIP_VEGA12:
1934 	case CHIP_VEGA20:
1935 	case CHIP_RAVEN:
1936 	case CHIP_RENOIR:
1937 	case CHIP_ARCTURUS:
1938 		device_queue_manager_init_v9(&dqm->asic_ops);
1939 		break;
1940 	case CHIP_NAVI10:
1941 	case CHIP_NAVI12:
1942 	case CHIP_NAVI14:
1943 	case CHIP_SIENNA_CICHLID:
1944 	case CHIP_NAVY_FLOUNDER:
1945 		device_queue_manager_init_v10_navi10(&dqm->asic_ops);
1946 		break;
1947 	default:
1948 		WARN(1, "Unexpected ASIC family %u",
1949 		     dev->device_info->asic_family);
1950 		goto out_free;
1951 	}
1952 
1953 	if (init_mqd_managers(dqm))
1954 		goto out_free;
1955 
1956 	if (allocate_hiq_sdma_mqd(dqm)) {
1957 		pr_err("Failed to allocate hiq sdma mqd trunk buffer\n");
1958 		goto out_free;
1959 	}
1960 
1961 	if (!dqm->ops.initialize(dqm))
1962 		return dqm;
1963 
1964 out_free:
1965 	kfree(dqm);
1966 	return NULL;
1967 }
1968 
deallocate_hiq_sdma_mqd(struct kfd_dev * dev,struct kfd_mem_obj * mqd)1969 static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev,
1970 				    struct kfd_mem_obj *mqd)
1971 {
1972 	WARN(!mqd, "No hiq sdma mqd trunk to free");
1973 
1974 	amdgpu_amdkfd_free_gtt_mem(dev->kgd, mqd->gtt_mem);
1975 }
1976 
device_queue_manager_uninit(struct device_queue_manager * dqm)1977 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1978 {
1979 	dqm->ops.uninitialize(dqm);
1980 	deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
1981 	kfree(dqm);
1982 }
1983 
kfd_process_vm_fault(struct device_queue_manager * dqm,u32 pasid)1984 int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid)
1985 {
1986 	struct kfd_process_device *pdd;
1987 	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1988 	int ret = 0;
1989 
1990 	if (!p)
1991 		return -EINVAL;
1992 	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
1993 	pdd = kfd_get_process_device_data(dqm->dev, p);
1994 	if (pdd)
1995 		ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
1996 	kfd_unref_process(p);
1997 
1998 	return ret;
1999 }
2000 
kfd_process_hw_exception(struct work_struct * work)2001 static void kfd_process_hw_exception(struct work_struct *work)
2002 {
2003 	struct device_queue_manager *dqm = container_of(work,
2004 			struct device_queue_manager, hw_exception_work);
2005 	amdgpu_amdkfd_gpu_reset(dqm->dev->kgd);
2006 }
2007 
2008 #if defined(CONFIG_DEBUG_FS)
2009 
seq_reg_dump(struct seq_file * m,uint32_t (* dump)[2],uint32_t n_regs)2010 static void seq_reg_dump(struct seq_file *m,
2011 			 uint32_t (*dump)[2], uint32_t n_regs)
2012 {
2013 	uint32_t i, count;
2014 
2015 	for (i = 0, count = 0; i < n_regs; i++) {
2016 		if (count == 0 ||
2017 		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
2018 			seq_printf(m, "%s    %08x: %08x",
2019 				   i ? "\n" : "",
2020 				   dump[i][0], dump[i][1]);
2021 			count = 7;
2022 		} else {
2023 			seq_printf(m, " %08x", dump[i][1]);
2024 			count--;
2025 		}
2026 	}
2027 
2028 	seq_puts(m, "\n");
2029 }
2030 
dqm_debugfs_hqds(struct seq_file * m,void * data)2031 int dqm_debugfs_hqds(struct seq_file *m, void *data)
2032 {
2033 	struct device_queue_manager *dqm = data;
2034 	uint32_t (*dump)[2], n_regs;
2035 	int pipe, queue;
2036 	int r = 0;
2037 
2038 	if (!dqm->sched_running) {
2039 		seq_printf(m, " Device is stopped\n");
2040 
2041 		return 0;
2042 	}
2043 
2044 	r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd,
2045 					KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE,
2046 					&dump, &n_regs);
2047 	if (!r) {
2048 		seq_printf(m, "  HIQ on MEC %d Pipe %d Queue %d\n",
2049 			   KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1,
2050 			   KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm),
2051 			   KFD_CIK_HIQ_QUEUE);
2052 		seq_reg_dump(m, dump, n_regs);
2053 
2054 		kfree(dump);
2055 	}
2056 
2057 	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
2058 		int pipe_offset = pipe * get_queues_per_pipe(dqm);
2059 
2060 		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
2061 			if (!test_bit(pipe_offset + queue,
2062 				      dqm->dev->shared_resources.cp_queue_bitmap))
2063 				continue;
2064 
2065 			r = dqm->dev->kfd2kgd->hqd_dump(
2066 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
2067 			if (r)
2068 				break;
2069 
2070 			seq_printf(m, "  CP Pipe %d, Queue %d\n",
2071 				  pipe, queue);
2072 			seq_reg_dump(m, dump, n_regs);
2073 
2074 			kfree(dump);
2075 		}
2076 	}
2077 
2078 	for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) {
2079 		for (queue = 0;
2080 		     queue < dqm->dev->device_info->num_sdma_queues_per_engine;
2081 		     queue++) {
2082 			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
2083 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
2084 			if (r)
2085 				break;
2086 
2087 			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
2088 				  pipe, queue);
2089 			seq_reg_dump(m, dump, n_regs);
2090 
2091 			kfree(dump);
2092 		}
2093 	}
2094 
2095 	return r;
2096 }
2097 
dqm_debugfs_execute_queues(struct device_queue_manager * dqm)2098 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm)
2099 {
2100 	int r = 0;
2101 
2102 	dqm_lock(dqm);
2103 	dqm->active_runlist = true;
2104 	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
2105 	dqm_unlock(dqm);
2106 
2107 	return r;
2108 }
2109 
2110 #endif
2111